{"schema_version":1,"dataset_version":"1.0.0","generated":"2026-08-18T20:36:57.532Z","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/","attribution":"The AI Tells Index, feedsquad.com/ai-tells"},"count":137,"entries":[{"id":"delve-excess-vocabulary","name":"Excess-vocabulary cluster (delve, underscore, intricate)","aka":["delve","AI vocabulary","excess vocabulary","AI words","WP:AIVOCAB","delves","delving","deep dive","tapestry","rich tapestry","vibrant","bustling","nestled","beacon","realm","in the realm of","intricate","intricacies","garner","bolster","interplay","boasts","showcasing","2023 marker vocabulary"],"category":"lexical","subcategory":"excess-vocabulary","description":"A set of ordinary English style words whose frequency jumped after late 2022. Kobak and colleagues measured the jump across more than 15 million PubMed abstracts, with delves at about 25 times its extrapolated pre-2022 baseline. The words are not new. What changed is how often several of them arrive in the same short passage, which is the effect Kousha and Thelwall tracked across six scholarly databases. The detection here is a density count, not a word ban: distinct markers drawn from the two published lists, counted per 1000 words. The threshold of 3 is a working default, not a published cutoff. It is set there because the published lists contain common academic words with real human base rates, so one hit carries almost no information and Wikipedia's own discipline rule is that co-occurrence is the signal. Recalibrate it against a measured corpus before treating it as a number rather than a starting point.","why_it_reads_ai":"One marker word means nothing. The cluster is what readers react to: several of these words in a paragraph with no proper noun or number between them. The studies behind the list measured populations, not documents. A sentence containing delve tells you about the corpus it came from and almost nothing about who typed it.","examples":[{"before":"This piece delves into the intricate landscape of remote onboarding, underscoring the pivotal part a robust welcome sequence plays in long-term retention. The interplay between structure and warmth is what most teams eventually garner from the exercise. Remote work remains a vibrant realm, and the organisations that treat it as one tend to bolster retention without ever quite saying how they did it. Practitioners would do well to interrogate their assumptions before wholesale adoption, since what proves efficacious in one setting may prove considerably less so in another.","after":"Our welcome sequence is four emails over nine days. The third one asks the reader to name the job they hired the product for. Those answers predict cancellations better than anything in our analytics.","note":"The repair is not synonym swapping. The specimen states nothing about the sequence at any point. The after text says how long it runs and what it is used for."}],"detection":{"type":"statistical","metric":"distinct-excess-vocabulary-markers-per-1000-words","threshold":3,"direction":"above","threshold_basis":"Working default, not a published cutoff. Kobak and colleagues measure excess vocabulary across 15 million abstracts at corpus scale and publish no per-document threshold, because the method is population-level by design. Three distinct markers per 1,000 words is a FeedSquad review trigger chosen so that one ordinary use of one word cannot fire it. It flags text for a human to read. It decides nothing."},"severity":"medium","status":"fading","status_history":[{"date":"2026-08-14","status":"fading","rationale":"The Washington Post measured delve in roughly 1 in 1,000 publicly shared ChatGPT messages by July 2025, well below its 2023 peak. Yakura and colleagues found the same words rising in unscripted human speech after ChatGPT's release, with a preregistered experiment showing adoption after brief exposure. The cluster still measures something at corpus scale. It no longer separates one author from another."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Biomedical and STEM academics used these words before 2022. Kobak's method measures excess over an extrapolated baseline, which means the baseline was real and nonzero. Nigerian English speakers publicly objected in April 2024 that delve into is ordinary business register for them, and no corpus study settling that question has been published, so the dispute stands open. Podcast speakers now produce the cluster in unscripted speech, per Yakura. Non-native English writers are the group with most to lose: Liang and colleagues measured a 61.3% average false-positive rate across seven detectors on human-written TOEFL essays, and the prompt that cleared the false accusation was one that made the vocabulary fancier.","model_attribution":"Corpus-level LLM-era signal, not a family marker. Neither Kobak nor Liang assigns any word to a model family. Wikipedia's era buckets suggest the vocabulary drifted between model generations, and that periodisation is editor observation rather than measurement.","platform_notes":[{"platform":"wikipedia","note":"Wikipedia keeps a per-word sourced list and states the rule literally: a word being overused by AI does not imply its synonyms are. Editors must corroborate a word from a non-pop-science source before adding it."},{"platform":"linkedin","note":"No LinkedIn policy names any word. The May 2026 announcement targets posts with no unique perspective and reduces distribution outside a person's network rather than removing the post."}],"languages":["en"],"sources":[{"kind":"external","title":"Kobak, Gonzalez-Marquez, Horvat, Lause: Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Same study, arXiv HTML v1 (excess frequency ratios and gaps)","url":"https://arxiv.org/html/2406.07016v1","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Kousha and Thelwall: How much are LLMs changing the language of academic papers after ChatGPT?","url":"https://arxiv.org/abs/2509.09596","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Yakura et al.: Empirical evidence of Large Language Model's influence on human spoken communication","url":"https://arxiv.org/abs/2409.01754","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Merrill, Chen, Kumer: What are the clues that ChatGPT wrote something? We analyzed its style, Washington Post","url":"https://www.washingtonpost.com/technology/interactive/2025/how-detect-chatgpt-em-dash/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Liang, Yuksekgonul, Mao, Wu, Zou: GPT detectors are biased against non-native English writers, Patterns 4:100779 (PubMed Central copy)","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC10382961/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Vanguard Nigeria: Nigerians tackle American author for claiming delve is only used by ChatGPT","url":"https://www.vanguardngr.com/2024/04/nigerians-tackle-american-author-for-claiming-delve-is-only-used-by-chatgpt/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Juzek and Ward: Why Does ChatGPT Delve So Much? (COLING 2025)","url":"https://arxiv.org/abs/2412.11385","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Geng and Trotta: coevolution of human and LLM writing (Findings of ACL 2025)","url":"https://aclanthology.org/2025.findings-acl.657/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"LexA-Index, CC0 per-language overuse dataset","url":"https://github.com/fsu-nlp/lexa-index","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-14","updated":"2026-08-17"},{"id":"underscores-the-importance","name":"Significance inflation (underscores the importance)","aka":["underscores the importance","highlights its significance","emphasizes the importance","WP:AILEGACY","emphasising its significance","highlighting the importance of","underscores the need"],"category":"lexical","subcategory":"importance-assertion","description":"A verb of emphasis bolted to an abstract noun of importance. The sentence asserts that something matters without saying to whom, by how much, or on whose account. Kobak measured underscores at about nine times its extrapolated pre-2022 baseline in biomedical abstracts, and Juzek and Ward list underscore among their 21 focal overrepresented words. Wikipedia files the exact phrasing under undue emphasis on significance and legacy.","why_it_reads_ai":"It is a claim about a claim. Nothing in the clause can be checked, so nothing in it can be wrong. Models produce it when the retrieved material will not support a specific statement, which is the same moment a human writer would stop and go find one.","examples":[{"before":"The launch underscores the importance of community feedback in modern product development. Teams that listen to their users early tend to build products those users stay with, and the ones that skip that step usually learn the same lesson later and at greater cost. Feedback is only valuable when someone acts on it.","after":"Fourteen people in the beta channel asked for keyboard shortcuts before we had built any. We shipped the shortcuts first and the settings page second because of those fourteen messages.","note":"Importance is not a property of the launch. It is a verdict the reader is asked to accept on no evidence. The repair names who said what and what changed because of it."}],"detection":{"type":"deterministic","pattern":"\\b(?:underscor(?:e|es|ed|ing)|highlight(?:s|ed|ing)?|emphasi[sz](?:e|es|ed|ing))\\s+(?:the|its|their)\\s+(?:importance|significance)\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-14","status":"active","rationale":"Wikipedia's era buckets keep emphasizing and highlighting in the mid-2025 onward set, and record that Grok still overuses underscore as of 2026. No source reports the phrasing declining."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Grant writers and science journalists write this sentence for a living, and it is house style in research-institution press releases and in the discussion section of medical papers. Reinhart and colleagues found that LLM prose sits closest to the dense academic register, which means a genuinely academic human writer matches this profile far more often than a blogger does. The pattern requires an emphasis object for a reason: underscore also names a typographic mark and a piece of incidental music, and Wikipedia flags both senses.","model_attribution":"Wikipedia's editor-maintained era buckets place emphasizing and highlighting in current output and underscore in Grok's. That periodisation is observation by editors rather than measurement, and should be read as a hypothesis about drift.","platform_notes":[{"platform":"wikipedia","note":"Filed under undue emphasis on significance, legacy and broader trends. Shortcut WP:AILEGACY."},{"platform":"google","note":"Google's helpful-content guidance asks whether a page provides original information, reporting, research or analysis. A sentence that asserts importance and supplies no evidence is the shape that question is written to catch."}],"languages":["en"],"sources":[{"kind":"external","title":"Kobak et al., arXiv HTML v1: excess frequency ratio for underscores","url":"https://arxiv.org/html/2406.07016v1","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Juzek and Ward: Why Does ChatGPT Delve So Much? Exploring the Sources of Lexical Overrepresentation in LLMs (COLING 2025)","url":"https://arxiv.org/abs/2412.11385","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Wikipedia: Signs of AI writing, section on undue emphasis on significance, legacy and broader trends","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Reinhart et al.: Do LLMs write like humans? PNAS 122(8) (PubMed Central copy)","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11874169/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Google Search Central: Creating helpful, reliable, people-first content","url":"https://developers.google.com/search/docs/fundamentals/creating-helpful-content","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Kobak et al.: Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Reinhart et al.: Do LLMs write like humans? (arXiv:2410.16107)","url":"https://arxiv.org/abs/2410.16107","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-14","updated":"2026-08-17"},{"id":"formal-transition-crutch","name":"Formal transition crutch","aka":["additionally","furthermore","moreover","consequently","nonetheless","thereby","accordingly","subsequently","connective stacking","AFL connectives"],"category":"lexical","subcategory":"transition-words","description":"Formal connectives used as the joint itself. Additionally, furthermore, moreover and consequently announce a logical relation between two sentences that were going to sit next to each other anyway. The taxonomy harvest behind this index normalised 13 independent public lists of AI writing markers. Three items were named by 11 of the 13: additionally, delve and underscore. The detection here counts sentence-initial connectives per 1,000 words rather than banning any of them, because each one has a real job in English and one occurrence carries almost no information.","why_it_reads_ai":"A model producing text one sentence at a time has a strong pull toward marking the seam. The connective arrives before the relation exists, so the reader is told that a consequence follows and then reads a restatement. Dense academic prose stacks connectives too. That is why this is a density measure, and a weak one.","examples":[{"before":"Additionally, the onboarding flow was updated. Furthermore, the copy was revised for clarity, and the revision was reviewed internally before anything went out. Consequently, users are now better supported throughout the signup process. Moreover, the change reflects a broader commitment to clarity that the organisation has been working toward for some time. Additionally, feedback gathered during the review has been logged and will inform the next phase of work. Subsequently, the same approach will be applied elsewhere.","after":"We cut onboarding from six screens to three on 4 June. The two screens we dropped asked for company size and job title, and nothing in the product ever read either field. Signups now finish about a minute faster.","note":"Figures in this repair are invented for the specimen. The repair works by naming what changed and when. Deleting the connectives on their own would have left the same empty sentences behind."}],"detection":{"type":"statistical","metric":"formal-connective-sentence-openers-per-1000-words","threshold":6,"direction":"above","threshold_basis":"No published per-document cutoff exists for connective density. Kobak and colleagues measure excess vocabulary across 15 million abstracts at corpus scale, and Juzek and Ward test where the overrepresentation comes from; neither publishes a document-level line, because both methods are population-level by design. Six sentence-initial formal connectives per 1,000 words is a FeedSquad review trigger, set high enough that one well-signposted academic paragraph does not fire on its own. Recalibrate against a measured corpus of the genre under review before treating the number as anything but a prompt to read the text."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Wikipedia keeps the formal connectives in its current list, vale-ai-tells ships machine-checkable rules for them, and additionally is one of the three items named by 11 of the 13 independent lists we normalised. Geng and Trotta show publicised markers decaying faster than unpublicised ones, so this entry should be expected to move."}],"evidence_grade":"corroborated","false_positive_notes":"This is taught vocabulary. The Academic Word List and the Academic Formulas List put these connectives in front of every student who has taken a course in English for academic purposes, and second-language writers produce them at higher rates than native speakers because that is what the instruction rewards. Legal drafting and standards documents use thereby and accordingly as terms of the trade. Technical specifications signpost on purpose, because a reader who skips a clause has to be able to find the thread again. A reviewer should read what sits between the connectives rather than counting them: honest signposting joins two different claims, and the crutch version joins a claim to its own paraphrase. Wikipedia lists transition words in isolation among its ineffective indicators, which is the same finding from the other direction.","model_attribution":"No family attribution. The connectives appear across the public list corpus without a vendor named, and Kobak measures the vocabulary at corpus scale in biomedical abstracts rather than per model.","platform_notes":[{"platform":"wikipedia","note":"Editors list transition words in isolation under ineffective indicators, and require corroboration from a non-pop-science source before any word joins the maintained list."},{"platform":"google","note":"Google rates content on effort, originality and added value rather than on tooling. Connective density is invisible to that question, which is the right way round."}],"languages":["en"],"sources":[{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Juzek and Ward, Why Does ChatGPT Delve So Much? COLING 2025 (arXiv:2412.11385)","url":"https://arxiv.org/abs/2412.11385","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Geng and Trotta, coevolution of human and LLM writing, Findings of ACL 2025","url":"https://aclanthology.org/2025.findings-acl.657/","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"importance-adjective-inflation","name":"Importance-inflation adjectives","aka":["crucial","vital","essential","paramount","pivotal","plays a crucial role","pivotal moment"],"category":"lexical","subcategory":"importance-assertion","description":"An adjective asserts that something matters and the sentence never says what turns on it. Crucial, vital, essential, paramount and pivotal all do the same work: they raise the stakes without naming them. This family sits in the released excess-word set that Kobak and colleagues measured across more than 15 million biomedical abstracts. The pattern is deliberately narrow. It matches the fixed formula plays a crucial role and a short list of stakes-free predicates, and it leaves the bare adjectives alone, because safety writing needs them and firing on them would punish the people who have the least room to soften.","why_it_reads_ai":"Importance is cheap to assert and expensive to demonstrate. A model that has not retrieved a consequence still has to finish the sentence, and the adjective closes it at no cost. The reader cannot disagree with the claim because no claim was made.","examples":[{"before":"Onboarding plays a crucial role in whether a new customer stays. Expectations get set early, and expectations are what people remember when renewal comes around. The principle is not in dispute anywhere, which is precisely why so little turns on stating it again. Doing something about it is a different conversation, and it is the one that keeps getting deferred.","after":"Customers who connect a second channel in their first week renew at roughly twice the rate of those who never do. Onboarding exists to get that second channel connected.","note":"Figures in this repair are invented for the specimen. The repair names the thing that changes and the number attached to it."},{"before":"A clear changelog is vital to the success of any release. Users want to know what changed, and telling them plainly builds the kind of trust that pays off downstream. The teams that write good release notes rarely have to explain themselves twice. The ones that skip them find out why eventually.","after":"We stopped publishing release notes for two months in spring. Support tickets asking what changed went up. The two questions we answered most became the first two lines of the changelog when it came back.","note":"No number is invented here. The repair is the incident."}],"detection":{"type":"deterministic","pattern":"\\bplay(?:s|ed|ing)?\\s+an?\\s+(?:crucial|vital|pivotal|critical|key|essential)\\s+role\\b|\\b(?:is|are|remains?)\\s+(?:absolutely\\s+)?(?:crucial|vital|paramount|pivotal)\\s+(?:to|for)\\s+(?:the\\s+)?(?:success|growth|future|survival|development|adoption|understanding)\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"The family is current on the Wikipedia list, carries rules in vale-ai-tells and slop-lint, and sits in the Kobak excess-word set. No source reports it declining."}],"evidence_grade":"corroborated","false_positive_notes":"Safety documentation asserts importance because that is the job. A checklist saying hand hygiene is essential is the instruction, not padding, and the same holds for clinical protocol, aviation procedure and any guidance written to be obeyed under time pressure. Plays a crucial role is separately a taught academic formula: it appears in the Academic Formulas List, and second-language writers produce it at high rates because every course puts it there. The discriminator is the consequence. A reviewer should ask what happens if the crucial thing is missing, and if the text answers within a sentence or two, the adjective is doing honest work. The pattern deliberately skips bare uses of essential for this reason, which means it under-fires on purpose.","model_attribution":"No family attribution. The adjectives sit in the Kobak released excess-word set, which is measured across abstracts without reference to any vendor.","platform_notes":[{"platform":"google","note":"The quality rater guidelines ask whether a page adds value beyond what is already available. A sentence that asserts stakes and supplies none is the shape that question catches."}],"languages":["en"],"sources":[{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"novelty-impact-superlative","name":"Unanchored novelty superlatives","aka":["groundbreaking","cutting-edge","state-of-the-art","game-changing","game-changer","revolutionize","unprecedented","best-in-class","transformative","innovative"],"category":"lexical","subcategory":"superlative","description":"Novelty and impact claimed at maximum strength with nothing to check the claim against. Groundbreaking, cutting-edge, state-of-the-art and best-in-class are comparative words used without a comparison, and revolutionize is a verb of scale used without one. Microsoft names the same vocabulary as unsupported claiming in guidance written for human copywriters, which is worth noting because it means the objection predates the tooling. The pattern matches the attributive slot in front of a generic product noun, so a specific claim about a specific instrument survives it.","why_it_reads_ai":"The words carry a promise the sentence never has to keep. They are also what a model reaches for when the retrieved material contains a product name and no measurement. A human writing about a real advance usually cannot resist naming what it beat.","examples":[{"before":"Our groundbreaking new platform will revolutionize the way teams plan their content. Planning is the part everybody puts off, and putting it off has a way of costing more than it saves. What we built takes that seriously and makes the weekly version of it something you do not have to think about. You will notice the difference sooner than you expect to.","after":"The planner now drafts the coming week from the last four weeks of your own posts instead of from an empty prompt. That is the only thing it does differently from the version before it.","note":"The repair replaces the scale claim with the mechanism, which the reader can check by using it."}],"detection":{"type":"deterministic","pattern":"\\b(?:groundbreaking|game-chang(?:ing|er)|best-in-class|state-of-the-art|cutting-edge)\\s+(?:new\\s+)?(?:solution|platform|technology|approach|tool|feature|innovation|product|suite)\\b|\\brevolutioniz(?:e|es|ed|ing)\\s+the\\s+way\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Named by four independent lists in the harvest and by Microsoft in first-party marketing guidance. Nothing reports it fading, and the low severity reflects how much ordinary promotional copy shares the register."}],"evidence_grade":"corroborated","false_positive_notes":"Institutional communications staff, grant writers and press officers are trained to write this way, and are often reviewed on whether they did. A university announcing a state-of-the-art facility is following a house style that predates any language model by decades. Sales copy in hardware and instrumentation uses best-in-class as a comparative with a real benchmark behind it, and in that case the benchmark is usually one sentence away. The pattern only fires in front of a generic product noun for exactly this reason: cutting-edge microscopy escapes it, because the noun carries the specificity the adjective lacks. A reviewer should look for the referent rather than the adjective, and treat a hit as a question about what the thing is being compared to.","model_attribution":"Not attributed to a family. Juzek and Ward find no established cause for lexical overrepresentation at all: architecture and training data are ruled out and the evidence for preference tuning is mixed.","platform_notes":[{"platform":"bing","note":"Microsoft advertising guidance for AI search visibility tells writers to anchor claims to evidence rather than to superlatives. That is advice with no ranking claim attached, and it should not be read as one."}],"languages":["en"],"sources":[{"kind":"external","title":"Juzek and Ward, Why Does ChatGPT Delve So Much? COLING 2025 (arXiv:2412.11385)","url":"https://arxiv.org/abs/2412.11385","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Microsoft Advertising, optimizing content for inclusion in AI search answers","url":"https://about.ads.microsoft.com/en/blog/post/october-2025/optimizing-your-content-for-inclusion-in-ai-search-answers","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"is-this-ai-slop word and phrase list","url":"https://github.com/didrod205/is-this-ai-slop","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"latinate-quantity-flourish","name":"Latinate quantity flourish","aka":["myriad","a myriad of","plethora","a plethora of","an array of","a wide range of"],"category":"lexical","subcategory":"ornate-diction","description":"An ornate quantity noun where a number or the word many would carry the same information. A myriad of options and a plethora of features tell the reader that the count is large and withhold the count. The pattern matches the fixed determiner phrases only. It leaves the bare noun alone, and it leaves an array of alone on purpose, because an array is a literal object in engineering and photography and the phrase there is a description rather than a flourish.","why_it_reads_ai":"The construction fills the slot where a number belongs. It reads as effort without being effort, which is the recurring shape of low-effort prose: the sentence performs care and carries no fact.","examples":[{"before":"The dashboard surfaces a myriad of signals from a wide range of connected sources. Some of them matter more than others, and which ones matter comes down to what you are trying to learn. Teams that look at it regularly build a feel for their own numbers that no single metric gives them. That feel is the part worth having.","after":"The dashboard reads five things: posting cadence, reply latency, follower change, link clicks, and the share of posts that got no reply at all. The last one is the only one that ever changed what we did.","note":"The repair names the count and then says which item mattered."}],"detection":{"type":"deterministic","pattern":"\\b(?:a\\s+myriad\\s+of|myriad\\s+of|a\\s+plethora\\s+of|a\\s+wide\\s+(?:range|array|variety)\\s+of)\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Named by three independent community lists including one whose own purpose is inverted, which is recorded on the source. No published measurement exists for this family, so the grade stays community-observed and the severity stays low."}],"evidence_grade":"community-observed","false_positive_notes":"Literary and humanities registers use these words without irony, and myriad has been ordinary English since the sixteenth century in both its adjectival and its nominal form. Writers taught to vary their diction reach for them because varying diction is what they were told to do, and that instruction falls hardest on second-language writers whose graders reward it. Catalogue copy and academic survey articles legitimately need a quantity word before a list the writer is about to give in full. A reviewer should check whether the count appears anywhere nearby: a plethora followed by seven named items is a stylistic choice, and a plethora followed by nothing is a missing number.","model_attribution":"Undocumented at family level. The evidence here is practitioner list convergence, not measurement, and the entry should not be read as saying more.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"is-this-ai-slop word and phrase list","url":"https://github.com/didrod205/is-this-ai-slop","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"keskinonur gist, words and phrases to avoid for ChatGPT","url":"https://gist.github.com/keskinonur/4b2d9b7f3311332cf60c91cb45efb362","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"capability-release-verb","name":"Capability-release verbs","aka":["unlock","unleash","unlock the potential","unlock the power","empower","supercharge","elevate"],"category":"lexical","subcategory":"metaphor-verbs","description":"Release and elevation metaphors bolted to ordinary improvements. Unlock the potential, unleash the power, supercharge and empower share one move: they describe a capability as having been trapped, so that the product can be the thing that frees it. The pattern matches the fixed collocations rather than the verbs, because unlocking a door and unlocking a game level are literal and common.","why_it_reads_ai":"The metaphor supplies drama that the underlying change does not. It also lets a sentence describe a benefit without naming a mechanism, which is precisely the position a model is in when it has a product name and no product.","examples":[{"before":"Our new templates unlock the full potential of your content team and supercharge your weekly workflow. A good template takes the decisions nobody wants to make and makes them once, so the work that follows is only the work. Teams that adopt them stop staring at empty documents. The output speaks for itself soon enough.","after":"The templates are five post shapes we saw working in our own drafts: a teardown, a number with the method attached, an apology for something we got wrong, a question with our answer underneath it, and a link with the reason we kept the tab open.","note":"The repair says what the templates are. The reader can now disagree with the list, which is the point."}],"detection":{"type":"deterministic","pattern":"\\bunlock(?:s|ing|ed)?\\s+(?:the\\s+)?(?:full\\s+)?(?:potential|power|value)\\b|\\bunleash(?:es|ing|ed)?\\s+(?:the\\s+)?(?:full\\s+)?(?:potential|power|creativity)\\b|\\bsupercharg(?:e|es|ed|ing)\\s+your\\b|\\bempower(?:s|ing|ed)?\\s+(?:you|your\\s+team|teams|users|customers)\\s+to\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Current across three independent community lists and present in the Kobak excess-word set. Growth marketing keeps producing it under human authorship, which is why the severity is low."}],"evidence_grade":"community-observed","false_positive_notes":"Growth marketing and games copy built this register and product marketers write it deliberately for audiences who expect it, so a hit inside a landing page tells you about the genre before it tells you about the author. Unlock is literal in access control, in games, and in anything with a key. Empower is a term of art in law and in governance, where empowering a body to act names an actual grant of authority. Community and development organisations use it in the sense their funders use it. The pattern requires the fixed collocations so that these survive, and a reviewer looking at a hit should ask what capability was supposedly trapped and by what.","model_attribution":"Not attributed to a family. Community lists name the family and no published study isolates it, which is what the community-observed grade means here.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tellsign word and phrase lists","url":"https://github.com/ctkrug/tellsign","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"domain-metaphor-noun","name":"Domain metaphor nouns","aka":["the competitive landscape","the AI landscape","ecosystem","arena","sphere","frontier"],"category":"lexical","subcategory":"metaphor-nouns","description":"A spatial or biological metaphor noun standing in where the plain name of the subject would go. The competitive landscape, the developer ecosystem and similar phrases name a shape rather than a set of companies, products or people. The status here is contested, and it should be. Strategy consultants and platform economists coined several of these usages and use them as terms of art with a definition behind them.","why_it_reads_ai":"The metaphor lets a sentence be about a whole field without knowing anything specific about it. That is a useful property when the writer has no particulars, and a wasteful one when they do.","examples":[{"before":"The competitive landscape for scheduling tools shifted again this quarter, and the developer ecosystem is responding in the ways you would expect it to. Consolidation follows attention, attention has not been in short supply, and where any of it settles is a question most of the people watching closely have already answered for themselves without waiting for the evidence.","after":"Two scheduling tools shipped an agent that drafts and publishes without a human step this quarter. Both charge per seat for it. Nobody has published what happens when the agent posts something wrong.","note":"The repair replaces the shape word with the two things that actually moved."}],"detection":{"type":"deterministic","pattern":"\\bthe\\s+(?:competitive|AI|digital|marketing|business)\\s+landscape\\b(?!\\s+of\\b)|\\bthe\\s+(?:startup|developer|content|partner|creator)\\s+ecosystem\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Contested from the start. Wikipedia lists the family, vale-ai-tells ships rules for it and the LexA per-language data carries the vocabulary, but ecosystem is a defined technical term in two separate fields and the phrase has a documented pre-2022 business register. Listed because the metaphor is worth noticing, graded so that nobody treats a single hit as meaning anything."}],"evidence_grade":"corroborated","false_positive_notes":"Biologists use ecosystem literally, and have since 1935. Platform economists and strategy researchers use it as a defined term for a set of firms with complementary products, and a paper in that field will use it forty times without a single one being decoration. Competitive landscape has been standard consulting and equity-research vocabulary since long before any language model existed. The pattern only fires on business-modified forms, and it explicitly excludes the opener phrase in the ever-evolving landscape of, which belongs to the scene-setting entry rather than this one. A reviewer should ask whether the writer could have named the members of the ecosystem, and treat the hit as a prompt to ask, never as a finding.","model_attribution":"Not attributed. Wikipedia editors observe era drift in this vocabulary, and that periodisation is editor observation rather than measurement.","platform_notes":[{"platform":"wikipedia","note":"The maintained list carries this family with the standing rule that a word being overused does not implicate its synonyms."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"LexA-Index, CC0 per-language overuse dataset","url":"https://github.com/fsu-nlp/lexa-index","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"journey-navigation-metaphor","name":"Journey and navigation metaphors","aka":["navigate","navigating the complexities","embark","embark on a journey","your journey","traverse"],"category":"lexical","subcategory":"metaphor-verbs","description":"Movement metaphors applied to reading, learning and ordinary work. Navigating the complexities, embarking on a journey and your content journey all describe sitting still. The family is fading rather than active: it is on every avoid list published since 2023, which is exactly the condition under which a marker stops separating anyone from anyone.","why_it_reads_ai":"A journey frame gives a paragraph an arc it has not earned, and it flatters the reader for showing up. Both are cheap, and both are what a text produces when it has no events to narrate.","examples":[{"before":"This guide helps you navigate the complexities of social scheduling as you embark on your content journey. Everyone arrives at this point with different habits, and the aim here is to meet you where you already are. Progress rarely looks like progress while it is happening. Give it time and the pieces start to sit together on their own.","after":"This guide covers connecting an account, the two settings that decide when a post goes out, and what to do when a platform rejects one. If you read only one part, read the rejection part, because that is where people lose posts.","note":"The repair states the contents and then tells the reader which part is load-bearing."}],"detection":{"type":"deterministic","pattern":"\\bnavigat(?:e|es|ing|ed)\\s+the\\s+(?:complexities|challenges|nuances|intricacies|world)\\b|\\bembark(?:s|ed|ing)?\\s+on\\s+(?:a|your|this|the)\\s+(?:journey|adventure)\\b|\\byour\\s+(?:learning|content|growth|fitness)\\s+journey\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"fading","status_history":[{"date":"2026-08-15","status":"fading","rationale":"Marked fading because the family appears on every published avoid list, including the two gists in the sources, and Geng and Trotta measured publicised markers losing frequency soon after they became famous. A word that everyone has been told to remove tells you about the editing, not the drafting."}],"evidence_grade":"community-observed","false_positive_notes":"User journey is the standard term in product design and service design, with mapped stages and an established literature; a designer writing it is using the vocabulary of the trade. Coaching, therapy and patient-experience writing use journey in a sense their readers ask for, and cancer and recovery narratives in particular have made it the accepted word. Travel writing uses embark literally, as does anything about ships. Navigation is literal in aviation, sailing, and interface design. The pattern targets only the fixed abstract collocations, and it does not touch user journey at all. A reviewer seeing a hit should check whether anything in the text actually moves.","model_attribution":"Undocumented at family level. Both gists in the sources carry avoid-list framing, which makes them evidence of what circulates rather than of what any model does.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tellsign word and phrase lists","url":"https://github.com/ctkrug/tellsign","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"chrisgherbert gist, ChatGPT cliches","url":"https://gist.github.com/chrisgherbert/c734ec50ae464135be57cd03b84281f9","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"},{"kind":"external","title":"LexA-Index, CC0 per-language overuse dataset","url":"https://github.com/fsu-nlp/lexa-index","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"institutional-abstraction-verb","name":"Institutional abstraction verbs","aka":["enhance","foster","fostering","facilitate","leverage","utilize","streamline","optimize"],"category":"lexical","subcategory":"abstraction-verbs","description":"Institutional verbs standing in for the specific verb for what was done. Enhance, foster, facilitate, leverage, utilize, streamline and optimize describe an improvement in the abstract, and the abstraction is the problem: the sentence survives without the writer knowing what happened. These verbs sit in the released excess-word set from Kobak and colleagues and in the LexA per-language data. The pattern matches verb plus abstract object, so facilitating a meeting and a machine that utilizes a counterweight both escape it.","why_it_reads_ai":"Each of these verbs takes almost any object, which makes them the safest possible choice when the object is unknown. A writer who watched the thing happen usually has a narrower verb available and uses it.","examples":[{"before":"The new workspace fosters collaboration and helps teams utilize a variety of tools more effectively. Work gets easier to see when it lives in one place, which is usually the thing that was missing.","after":"The workspace has one shared calendar. Before it existed, two people published to the same channel twenty minutes apart on 12 May and neither knew the other had scheduled anything.","note":"Figures in this repair are invented for the specimen. The repair names the failure the change was made against."}],"detection":{"type":"deterministic","pattern":"\\b(?:foster(?:s|ed|ing)?|facilitat(?:e|es|ed|ing))\\s+(?:an?\\s+|the\\s+)?(?:culture|environment|collaboration|innovation|engagement|dialogue|growth|community)\\b|\\bleverag(?:e|es|ed|ing)\\s+(?:our|their|your|its)\\s+\\w+\\s+(?:to|for)\\b|\\butiliz(?:e|es|ed|ing)\\s+(?:an?\\s+|the\\s+)?(?:variety|range|approach|framework|strategy|tools?|resources)\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active. The verbs are in the Kobak excess-word set and carry current rules in vale-ai-tells, and nothing in the evidence base reports them declining. Severity is medium because the abstraction hides content rather than merely decorating it."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Education, development and grant writing use foster and facilitate as the standard terms, and a funder who asked for a proposal about fostering literacy will expect the proposal to say fostering literacy. Facilitation is a named profession with its own certification, so a facilitator facilitating is not a euphemism. Consulting and finance adopted leverage decades before any model existed and use it with a precise meaning about borrowed capital. Optimize is a mathematical term. The pattern therefore requires an abstract object, which lets the concrete uses through: the crane that utilizes a counterweight and the meeting that was facilitated on 3 March both escape. A reviewer should ask what the narrower verb would have been, and whether the writer could have supplied it.","model_attribution":"No vendor attribution. Membership in the Kobak released excess-word set is a corpus-level fact about biomedical abstracts, and the paper assigns nothing to any model family.","platform_notes":[{"platform":"wikipedia","note":"The maintained list keeps these verbs in its current set and pairs them with the corroboration rule that governs every word on it."}],"languages":["en"],"sources":[{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"LexA-Index, CC0 per-language overuse dataset","url":"https://github.com/fsu-nlp/lexa-index","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"canned-conclusion-phrase","name":"Canned conclusion phrases","aka":["in conclusion","in summary","to summarize","to sum up","all in all"],"category":"lexical","subcategory":"conclusion-marker","description":"A phrase that announces an ending instead of being one. In conclusion, in summary and to sum up were taught as essay scaffolding long before any of this, and they remain correct in genres that require an explicit summary move. Contested for that reason. The pattern is sentence-initial and requires the punctuation that follows the phrase in its scaffolding use, so in summary judgment and a stray all in all mid-sentence both escape.","why_it_reads_ai":"The phrase does the work of ending without the work of concluding. What follows it is usually the opening restated, which is a separate tell with its own entry: this one is about the announcement, and the structural entry on signposted conclusions is about the restatement.","examples":[{"before":"In conclusion, consistency matters more than volume when you are building an audience. Showing up on a schedule you can keep is worth more than a burst you cannot repeat, and most people learn that the expensive way. An audience notices the rhythm long before it notices any single post. That is the part worth holding on to.","after":"We posted three times a week for a quarter, then nine times a week for the next one. The nine-a-week quarter got more impressions and fewer replies. We went back to three.","note":"Figures in this repair are invented for the specimen. The repair ends by stopping, which is what an ending is."}],"detection":{"type":"deterministic","pattern":"^\\s*(?:in\\s+conclusion|in\\s+summary|to\\s+summari[sz]e|to\\s+sum\\s+up|all\\s+in\\s+all)\\s*[,:]","flags":"i","scope":"sentence"},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Contested. Named by four independent lists, and simultaneously the standard closing move taught in school essay instruction and required by the structure of an abstract, an executive summary or a legal brief. The signal is about the genre it appears in more than about the sentence."}],"evidence_grade":"corroborated","false_positive_notes":"Taught essay scaffolding puts these phrases in place and rewards them, so student writing and the writing of anyone recently graded is full of them. Several genres require an explicit summary move: abstracts, executive summaries, appellate briefs and any document whose reader may only read the last paragraph. Spoken registers put the phrase in for a listener who cannot scroll back. The pattern requires sentence-initial position and the comma or colon that follows the scaffolding use, which is why in summary judgment the court found for the tenant does not fire, and why an ordinary all in all with no punctuation after it does not either. A reviewer should read what follows the phrase, because the announcement is harmless and the restatement is the real problem.","model_attribution":"No family attribution, and the phrase is old enough that any attribution would be wrong.","platform_notes":[{"platform":"wikipedia","note":"Editors record the closing formula among current signs while keeping their own caution that formal prose alone indicates nothing."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"chrisgherbert gist, ChatGPT cliches","url":"https://gist.github.com/chrisgherbert/c734ec50ae464135be57cd03b84281f9","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"},{"kind":"external","title":"is-this-ai-slop word and phrase list","url":"https://github.com/didrod205/is-this-ai-slop","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"precision-hedge-cluster-2026","name":"2026 precision-hedge cluster","aka":["comparatively","markedly","attributable","typically","modest","stratify","quantify","standardize","minimize"],"category":"lexical","subcategory":"era-2026","description":"The current frontier vocabulary profile, which is nothing like the 2023 one. We extracted the CC0 LexA English science split for GPT-5.2 on 2026-08-14 and ranked words by their model-to-human occurrence ratio. Delve, tapestry and vibrant do not appear in the split at all, and meticulous appears once on each side. What sits at the top is precision and hedge vocabulary: comparatively at about 25 times the human rate in that split, modest at about 10, minimize at about 9, attributable at about 8, typically at about 8. That ranking is ours. No third party has published this profile, which is why the grade is community-observed and why this entry counts markers rather than banning any of them.","why_it_reads_ai":"The 2023 flourishes were trained down and something replaced them. What replaced them reads as careful rather than ornate, which makes it harder to notice and much more dangerous to flag. A directory built on the old word list is a museum piece, so this entry exists mainly to date-stamp the problem.","examples":[{"before":"Results were comparatively modest and largely attributable to seasonal variation, which we typically observe in this period. Performance across segments was not markedly different once the data had been standardized, and the residual variation is difficult to quantify with any confidence. Steps have been taken to minimize the effect going forward, though it should be acknowledged that the underlying drivers are only partially understood at this stage. Management remains cautiously optimistic, subject to the usual caveats around comparability. The picture remains broadly stable.","after":"Signups fell 12 percent in July. They fell 14 percent last July. We think it is the summer holiday and we will know in September, when the comparison stops being a guess.","note":"Figures in this repair are invented for the specimen. The repair swaps the hedge for the two numbers the hedge was standing in front of."}],"detection":{"type":"statistical","metric":"precision-hedge-markers-per-1000-words","threshold":4,"direction":"above","threshold_basis":"Ours, and provisional. The word ranking comes from our own extraction of the CC0 LexA English science split for GPT-5.2 on 2026-08-14; nobody has published a per-document threshold for this vocabulary, because nobody has published the profile at all. Four distinct markers per 1,000 words is a FeedSquad review trigger chosen so that one hedged sentence cannot fire it. The false-positive risk here is the worst in the directory, since this is also how careful quantitative writing sounds. Treat the number as a prompt to read the text, and recalibrate it against a corpus in the genre before treating it as anything else."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active, and new. The extraction on 2026-08-14 puts precision and hedge vocabulary at the top of the frontier science split while the 2023 markers are absent from it. Juzek documents post-2022 lexical uptake continuing across languages, and Geng and Trotta explain the substitution mechanism: publicised markers decay and unpublicised ones keep rising."}],"evidence_grade":"community-observed","false_positive_notes":"Epidemiologists, statisticians, clinical trialists and anyone writing a results section produce this register by obligation, and the register is a virtue there. Attributable and stratify are technical terms with defined meanings in that context. Regulatory and pharmacovigilance writing is hedged because overclaiming is a compliance failure. This makes the cluster the most dangerous one in the directory to point at technical prose, and it should not be pointed at technical prose. The intended use is marketing copy, social posts and general commentary, where a stack of hedges usually means the writer had no measurement. A reviewer should check whether the numbers the hedges qualify appear anywhere; hedging around a stated figure is careful writing, and hedging around nothing is filler.","model_attribution":"Extracted from the LexA science split for GPT-5.2 specifically. Treat it as a dated snapshot of one model in one register, not as a property of frontier models in general, and expect it to move.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"LexA-Index, CC0 per-language overuse dataset","url":"https://github.com/fsu-nlp/lexa-index","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Juzek, LLM lexical uptake across 34 languages, preprint (arXiv:2605.25358)","url":"https://arxiv.org/abs/2605.25358","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Geng and Trotta, coevolution of human and LLM writing, Findings of ACL 2025","url":"https://aclanthology.org/2025.findings-acl.657/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"feedsquad-observed","title":"FeedSquad extraction of the LexA GPT-5.2 English science split","observed":"2026-08-14","corpus":"CC0 LexA-Index release, English science register, GPT-5.2 split: 66,865 word rows with model and human counts, ranked by log-probability-ratio. Delve, delves, tapestry and vibrant absent from the split; meticulous present once on each side. Ranking performed by us and not published elsewhere."}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"gpt4o-creative-cluster","name":"GPT-4o creative-register cluster","aka":["camaraderie","palpable","fleeting","unspoken","whimsical","gossamer","amidst","pang","unease"],"category":"lexical","subcategory":"era-2024","description":"Creative-writing vocabulary that was overrepresented in the GPT-4o era and is decaying as a group. Camaraderie, palpable, fleeting, unspoken, whimsical, gossamer and amidst arrive together in narrative and reflective writing, and Reinhart and colleagues measured per-word overrepresentation for this family against human writing in PNAS. Their multipliers sit in the paper body and we could not confirm them at page level, so this entry keeps the ranking qualitative rather than printing a number we have not verified.","why_it_reads_ai":"The words are how a text signals that a moment was meaningful without describing the moment. They cluster in exactly the passages that have no event in them. Fiction writers use the same vocabulary with events attached, which is what separates the two.","examples":[{"before":"There was a palpable camaraderie in the room, a fleeting and unspoken sense that something had shifted. Amidst the noise a whimsical thought surfaced and dissolved again before anyone could name it. Underneath it all sat a pang of unease, gossamer thin, the kind of feeling that lingers long after the moment that made it has gone. The air held a certain electric quality, ineffable and faintly bittersweet, and the whole scene seemed to shimmer at its edges in a way that nobody present would have been able to articulate afterwards.","after":"Four of us stayed after standup and argued about the pricing page for an hour. Nobody won. The page shipped with the old headline and a new second line that somebody muttered near the end.","note":"The repair supplies the event the atmosphere words were standing in for."}],"detection":{"type":"statistical","metric":"creative-register-markers-per-1000-words","threshold":2,"direction":"above","threshold_basis":"Reinhart and colleagues measure per-word overrepresentation against human writing at population level and publish no per-document cutoff; the fold changes for this family are body-level figures in the paper and we did not confirm them at page level, so no multiplier is printed here. Two distinct markers per 1,000 words is a FeedSquad review trigger, set low because the family clusters tightly when it appears at all. It flags a passage for reading and decides nothing, and it should not be run over fiction."},"severity":"low","status":"fading","status_history":[{"date":"2026-08-15","status":"fading","rationale":"Fading. The family is tied to the GPT-4o generation that Reinhart measured, and the frontier science split we extracted in August 2026 shows a different profile entirely. The vocabulary also sits on published avoid lists, which is the condition under which a marker stops separating anyone from anyone."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Fiction, memoir and literary journalism use this vocabulary as ordinary equipment, and a novelist writing palpable is doing their job. Reinhart measures model output against human writing at population level and makes no claim about any individual text, which is the correct reading of the entry too. Amidst is standard in British and Indian English and carries no register signal in either. Reflective and devotional writing is built from unspoken and fleeting. The discriminator is whether an event is attached: a described room with named people in it can carry any of these words, and a room with only atmosphere in it is the thing worth flagging. Never run this entry over fiction, where a hit means only that the text is fiction.","model_attribution":"Tied to the GPT-4o generation by Reinhart, who report larger differences from human writing for instruction-tuned models than for base models. Later generations show a different vocabulary, so read this entry as dated.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Reinhart et al., Do LLMs write like humans? PNAS 122(8) (arXiv:2410.16107)","url":"https://arxiv.org/abs/2410.16107","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"chrisgherbert gist, ChatGPT cliches","url":"https://gist.github.com/chrisgherbert/c734ec50ae464135be57cd03b84281f9","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"review-praise-register","name":"Peer-review praise register","aka":["commendable","meticulous","meticulously","compelling","noteworthy","invaluable"],"category":"lexical","subcategory":"evaluative-praise","description":"Evaluative praise vocabulary that rose sharply in conference peer review after 2022. Liang and colleagues measured review text at four AI conferences and estimated that between 6.5 and 16.9 percent of it was substantially modified by a language model after ChatGPT, against a pre-ChatGPT control of 1.6 to 2.4 percent. The control is what makes that number citable, and it is the number this entry rests on. Per-word fold changes appear in the paper body and we did not confirm them at page level, so they are not printed here.","why_it_reads_ai":"Praise vocabulary is what a text produces when it has read something and has nothing to say about it. A commendable effort and a compelling narrative are verdicts with no reasons attached, and a reviewer with a reason usually leads with the reason.","examples":[{"before":"The authors have done a commendable job, and the meticulously crafted methodology makes for a compelling read. The framing is timely and the argument is organised well throughout. The literature review is thorough and situates the contribution appropriately within the wider body of work. A few small clarifications would strengthen the presentation, though none of them touch the substance of what is claimed, and the reviewer has no material concerns to raise on that front. The manuscript would make a welcome addition to the issue.","after":"The sampling frame excludes anyone who left the platform before 2024, which is the population the paper is about. Table 3 would settle it if it reported the attrition count.","note":"The repair is the objection. It is also shorter, which is usual."}],"detection":{"type":"deterministic","pattern":"\\bcommendable\\s+(?:effort|work|job|attempt|contribution)\\b|\\bmeticulous(?:ly)?\\s+(?:crafted|curated|researched|documented|designed)\\b|\\ba\\s+(?:truly\\s+)?compelling\\s+(?:case|narrative|argument|read)\\b|\\b(?:is|are|was|were)\\s+(?:a\\s+)?(?:noteworthy|invaluable)\\s+(?:contribution|addition|resource)\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"fading","status_history":[{"date":"2026-08-15","status":"fading","rationale":"Fading. Liang measured the rise in reviews written between 2023 and 2024, and the vocabulary has since been named in every venue that discusses this, including conference policies. Yakura documents the same words entering unscripted human speech, which drains a word-level marker of information regardless of what any model does next."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Reviewers, referees, examiners and anyone writing a school report are required to evaluate, and evaluation vocabulary is the tool for it. Recommendation letters are built from this register, and a letter writer who avoids it damages the candidate. Meticulous is literal in conservation, watchmaking, archaeology and laboratory technique, where it describes an observable working method rather than a compliment. Yakura and colleagues measured these words rising in spontaneous podcast speech after 2022, which means a human writing commendable in 2026 may simply have heard it. The discriminator is whether a reason follows the verdict within a sentence or two. A hit on a peer review is a prompt to ask what the reviewer noticed, and never a finding about who wrote it.","model_attribution":"Measured across review text at four machine learning conferences without attribution to a vendor. Liang estimates the share of substantially modified text in a corpus and makes no claim about any single review.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Liang et al., Monitoring AI-Modified Content at Scale, ICML 2024 (arXiv:2403.07183)","url":"https://arxiv.org/abs/2403.07183","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Yakura et al., Empirical evidence of LLM influence on human spoken communication (arXiv:2409.01754)","url":"https://arxiv.org/abs/2409.01754","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"quality-assertion-adjective","name":"Quality-assertion adjectives","aka":["seamless","seamlessly","robust","comprehensive","versatile","scalable","holistic"],"category":"lexical","subcategory":"product-adjectives","description":"Adjectives that assert quality in a slot where the sentence could have demonstrated it. Seamless, robust, comprehensive, versatile, scalable and holistic are the standard product-page vocabulary, and they are also the standard vocabulary of a page written by someone who has not used the product. Contested, because half of these words are terms of art elsewhere. The pattern matches the fixed marketing collocations only, which lets the technical senses through.","why_it_reads_ai":"These adjectives are unfalsifiable in the slot they occupy. A seamless integration cannot be checked and a two-click connection can, so the version with the number in it is the version somebody tested.","examples":[{"before":"A robust platform with seamless integration and a comprehensive set of publishing features. Everything is built to work the way you already work, so there is nothing new to learn and nothing at all to configure.","after":"It connects to LinkedIn, X, Threads, Bluesky and Mastodon. Instagram needs a business account. There is no TikTok connector and no date for one.","note":"The repair is a list of what exists, including the gap. The gap is the part that makes it credible."}],"detection":{"type":"deterministic","pattern":"\\bseamless(?:ly)?\\s+(?:integrat(?:e|es|ed|ion)|experience|workflow)\\b|\\ba\\s+(?:robust|comprehensive|holistic|scalable)\\s+(?:solution|platform|approach|framework|strategy|suite)\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Contested. Named across four sources including the Kobak excess-word set and the LexA data, and simultaneously the native vocabulary of product marketing, statistics and systems engineering. The word alone signals genre rather than authorship, which is why the pattern is restricted to the marketing collocations."}],"evidence_grade":"corroborated","false_positive_notes":"Robust is a precise term in statistics, control theory and structural engineering, where a robust estimator has a defined breakdown point and a robust design has a stated tolerance. Scalable has a specific meaning in distributed systems. Holistic is standard in nursing, education and several therapeutic traditions. Product and user-experience writing owns seamless outright, and a designer describing a handoff as seamless usually means they removed a step and can tell you which one. Comprehensive is required in insurance and in the title of a great many reference works. The pattern therefore fires only on the marketing collocations, and a reviewer should check whether the sentence names a mechanism anywhere near the adjective.","model_attribution":"No vendor attribution. These adjectives predate the current models by decades in product copy, and their presence in the excess-word set is a statement about frequency change, not origin.","platform_notes":[{"platform":"google","note":"Quality rating turns on effort, originality and added value. A page of quality adjectives with no demonstrated mechanism is the shape those criteria are written against, whatever produced it."}],"languages":["en"],"sources":[{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"LexA-Index, CC0 per-language overuse dataset","url":"https://github.com/fsu-nlp/lexa-index","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"depth-claim-adjective","name":"Depth-claim adjectives","aka":["profound","nuanced","multifaceted","deeply rooted"],"category":"lexical","subcategory":"depth-claims","description":"Adjectives that claim analytical depth for analysis that never arrives. A nuanced understanding, a profound impact and a multifaceted challenge each promise a complication and then decline to name one. The pattern matches the fixed noun phrases rather than the adjectives, because profound has an ordinary sense and multifaceted is often the finding itself.","why_it_reads_ai":"The claim of depth is the cheapest available substitute for depth. It is also the move a text makes when it has summarised two positions and cannot adjudicate between them.","examples":[{"before":"The report offers a nuanced understanding of a multifaceted problem deeply rooted in team culture. The findings resist easy summary, which is part of what makes them worth sitting with. Culture is not fixed in a quarter, and treating it as a project rather than a habit is a mistake organisations make more than once before anything about it registers.","after":"Two teams missed the same deadline. One had nobody assigned to the final step and the other had three people assigned to it. Naming one owner fixed the first case in a day. The second case is still open.","note":"The repair supplies the complication that the word nuanced was promising."}],"detection":{"type":"deterministic","pattern":"\\ba\\s+(?:more\\s+)?nuanced\\s+(?:understanding|view|perspective|take|approach)\\b|\\ba\\s+profound\\s+(?:impact|effect|shift|change)\\b|\\bmultifaceted\\s+(?:approach|challenge|issue|problem)\\b|\\bdeeply\\s+rooted\\s+in\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active. Present in the Kobak excess-word set and current across three community taxonomies, with nothing reporting decline. Severity is low because the adjectives are common in honest writing and the pattern only catches the fixed phrases."}],"evidence_grade":"corroborated","false_positive_notes":"Philosophy, theology and obituary writing use profound in its ordinary sense, and a profound effect is standard clinical phrasing for an effect of large magnitude. Multifaceted is frequently the finding rather than a decoration: much social science exists to establish that a problem has several dimensions, and the paper that establishes it has earned the word. Deeply rooted is literal in botany and forestry. Nuance is the object of study in linguistics and in law. The pattern targets the fixed phrases for that reason, and a reviewer should ask what the second facet is, because a text that can name it is doing the work the adjective claims.","model_attribution":"No family attribution documented. The adjectives appear in the Kobak set as corpus-level frequency changes.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tellsign word and phrase lists","url":"https://github.com/ctkrug/tellsign","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"testament-formula","name":"A testament to","aka":["a testament to","testament to the power of","indelible mark"],"category":"lexical","subcategory":"significance-formula","description":"A formula that converts any fact into evidence of greatness. A testament to says that the thing just mentioned demonstrates a quality, and the demonstration is never shown. Wikipedia files it with the puffery family, and three independent community lists carry it. Fading, because it is one of the most publicised phrases in this whole area and human editors now remove it on sight.","why_it_reads_ai":"The formula upgrades a fact into a verdict without an argument in between. That upgrade is free, and free is what low-effort prose is made of.","examples":[{"before":"The response from the community is a testament to the power of building in public. People want to see the work and not only the outcome, and showing it costs a good deal less than most founders assume before they try it. Teams that share early hear the useful criticism while it is still cheap to act on, which is worth considerably more than applause.","after":"Forty-one people replied to the roadmap post. Nine asked for the same thing, an export button, and it is the next thing we build.","note":"Figures in this repair are invented for the specimen. The repair replaces the verdict with the count and what the count changed."}],"detection":{"type":"deterministic","pattern":"\\ba\\s+testament\\s+to\\b|\\bleft\\s+an\\s+indelible\\s+mark\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"fading","status_history":[{"date":"2026-08-15","status":"fading","rationale":"Fading. Named by four sources including the Wikipedia list, where editors treat it as a routine removal, and by the community linters. A phrase that trained editors strip automatically stops distinguishing between drafts."}],"evidence_grade":"corroborated","false_positive_notes":"Eulogies, tributes, award citations and retirement speeches use the formula because the occasion calls for exactly that move, and removing it from a eulogy would be a mistake. Sports writing and obituaries share the register. The phrase is also ordinary in religious writing in its literal sense. There is a boundary to respect with the copula-avoidance entry, which owns the copula-substitute verbs: a sentence that pairs one of those verbs with the testament noun trips that rule rather than this one, and this pattern deliberately matches the bare formula only. A reviewer looking at a hit should treat genre as the first question. In a wedding speech it is correct. In a product changelog it is a missing number.","model_attribution":"Not attributed to a family. The phrase is old and its recent frequency is a corpus observation from community lists rather than a measurement.","platform_notes":[{"platform":"wikipedia","note":"Filed with promotional language. Editors remove it as puffery under the same guidance that governs unsupported praise from any source."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"is-this-ai-slop word and phrase list","url":"https://github.com/didrod205/is-this-ai-slop","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"figurative-verb-family","name":"Figurative verb family","aka":["carries weight","casts a shadow","strikes a chord","sits at the heart of","lends itself to"],"category":"lexical","subcategory":"figurative-verbs","description":"A small set of figurative verbs recycled as the default way to say that something matters. Carries real weight, casts a long shadow, strikes a chord, sits at the heart of and lends itself to are live metaphors in the hands of a feature writer and dead ones everywhere else. The evidence is two independent linters and nothing measured, which is what the community-observed grade records.","why_it_reads_ai":"The set is small and the substitution is frictionless. Any of these verbs fits any subject, so the choice between them carries no information about the subject.","examples":[{"before":"The pricing change carries real weight, and the reasoning behind it sits at the heart of how we think about value. A price is a statement about who a product is for, and statements like that are hard to take back. Getting one wrong casts a long shadow over everything that follows it. We would rather be clear now than clever later.","after":"We raised the price of an agent from 29 to 39 euros in June and kept the old price for everyone who had already paid. Two customers left. Both said the same thing, that they used one agent and were being priced as though they used four.","note":"Figures in this repair are invented for the specimen. The repair states the change, the exception, and the cost of it."}],"detection":{"type":"deterministic","pattern":"\\bcarries\\s+(?:real|significant|extra|more)\\s+weight\\b|\\bcasts?\\s+a\\s+long\\s+shadow\\b|\\bstrikes?\\s+a\\s+chord\\b|\\bsits\\s+at\\s+the\\s+heart\\s+of\\b|\\blends\\s+itself\\s+(?:well\\s+)?to\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active, on thin evidence. Two independent rule sets carry the family and no published study measures it, so the entry ships at low severity with the grade stated. It is listed because the substitution test is easy to run and useful, not because anyone has counted it."}],"evidence_grade":"community-observed","false_positive_notes":"Sports writing, feature journalism and music and theatre criticism use these verbs as working metaphors, and a critic writing that a performance struck a chord is describing an audience reaction they watched happen. Casts a long shadow is standard in history and biography for an influence that outlasts a person. Lends itself to is ordinary technical English for a material or method that suits a purpose. The pattern requires the fixed collocations, so a crane that carries weight up to nine tonnes and a guitarist playing a chord both escape it. A reviewer should try the substitution test: if any other verb in the family would fit the sentence equally well, the verb is carrying no meaning, and if only one fits, it is.","model_attribution":"Undocumented. The family comes from two rule sets built by practitioners, and no measurement exists to attribute it to anything.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"emphatic-adverb-tic","name":"Emphatic adverbs","aka":["truly","genuinely","straightforward","really","simply put"],"category":"lexical","subcategory":"vendor-suppressed","description":"Intensifying adverbs that add emphasis and no content. What lifts this entry above list folklore is first-party documentation: Anthropic publishes its product system prompts with dates, and the Claude Opus 5 prompt of 2026-07-24 instructs the model to avoid saying genuinely, honestly and straightforward, with the 4.8 prompt of 2026-05-28 naming genuinely, honestly and actually. A vendor writing a suppression instruction is documenting the tendency it is suppressing, which is dated evidence about a default pull rather than about what the shipped product emits.","why_it_reads_ai":"The adverb performs sincerity in a sentence that has not earned any. It is also load-bearing for nothing: remove it and the claim is unchanged, which is the test.","examples":[{"before":"This is truly remarkable work. Simply put: nobody else is doing it, and the care behind it is genuinely rare in a category this crowded.","after":"We tried five other tools in July and none of them let you approve a week of posts from one screen. If we missed one, tell us and we will correct this page.","note":"The repair replaces the intensifier with the count, the month, and an invitation to be shown wrong."}],"detection":{"type":"deterministic","pattern":"\\b(?:truly|genuinely)\\s+(?:remarkable|exciting|special|impressive|unique|humbled|grateful)\\b|\\bsimply\\s+put\\s*[,:]","flags":"gi","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active, and unusually well evidenced for a word-level entry. Anthropic dated the suppression instruction across three model releases between May and July 2026, which places the tendency in current systems by the vendor own account. The pattern is nonetheless narrow, because the words themselves are ordinary English."}],"evidence_grade":"primary-doc","false_positive_notes":"Conversational essayists, newsletter writers and anyone putting speech on the page use these adverbs naturally, because spoken English is full of them and a transcript that removes them stops sounding like a person. Straightforward is plain English and is often the most accurate word available, which is why this pattern does not match it at all despite the vendor instruction naming it. Sincerity markers also do real work in apologies and in bad news, where a flat register reads as indifference. The pattern requires an intensifier plus a specific evaluative adjective for that reason. A reviewer should apply the deletion test: if the sentence means exactly the same thing without the adverb, the adverb was decoration, and if it does not, leave it alone.","model_attribution":"Anthropic names genuinely, honestly and actually in the Claude Opus 4.8 system prompt dated 2026-05-28, and genuinely, honestly and straightforward in the Claude Opus 5 prompt dated 2026-07-24. That is evidence about a default pull the vendor is correcting, and it says nothing about what any other family does.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Anthropic dated system prompts, release notes","url":"https://platform.claude.com/docs/en/release-notes/system-prompts","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"multilingual-lexical-uptake","name":"Multilingual lexical uptake","aka":["cross-language overuse profile","non-English marker lists","LexA language coverage"],"category":"lexical","subcategory":"non-english","description":"The post-2022 lexical shift is not an English problem. Juzek tested 34 languages and found post-ChatGPT lexical uptake in 26 of them, with a mean prevalence increase of 15.1 percent. The companion CC0 dataset covers those languages in a news register. The English word lists that circulate do not transfer: a Finnish or German text does not become suspect because it contains a translation of delve, and no per-language marker list is published that we could responsibly reprint. This entry exists to record the finding and to stop the English list being applied where it does not belong.","why_it_reads_ai":"The same pull produces the same shape in any language: a verb of depth, a noun of significance, and no date anywhere in the paragraph. What changes across languages is the vocabulary, not the emptiness.","examples":[{"before":"Tämä artikkeli syventyy etätyön monimutkaiseen kokonaisuuteen ja korostaa selkeän viestinnän merkitystä. Luottamus rakentuu hitaasti ja katoaa nopeasti, minkä useimmat tiimit oppivat vasta jälkikäteen. Toimivat käytännöt syntyvät harvoin kerralla, vaan ne muotoutuvat matkan varrella. Lopulta ratkaisee se, miten arjessa oikeasti toimitaan, ei se, mitä ohjeeseen on kirjoitettu.","after":"Siirryimme etätyöhön maaliskuussa. Palaverit vähenivät neljästä kahteen viikossa, ja kirjoitettu muistio korvasi loput. Muistio on nyt ensimmäinen asia, jonka uusi työntekijä lukee.","note":"The specimen is Finnish. Figures in this repair are invented for the specimen. The before text makes the same move as the English version: a verb of depth and a noun of significance where a month and a number would do."},{"before":"Dieser Beitrag beleuchtet die vielschichtigen Herausforderungen der Fernarbeit und unterstreicht die Bedeutung klarer Kommunikation. Vertrauen entsteht langsam und geht schnell verloren, was die meisten Teams erst im Nachhinein bemerken. Gute Gewohnheiten lassen sich nicht verordnen, sie wachsen mit der Zeit. Am Ende zählt, wie im Alltag tatsächlich gearbeitet wird, und nicht, was in einem Leitfaden steht.","after":"Wir haben im März auf Fernarbeit umgestellt. Aus vier Meetings pro Woche wurden zwei, den Rest übernimmt ein schriftliches Protokoll.","note":"The specimen is German. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"distinct-per-language-uptake-markers-per-1000-words","threshold":3,"direction":"above","threshold_basis":"Carried over from the English density trigger, and weaker here. Juzek reports 26 of 34 languages showing post-2022 uptake with a mean prevalence increase of 15.1 percent, all at corpus level, and publishes no per-document cutoff and no per-language marker list. The metric is named separately from the English one because the marker list differs by language and the two numbers are not comparable. Three distinct markers per 1,000 words is therefore a FeedSquad review trigger with no per-language calibration behind it at all. Anyone applying this outside English must build the marker list from a measured corpus in that language first. Running the English list through a translator produces nothing worth reading."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active. Juzek measures uptake in 26 of 34 languages and the companion dataset covers them; Liang and colleagues find the same shift in consumer complaints, corporate communications and United Nations press releases at population scale. The source is a preprint committed to EMNLP 2026 and is labelled as such on the entry."}],"evidence_grade":"corroborated","false_positive_notes":"Translators and staff at multilingual institutions write a formal register in every language they publish in, and that register is a large part of what the measurement picks up. European Union and United Nations texts are drafted to be translatable, which flattens exactly the features a marker list keys on. Second-language writers in any target language produce the most predictable register available to them, which is the same mechanism behind the 61.3 percent average false-positive rate Liang and colleagues measured across seven detectors on non-native English essays. Languages with rich morphology break naive word matching outright, so a Finnish marker count that ignores inflection measures tokenisation rather than writing. A reviewer working outside English should treat this entry as a warning about method, not as a detector.","model_attribution":"Corpus-level and cross-model. The LexA news data was generated with a small 2025-era model and the science data with several families, so no single vendor sits behind the finding.","platform_notes":[],"languages":["en","af","ar","bg","cs","de","el","es","et","fa","fi","fr","hi","hr","id","is","it","ja","kk","ko","ky","lt","lv","mr","nl","pl","pt","ro","ru","sr","ta","tr","uk","zh"],"sources":[{"kind":"external","title":"Juzek, LLM lexical uptake across 34 languages, preprint (arXiv:2605.25358)","url":"https://arxiv.org/abs/2605.25358","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"LexA-Index, CC0 per-language overuse dataset","url":"https://github.com/fsu-nlp/lexa-index","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Kobak et al., Delving into LLM-assisted writing in biomedical publications, Science Advances 11(27)","url":"https://www.science.org/doi/10.1126/sciadv.adt3813","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Liang et al., Widespread LLM adoption in society, Patterns (arXiv:2502.09747)","url":"https://arxiv.org/abs/2502.09747","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Liang, Yuksekgonul, Mao, Wu, Zou: GPT detectors are biased against non-native English writers, Patterns 4:100779 (PubMed Central copy)","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC10382961/","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"copula-avoidance-serves-as","name":"Copula avoidance (serves as, stands as, functions as)","aka":["serves as a","stands as a","copula avoidance","is/are avoidance","WP:AINOCOPULA","the serves-as dodge","represents in place of is","WP:AINOCOPULA shortcut"],"category":"structural","subcategory":"copula-avoidance","description":"Where a plain sentence would use is or was, the text reaches for a heavier verb plus as plus an article. The claim does not change. The sentence gets longer and the subject gets vaguer. Wikipedia files this under avoidance of basic copulatives and attributes a post-2022 decline in is and are to Geng and Trotta. That figure is not stated in the paper's abstract, so treat it as reported rather than confirmed. Huang and colleagues report the same decline in Wikipedia articles, holding lead paragraphs constant.","why_it_reads_ai":"The heavier verb carries no extra information. It lifts the register and leaves the claim exactly where it was. Do it twice in a paragraph and every noun in the piece starts to read like an exhibit label.","examples":[{"before":"The old mill serves as a reminder of what the town used to be and stands as a focal point for the community. Buildings like it tend to hold a place in local memory long after the work that filled them has gone. What matters is that the space still brings people together, and that the story behind it is not forgotten.","after":"The grain mill closed in 1974 and reopened as a market hall in 2019. The Saturday produce market runs out of the old drying floor, which is the only part of the building that kept its original floorboards.","note":"Sixty words about a building, and not one of them says what happened to it. The repair says what happened to it. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:serves|stands|functions|operates)\\s+as\\s+(?:a|an|the)\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-14","status":"active","rationale":"Wikipedia lists the construction under language and grammar with no decay note, and the copula decline is reported in two separate corpora, academic abstracts and Wikipedia articles. No published measurement shows it fading."}],"evidence_grade":"corroborated","false_positive_notes":"Heritage and travel writing use this construction to describe function without asserting change, and it is standard in that register. Bare job titles after \"as\" do not match the pattern, but articled descriptions do: \"serves as the director of the county health department\" fires, and that phrasing is ordinary in bios, obituaries and board announcements. Treat a hit in biographical copy as noise unless the surrounding text also avoids dates and events.","model_attribution":"Observed across chat assistants generally. No study assigns the construction to a model family. The two corpus measurements are of human academic and encyclopedic writing after 2022, not of any single product's output.","platform_notes":[{"platform":"wikipedia","note":"Shortcut WP:AINOCOPULA. Wikipedia pairs it with the inverted tell: simple is and has phrasing appears under signs of human writing, based on 25 years of article text."},{"platform":"linkedin","note":"Nothing in LinkedIn's policy addresses sentence construction. The stated target is content that restates without adding, which this construction tends to accompany rather than cause."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing, section on avoidance of basic copulatives","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Geng and Trotta: Is ChatGPT Transforming Academics' Writing Style?","url":"https://arxiv.org/abs/2404.08627","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Huang, Xu, Geng, Wan, Chen: Wikipedia in the Era of LLMs, Evolution and Risks","url":"https://openreview.net/pdf?id=ahVmnYkVLt","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Reinhart et al.: Do LLMs write like humans? Variation in grammatical and rhetorical styles, PNAS 122(8) (PubMed Central copy)","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11874169/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Birattari: Come evitare le ripetizioni moleste quando scriviamo? (Italian style norm against repetition)","url":"https://www.illibraio.it/news/grammatica/come-evitare-ripetizioni-quando-scriviamo-540892/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-14","updated":"2026-08-17"},{"id":"participial-tackon","name":"Participial tack-on padding","aka":["WP:SUPERFICIAL","trailing participial clause","-ing trailer"],"category":"structural","subcategory":"participial-padding","description":"A sentence finishes. Then a comma arrives, and an -ing clause says the same thing again in the register of analysis. Reinhart and colleagues counted present participial clauses in GPT-4o output at 5.3 times the human rate, with a Cohen's d of 1.38, the largest effect in their grammatical table. That figure sits in the paper body rather than the abstract, and it compares two corpora. It licenses nothing about one sentence.","why_it_reads_ai":"The clause has the shape of a conclusion and none of the weight. It hangs off the main clause, adds no fact, and lets the sentence end on a note of consequence the writer never paid for. Wikipedia files the same habit under superficial analysis. Two of them in a paragraph and the prose starts reading like captions under photographs.","examples":[{"before":"The council approved the new cycle lane, marking a significant step forward for the city's transport plan. The decision was welcomed by residents, reflecting a growing appetite for safer streets. The scheme was described as a priority, underlining a commitment that has been building for some time.","after":"The council approved the new cycle lane in March. Building it closes the northbound lane outside the market hall for eleven weeks from September, and the bus route moves two streets east while it does.","note":"Three sentences, three trailing clauses, and every one of them restates the clause it hangs off. Delete all three and nothing is lost. The repair says what the approval costs. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"trailing-participial-clauses-per-1000-words","direction":"above","threshold_basis":"No threshold ships. The published number is 5.3 times the human rate with a Cohen's d of 1.38, and it is a ratio between corpora printed in a paper body, not a per-document cutoff. Nobody has published a human distribution for this rate, so any line drawn here would be invented. Measure the draft, compare it against the same writer's earlier work, and treat a jump as a reason to reread. That is a FeedSquad review trigger and not a finding."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Reinhart measured 2024-era models and no later study reports the gap closing. The construction is still carried on Wikipedia's signs-of-AI-writing page with no decay note, and vale-ai-tells ships 68 surface forms of it."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Academic and technical writers use trailing -ing clauses that carry a real result, and the clause is then load-bearing: a valve failed at 40 bar, flooding the pump room. The room flooded, which the main clause never said. The test is whether the clause introduces a noun, a figure or an event absent from the sentence it hangs off. Delete it; if nothing is lost, it was padding, and if a fact is lost, it was a sentence in the wrong clothes.","model_attribution":"Reinhart and colleagues measured GPT-4o and Llama 3 70B Instruct. The instruction-tuned models diverged from human prose further than the base models did, which is the opposite of what instruction tuning is usually assumed to do. No study assigns the habit to one vendor.","platform_notes":[{"platform":"wikipedia","note":"Listed under superficial analysis on the signs-of-AI-writing page, where the shortcut WP:SUPERFICIAL points. The page treats the trailing clause as text that performs analysis without doing any."}],"languages":["en"],"sources":[{"kind":"external","title":"Reinhart et al., Do LLMs write like humans? PNAS 122(8) (arXiv:2410.16107)","url":"https://arxiv.org/abs/2410.16107","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"nominalization-overload","name":"Nominalization overload","aka":["the implementation of","utilization","noun-heavy construction"],"category":"structural","subcategory":"nominalization","description":"Verbs freeze into abstract nouns, and then a weak verb has to carry them. Reinhart and colleagues report the construction in GPT-4o output at roughly twice the human rate, with a Cohen's d of 1.23. The number is a table value in the paper body, and the abstract does not carry it. This entry rests on that single measurement, which is the only published count of the pattern we found.","why_it_reads_ai":"The nominal version reads as institutional because that is where it comes from. It also loses the actor: once the doing becomes a thing, nobody has to be named as having done it. Writing that is trying to be exact tends to move the other way, from the noun back to the verb, and gets shorter as it goes.","examples":[{"before":"The implementation of the new review process resulted in an improvement in the identification of defects prior to release. The introduction of clearer expectations led to an increase in participation across the wider organisation. The reduction of duplication and the standardisation of terminology contributed to a general improvement in the quality of the documentation produced. The continuation of the initiative is dependent upon the allocation of resource in the next planning cycle, and a recommendation regarding the extension of the process to adjacent teams is under preparation.","after":"We changed the review rota on 4 May. Two people now read every pull request before it merges, and the number of bugs caught before release doubled over the following month.","note":"Eleven frozen verbs across three sentences, and nobody in any of them is doing anything. The repair names the actor and the date, and the verbs do the work the nouns were doing. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"nominalizations-per-1000-words","direction":"above","threshold_basis":"No threshold ships. The 2.1 times figure and the Cohen's d of 1.23 are body-level table values comparing model corpora against human corpora, and no per-document human distribution has been published. A number set here would be the fabricated precision this index catalogues. Measure it, compare it against the writer's own earlier drafts, and treat a rise as a FeedSquad review trigger."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"The only published measurement of the pattern is Reinhart, which found it at roughly twice the human rate in 2024-era models. Nothing since reports the gap closing, and no vendor documentation names the construction as suppressed."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Legal and administrative drafting is built from nominalizations, and work by Martinez and colleagues found that laypeople asked to write law produce the same construction with no training at all, which makes it a property of the genre rather than of the writer. Grant applications, standards documents and clinical protocols reward the nominal form because it names a defined thing that other clauses point back to. Check whether the actor is also missing; a named actor sitting beside nominal vocabulary is usually a lawyer, not a thin draft.","model_attribution":"Measured in GPT-4o at roughly twice the human rate and in Llama 3 70B Instruct at between 1.5 and 2 times. No vendor documents suppressing it.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Reinhart et al., Do LLMs write like humans? PNAS 122(8) (arXiv:2410.16107)","url":"https://arxiv.org/abs/2410.16107","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"that-clause-subject","name":"That-clause subjects","aka":["The fact that X shows","fronted clausal subject"],"category":"structural","subcategory":"clausal-subjects","description":"A sentence opens with a that-clause standing in as its subject, and the main verb then evaluates the proposition. Reinhart and colleagues count the construction at 2.6 times the human rate in GPT-4o output, with a Cohen's d of 0.77, the weakest effect in their grammatical table. The figure is body-level. Low severity is deliberate: the evidence here is strong and the signal is thin.","why_it_reads_ai":"Fronting a proposition and then judging it is the shape of a sentence about to announce that something matters. On its own the construction means close to nothing, and plenty of good writing uses it. It earns a place in this index as a cadence marker that travels with the heavier tells, never as a finding by itself.","examples":[{"before":"That the pilot ran without a single rollback shows that the deployment process was sound. That nobody raised a concern during the review suggests the team was comfortable with the approach. What the wider rollout will show is harder to say at this remove. That confidence is earned rather than announced is a thing organisations discover for themselves, usually more than once.","after":"The pilot ran six weeks with no rollbacks, after the team moved deploys to Tuesday mornings and added a smoke test on the payments path.","note":"Every sentence fronts a proposition and then grades it, and the grading is all the passage contains. The repair states the fact and names the two changes that produced it. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"that-clause-subjects-per-1000-words","direction":"above","threshold_basis":"No threshold ships. The 2.6 times ratio is a body-level population comparison and the effect size is the smallest in the source table, so a cutoff drawn from it would separate genres rather than drafts. Use the count as a FeedSquad review trigger alongside the other structural measures, and never on its own."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Published at 2.6 times the human rate in 2024-era model output, with no later study reporting a change. Severity is set low because the effect size is the weakest in the source table and the construction is ordinary in several genres."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Philosophy, linguistics and mathematics prose uses clausal subjects as standard equipment, because those sentences really are about propositions rather than about events. Any writing that discusses claims will produce them at a high rate, and so will a paper reporting what a result shows. Treat a hit as noise unless the surrounding text is narrative or promotional copy, where the construction has no work to do and is standing in for a plain statement.","model_attribution":"Measured in GPT-4o output. Reinhart reports the gap widening for instruction-tuned models relative to base models. No vendor names the construction in published guidance.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Reinhart et al., Do LLMs write like humans? PNAS 122(8) (arXiv:2410.16107)","url":"https://arxiv.org/abs/2410.16107","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"phrasal-coordination-habit","name":"Paired-phrase coordination","aka":["noun-pair habit","relentless X and Y pairing"],"category":"structural","subcategory":"coordination","description":"Nouns and adjectives arrive in pairs where one member would carry the whole meaning. Reinhart and colleagues measure phrasal coordination at 1.9 times the human rate in GPT-4o output, with a Cohen's d of 0.81, again a body-level table value. Pairs are two members. The three-member version belongs to the rule-of-three entry, and the two patterns do not overlap.","why_it_reads_ai":"A pair looks like care and costs nothing to produce. The second member is usually a near-synonym of the first, so the sentence lengthens without becoming more exact. Read a paragraph back and count how many pairs survive the deletion of one member with no loss of meaning. In most low-effort drafts the answer is none.","examples":[{"before":"The platform provides clarity and insight into your data, and gives teams the tools and resources they need for informed and confident decisions. It is designed to be simple and intuitive, so that everyday users and seasoned analysts alike can find what they need. The result is a smoother and more consistent workflow, and a team that feels supported and equipped at every stage. Adoption is quick and low-risk, and the value shows up early and keeps showing up.","after":"The platform shows which of your posts drew replies from accounts you do not already follow. It ranks them by reply count and exports the list as a CSV.","note":"Seven pairs went out and one checkable behaviour came in. Delete the second member of every pair and the passage says exactly what it said before, which is nothing that could be tested against the product."}],"detection":{"type":"statistical","metric":"coordinated-noun-or-adjective-pairs-per-1000-words","direction":"above","threshold_basis":"No threshold ships. The 1.9 times ratio is a body-level comparison between corpora, and legal and ceremonial registers sit far above any line that could be drawn from it. The count is a FeedSquad review trigger for a reread, not a measurement with a published human baseline behind it."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Measured at 1.9 times the human rate in 2024-era model output and carried by vale-ai-tells as a checkable rule. No published work reports the habit fading."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Legal drafting runs on doublets, and aid and abet, terms and conditions, null and void are fixed pairs that a lawyer cannot unpick without changing what the clause does. Liturgy, ceremonial writing and older English pair by convention. The check is deletion: remove one member and see whether any meaning goes with it. If meaning is lost, the writer is using a term of art. If nothing is lost, the pair is decoration, and a page full of decoration is the signal.","model_attribution":"Measured in GPT-4o output. No vendor documentation names the habit, and no study attributes it to a specific model family.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Reinhart et al., Do LLMs write like humans? PNAS 122(8) (arXiv:2410.16107)","url":"https://arxiv.org/abs/2410.16107","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"sentence-cadence-uniformity","name":"Uniform sentence cadence","aka":["isocolon cadence","metronome prose","sentence-length CV"],"category":"structural","subcategory":"cadence","description":"Sentence lengths sit in a narrow band and the prose keeps time. This entry reports a coefficient of variation over sentence word counts, calls it that, and stops. The widely repeated claim that human writing scores between 0.65 and 0.85 on burstiness is arithmetically impossible under the Goh and Barabasi formula it cites. Solving that formula for those values needs a sentence-length standard deviation between 4.7 and 12.3 times the mean; at a mean of 18 words that is a standard deviation of 85 to 222 words. Real English prose runs a coefficient of variation of roughly 0.4 to 0.8, which gives a negative value under the same formula. The circulated numbers are coefficients of variation wearing the wrong name.","why_it_reads_ai":"Even cadence is what comes out when nothing in the draft pushes back. Human paragraphs tend to carry one sentence that ran long because the thought did, and one that stopped early because the writer lost patience with it. The measurement is worth reporting and worthless as a verdict, which is why this entry ships contested and with no number attached.","examples":[{"before":"The team reviewed the numbers on Monday. The results were better than the previous quarter. The marketing spend stayed flat across every channel. The new pricing page brought in more signups than before.","after":"The team read the numbers on Monday. Signups were up 14 percent on the previous quarter with marketing spend flat, and almost all of the lift traced back to one change: the pricing page now shows the annual price first. Nobody had expected that.","note":"Four sentences of near-identical length became three of very different lengths, and the repair adds the figure and the cause. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"sentence-length-coefficient-of-variation","direction":"below","threshold_basis":"No threshold ships, and that is the finding. No published corpus baseline for human sentence-length coefficient of variation exists to anchor one. Munoz-Ortiz and colleagues report that human texts exhibit more scattered sentence length distributions than the output of six models, and publish no mean, no standard deviation and no coefficient for either side. Any figure quoted as burstiness in the range 0.65 to 0.85 is incoherent under the formula it cites, so this entry publishes the arithmetic instead of a cutoff. Report the coefficient, label it a coefficient, and use a low value as a prompt to read the draft aloud. That is a FeedSquad review trigger."},"severity":"medium","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Ships contested. The direction of effect has published support, no human baseline exists to set a line, and the number most often attached to this measurement is arithmetically impossible under the formula it is credited to. Detector work also shows structural features of this kind carrying corpus artifacts rather than authorship."}],"evidence_grade":"corroborated","false_positive_notes":"Plain-language mandates produce short even sentences on purpose, and any writer working to a government or medical readability standard will land in a narrow band by policy rather than by habit. Second-language writers drawing on a narrow set of lexical bundles produce the same shape, and so does anyone composing on a phone. A low coefficient is a fact about the sentences and says nothing about who assembled them. Read the draft aloud before drawing any conclusion from the number.","model_attribution":"No study reports a per-model coefficient of variation. Published work on stylistic homogeneity compares populations of texts rather than single documents, and the homogeneity finding does not convert into a per-document statistic.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"El Attar et al., lexical richness robustness across 27 models (arXiv:2606.04177)","url":"https://arxiv.org/abs/2606.04177","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Pudasaini et al., Why AI-Generated Text Detection Fails: Evidence from Explainable AI Beyond Benchmark Accuracy (arXiv:2603.23146)","url":"https://arxiv.org/abs/2603.23146","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Munoz-Ortiz, Gomez-Rodriguez, Vilares: Contrasting Linguistic Patterns in Human and LLM-Generated News Text (arXiv:2308.09067)","url":"https://arxiv.org/abs/2308.09067","accessed":"2026-08-15","tier":"peer-reviewed"},{"kind":"external","title":"Russell, Karpinska, Iyyer: People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (ACL 2025)","url":"https://aclanthology.org/2025.acl-long.267/","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"sentence-opener-repetition","name":"Sentence-opener repetition","aka":["low start entropy","The and This openers","transition repetition"],"category":"structural","subcategory":"openers","description":"Consecutive sentences starting from the same small set of words, usually a determiner or one of a handful of connectives. vale-ai-tells implements this as two separate checks, one measuring start entropy and one catching a repeated opener. The anti-slop prompt names the same habit from the other side, as something to suppress, which makes it inverted evidence; it is marked as such in the sources and it does not stand alone here.","why_it_reads_ai":"Openers are where a writer's rhythm shows. Four sentences in a row beginning with the same determiner means the draft was produced and not read back. The signal is weak on its own, because the same shape falls out of a plain declarative style that some people write on purpose. Anaphora abuse is the deliberate version, three or more clauses built to match; this entry is the accidental one, measured over function words.","examples":[{"before":"The report landed on Friday. The numbers were lower than expected. The team met on Monday. The plan is being revised.","after":"The report landed on Friday, and signups came in 12 percent under forecast. On Monday the team cut the paid channel and moved that budget to the newsletter, which had been the cheapest source all year.","note":"Four sentences opening on the same word became two that open differently, and the repair names the number and the decision. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"share-of-consecutive-sentences-sharing-an-opening-word","direction":"above","threshold_basis":"No threshold ships. No published baseline exists for opener repetition in human prose, and detector work shows that structural features of this family carried corpus artifacts rather than authorship. Report the share and the longest run, and treat a run of four as a FeedSquad reading cue that sends the draft back for a pass. It is not a measurement anyone has validated."},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Ships contested. Two community rule sets carry the pattern and one of them exists to suppress it, but no study measures it, and the shape is produced deliberately by writers with a plain declarative style."}],"evidence_grade":"community-observed","false_positive_notes":"Writers with a deliberately plain declarative style produce this shape on purpose, and it is standard in instructional and safety writing where every sentence names the same subject for a reason. University teaching centres list consistent sentence structure among the traits that get neurodivergent students falsely accused, and nobody has published a rate for that, so the measure has to stay a reading cue. Check whether the repeated opener refers to the same thing each time; if it does, the repetition is doing reference work and the prose is clearer for it.","model_attribution":"No study reports opener repetition by model family. The habit is named in community rule sets and in one prompt written to suppress it, which dates the concern rather than measuring the behaviour.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"anti-slop-writing system prompt and pattern list","url":"https://github.com/adenaufal/anti-slop-writing","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"paragraph-fragmentation","name":"Paragraph over-fragmentation","aka":["one-sentence paragraphs everywhere","fragmented body copy"],"category":"structural","subcategory":"paragraph-shape","description":"A body built almost entirely from paragraphs of one or two sentences. Pudasaini and colleagues took apart what detector features were actually keying on and found paragraph count dominating feature importance on one corpus and near-absent on another, which means those models had learned the corpus rather than the writer. On the corpus where it dominated, the association the classifier had learned ran the opposite way from this entry: single-paragraph texts read as machine-made, at a median paragraph count of 1 for the machine text it caught and 17 for the human text it cleared. The human text it flagged wrongly had a median of 1 as well. Paragraph shape was a corpus artifact there before it was a signal. That is why this entry ships without a number.","why_it_reads_ai":"The shape comes from a model writing for a scroll, and from a person pasting model output into a post box without joining anything back up. It also comes from every mobile-first style guide written since 2015. Alone it means very little. It earns its place because it travels with the other structural tells and because it makes a piece look longer than the thinking behind it.","examples":[{"before":"We shipped the new onboarding last week. It went well.\n\nThe team learned a lot from the process. Everyone contributed.\n\nThere is more to do here. We will keep going.","after":"We shipped the new onboarding last week: five screens became two, and email verification now runs in the background instead of blocking the first login. Completion went from 61 to 78 percent over the following six days, the first movement in that number since February. The import step is next, and it still loses about a third of the people who reach it.","note":"Six short paragraphs of nothing became one that carries two figures, a date and the next problem. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"share-of-paragraphs-under-three-sentences","direction":"above","threshold_basis":"No threshold ships. The error-analysis section of the Pudasaini paper found detectors keying on paragraph shape had learned the corpus they were trained on, with paragraph count carrying the largest gap in feature contribution between the two error types. The human text that analysis flagged wrongly also ran short, averaging 221 words against 421 for the human text the same model cleared. A cutoff here would inherit both failures. Report the share as context beside the other structural measures, as a FeedSquad reading cue and nothing more."},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Ships contested on the strength of the published error analysis, which shows this exact feature behaving as a corpus artifact. Paragraph count carried the largest gap in feature contribution between the classifier's two error types, and the human writing it misfired on had the same paragraph shape as the machine text it caught."}],"evidence_grade":"community-observed","false_positive_notes":"Mobile-first editorial style guides mandate this shape, and newsrooms writing for phones have followed them for a decade. The broetry format on LinkedIn is a documented human invention that predates chat assistants by about five years, and it is written that way on purpose by people who know exactly what they are doing. Published error analysis shows detectors that keyed on paragraph shape were learning their training corpus, and the human writing they flagged wrongly had the same paragraph shape as the machine text they caught, so a hit here is a reading cue and never a conclusion.","model_attribution":"No study attributes paragraph shape to a model family. Chat surfaces shape it more than models do: formatting defaults differ between a chat product and the same model called through an API.","platform_notes":[{"platform":"linkedin","note":"The one-sentence-paragraph post format on LinkedIn predates chat assistants and remains in wide human use. No published LinkedIn policy addresses line breaks or paragraph length."}],"languages":["en"],"sources":[{"kind":"external","title":"Pudasaini et al., Why AI-Generated Text Detection Fails: Evidence from Explainable AI Beyond Benchmark Accuracy (arXiv:2603.23146)","url":"https://arxiv.org/abs/2603.23146","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"anti-slop-writing system prompt and pattern list","url":"https://github.com/adenaufal/anti-slop-writing","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"fractal-summaries","name":"Fractal summaries","aka":["summary at every level","recap stacking"],"category":"structural","subcategory":"recap","description":"Every paragraph closes by summarising itself, every section closes by summarising its paragraphs, and the document closes by summarising the sections. tropes.fyi names the pattern; vale-ai-tells carries the sentence-level markers that usually come with it. No study has measured it, and the evidence grade says so.","why_it_reads_ai":"A recap is cheap to produce and reads as thoroughness. Stacked at three levels it tells the reader the same thing three times at three sizes while the piece never moves. The cost is attention: the reader spends it on repetition instead of on the one paragraph that carried the new fact.","examples":[{"before":"The pilot ran in two offices for six weeks. Helsinki logged fewer support tickets and Tampere logged none. Support tickets fell in both offices.\n\nThe rollout will cover four more offices, six weeks each. The rollout follows the same shape as the pilot. Both the pilot and the rollout point the same way.","after":"The pilot ran in two offices for six weeks. Helsinki logged 40 percent fewer support tickets. Tampere logged none at all, which turned out to be a broken tagging rule rather than a result. The rollout reaches four more offices from 1 September, and the Tampere queue is being audited first.","note":"The recap sentences carried no noun that was not already in the paragraph above them. The repair spends that space on the number and on the thing that went wrong. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Scan the passage at three levels: the last sentence of each paragraph, the last paragraph of each section, and the last section of the document. For each, decide whether it introduces a noun, figure, date, name or commitment that does not already appear in the unit it closes. Count how many of the three levels close on nothing new. Escape hatches: instructional and training material that states a learning objective, standards documents that repeat so a reader can enter anywhere, and any recap carrying a number, a date or a next action, all of which count as new. Return the count of empty levels out of three, and quote the closing sentence you judged at each level."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Named in two independent community rule sets and unmeasured anywhere. Nothing in the vendor record suggests it has been trained out, and the sentence-level markers that accompany it are still shipped as checkable rules."}],"evidence_grade":"community-observed","false_positive_notes":"Textbook and training-material authors write per-section recaps because the pedagogy asks for them, and a course reader without them would be judged incomplete. Standards documents and safety manuals repeat at every level so that a reader entering in the middle still gets the whole instruction. Technical documentation does the same for people who skim. The discriminator is content, not position: a recap that carries a figure, a date or a next action is doing work, and one that renames what was just said is filling space.","model_attribution":"Undocumented by vendor. The habit tracks length targets more than model family: a draft asked for a word count and given a short idea will spend the difference on recaps.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"signposted-conclusion","name":"Signposted conclusion","aka":["the announced ending","wrap-up heading","restated opening"],"category":"structural","subcategory":"conclusion-shape","description":"A closing section that announces its own arrival and then hands back the opening claim unchanged. tropes.fyi lists it, vale-ai-tells checks the heading forms it arrives under, and Wikipedia's signs-of-AI-writing page treats the announced wrap-up as a marker. This entry judges what the closing section contains. The phrases that announce an ending belong to the canned-conclusion-phrase entry, and the specimen below deliberately carries none of them.","why_it_reads_ai":"An ending has to be paid for, with a consequence, a decision or an admission. The announced version pays with restatement. The test is whether a reader who skips the last section loses anything, and in a low-effort draft the answer is no, because the section was generated from the piece rather than from any further thinking.","examples":[{"before":"The review process needed changing, and everything above explains why it needed changing. The team should take a look at it. Change is worth the effort when a process is holding people back.","after":"From 1 September two named reviewers sign off every pull request. If the median review time is still over a day at the end of October, we go back to the rota and I write up why this failed.","note":"The first version returns the opening claim. The second names a date, an owner, a number and the condition under which the decision gets reversed. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Read the last section, or the last two paragraphs if the piece has no sections. List every noun, figure, date, name and commitment in it. Check each against the opening two paragraphs. Return the count of items in the closing that do not appear in the opening; zero means the ending restates. Escape hatches: abstracts, executive summaries, legal briefs and scientific discussion sections, where a restatement move is required by the genre and judged on how well it is done, and any closing that names a next action, a date or a person. Report the count and quote the one item you judged new, or say there was none."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Carried by three independent taxonomies, including the largest maintained public one, with no decay note on any of them. The heading forms remain shipped as checkable rules."}],"evidence_grade":"corroborated","false_positive_notes":"The five-paragraph essay teaches this shape explicitly, so student work and anything written by someone recently taught that way will produce it under time pressure. Abstracts, executive summaries, legal briefs and scientific discussion sections require a restatement move by genre, and a good one is judged on how economically it does the job. The check is whether the ending carries a decision, a date or a number the opening did not; genre-required restatement usually does, and padding does not.","model_attribution":"Undocumented by vendor. The shape is what a model produces when asked for a complete piece, because a complete piece in most training material has a conclusion.","platform_notes":[{"platform":"wikipedia","note":"The signs-of-AI-writing page treats the announced wrap-up as a marker and pairs it with the observation that encyclopedic articles rarely need a conclusion section at all."}],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"cataphoric-forecasting","name":"Cataphoric forecasting","aka":["three things stand out","numbered lead-in","announced count"],"category":"structural","subcategory":"signposting","description":"The text announces how many points are coming before it makes any of them. vale-ai-tells implements the numbered lead-in as a rule with thirteen surface forms, and tropes.fyi carries the analogue. The count is usually accurate and almost always unnecessary, because on a page the items are about to appear where the reader can see them.","why_it_reads_ai":"Announcing a count buys nothing in writing. A reader can see how many items follow. In speech the move does real work, because a listener cannot scroll back, and it reaches text through material that was spoken first or written to sound spoken. Severity stays low for exactly that reason: the habit is ordinary, and it only reads as low effort when the list is already visible below it.","examples":[{"before":"There are three key reasons the migration slipped, and each one deserves its own paragraph. The first comes down to planning, or rather the lack of it. The second is about communication between the teams involved. The third is the one nobody wants to talk about, and it is probably the most important of them.","after":"The migration slipped twice. The vendor export endpoint rate-limited at 200 rows a minute, and the staging disk filled on 3 March while nobody was watching it.","note":"The count is announced, the three items arrive, and all three are empty. The repair drops the announcement and names the two causes. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:(?:There|Here|Below) are (?:the )?(?:two|three|four|five|six|seven) (?:\\w+ ){0,2}(?:reasons|ways|things|factors|points|steps|lessons|takeaways|pillars|principles)|(?:Two|Three|Four|Five|Six|Seven) (?:\\w+ ){0,2}(?:things|factors|reasons|forces|patterns) (?:stand out|matter here|explain|drive this|define))\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Implemented as a live rule in one community rule set and listed in another, with no report of the habit fading. Severity is low because the move is ordinary in speech and in anything transcribed from it."}],"evidence_grade":"community-observed","false_positive_notes":"Presentation trainers teach this move as the standard way to open a section, and preachers, lecturers and anyone speaking to a room that cannot scroll back use it because a listener needs the count in advance. Conference abstracts and structured reports announce their own contributions by convention. Treat a hit in a transcript, a talk script or a slide deck as expected noise. On a page where the list is already visible below the sentence, the count is doing no work at all.","model_attribution":"Undocumented by vendor. The pattern is heavy in chat-surface output, where a reply cannot rely on the reader seeing the whole answer at once, and it survives into pasted text that no longer has that problem.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"structure-announcement","name":"Structure announcement","aka":["Let's break this down","Let's explore","without further ado","let's dive in"],"category":"structural","subcategory":"metacommentary","description":"The text narrates what it is about to do instead of doing it. vale-ai-tells carries the explainer invitations as one rule and the opening cliches as another, tropes.fyi lists the same move, and the is-this-ai-slop list repeats it. This entry owns the invitation forms. The noun deep dive belongs to the shipped excess-vocabulary cluster, and the podcast opener belongs to its own platform entry.","why_it_reads_ai":"The line is spent on stage directions. In a chat window the move has a function, because a reply that starts working immediately can read as curt, and assistants are tuned to sound cooperative. Pasted into a post, the stage direction is the only part of that conversation that survives, and it now sits above a paragraph the reader can already see.","examples":[{"before":"Let's break this down. First we will look at what changed about the pricing, then at why it changed, and then at what it means for you going forward. There is a lot to cover here, so it helps to take it one piece at a time. By the end you should have a much clearer picture of the whole thing.","after":"Pricing changed on 1 June. The starter plan went from 19 to 24 euros a month, existing customers keep the old rate until renewal, and the overage rate did not move.","note":"Sixty-two words of stage direction, and the piece has still not started. The repair is the thing the announcement was announcing. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"(?:^|[.!?][\"'’”)]?\\s+)(?:Let(?:'|’)?s\\s+(?:break\\s+(?:this|it|that)\\s+down|dive\\s+into|dive\\s+in|unpack|dig\\s+into|walk\\s+through|explore)|Let\\s+me\\s+(?:break\\s+(?:this|it|that)\\s+down|unpack|walk\\s+through|explain\\s+this)|Without\\s+further\\s+ado)","flags":"gm","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Carried by three independent lists and implemented as a live rule in one of them. Vendor prompts suppress over-formatting and several adjacent habits, but no published instruction names these invitations, so nothing suggests the pattern is fading."}],"evidence_grade":"corroborated","false_positive_notes":"Video explainers, classroom teaching and conference talks use spoken signposting because a listener cannot scroll back and needs the map said out loud. A transcript of good teaching will fire this rule at every section boundary, and that is expected noise rather than a finding about the teacher. The check is the medium: on a page where the next paragraph is already visible, the announcement replaces that paragraph instead of introducing it. The rule is anchored to sentence-initial position for the same reason, so the move used mid-sentence inside reported speech passes untouched.","model_attribution":"Heaviest in chat-surface output across assistants. No vendor documents suppressing these specific invitations, which is why the entry treats them as a register habit rather than as a product signature.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"is-this-ai-slop word and phrase list","url":"https://github.com/didrod205/is-this-ai-slop","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"restatement-intro","name":"Restatement intro","aka":["intro paraphrases the title","newsletter restatement opener","takeaways block that restates the title"],"category":"structural","subcategory":"opener","description":"The first paragraph says the headline again in longer words, or a takeaways block above the article does it in bullets. Google's helpful-content guidance asks writers whether they are mainly summarising rather than adding value, and an opening that paraphrases its own headline is the smallest version of that failure. This entry absorbs the three-bullet takeaways block that sits above an article and repeats it. The formatting of such a block belongs to the bold-lead-list-items entry; the emptiness of it belongs here.","why_it_reads_ai":"The opening paragraph is the most valuable space on the page and the easiest to fill without thinking. A model handed a title will write about the title, because the title is what it was given. The reader arrives already knowing the title, so the paragraph is spent before it starts.","examples":[{"before":"Headline: why our onboarding email sequence stopped working.\n\nOnboarding email sequences can stop working, and understanding why they stop working matters for any team that depends on them. This piece looks at the reasons an onboarding sequence stops working.","after":"Headline: why our onboarding email sequence stopped working.\n\nOpen rates on the day-two email fell from 38 percent to 11 percent between March and June. Nothing in the sequence changed. What changed was that our largest customer segment moved to a mail client that files bulk senders into a separate tab, and we did not notice for eleven weeks.","note":"The repair keeps the headline and spends the opening on the number, the cause and the delay. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Compare the headline or subject line with the first paragraph, and with any takeaways block placed above the body. Extract the content words from the headline. Then check whether the opening introduces at least one figure, date, name, place or claim that the headline does not already contain. Escape hatches: a deck or standfirst that a house style requires, a newsletter that repeats its subject line so a forwarded copy still makes sense, and search-trained repetition of the target phrase, which is a real published instruction and only counts as a hit when the paragraph adds nothing else. Return the new item you found, quoted, or the verdict that the opening is a paraphrase."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Carried as a checkable rule in one community rule set, and consistent with published search guidance on summarising without adding value. Ships active because the shape is produced by the workflow rather than by any single model habit, and nothing in the record suggests the workflow has changed."}],"evidence_grade":"community-observed","false_positive_notes":"Search-trained writers are instructed to place the target phrase in the first paragraph, and that instruction predates chat assistants by well over a decade. House styles at magazines require a standfirst that restates the headline for readers arriving from a contents page. Newsletter formats repeat the subject line at the top of the body so that a forwarded copy still makes sense on its own. The check is whether anything else is in the paragraph; the phrase may repeat as long as a fact arrives with it.","model_attribution":"Undocumented by vendor. The pattern follows the prompt rather than the model: a generator handed a title and a word count has nothing else to put in the first paragraph.","platform_notes":[{"platform":"google","note":"The helpful-content guidance asks whether a page mainly summarises what others have said without adding much value. Nothing in it addresses how the text was produced."}],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Google, creating helpful, reliable, people-first content","url":"https://developers.google.com/search/docs/fundamentals/creating-helpful-content","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"feedsquad-observed","title":"FeedSquad newsletter and takeaways-block openers, read during the AI Tells Index build","observed":"2026-08-15","corpus":"Newsletter drafts and article openers produced by FeedSquad publishing agents for feedsquad.com, read by hand on 14 and 15 August 2026 during the AI Tells Index build. The reading was qualitative, so no count and no prevalence figure is published."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"empty-preamble-padding","name":"Empty preamble padding","aka":["before we dive in","why this matters preamble","filler intro"],"category":"structural","subcategory":"padding","description":"Throat-clearing paragraphs sitting where the content should start. Google's quality rater guidelines carry a section on filler as a poor user experience, at 5.2.2, describing low-effort content that occupies prominent space without providing value and naming placement ahead of the main content as what makes it worse. vale-ai-tells implements the opening forms; tropes.fyi lists the same move.","why_it_reads_ai":"The opening is where a piece proves it has something. Filler there is the cheapest thing a generator makes and the most expensive thing a reader can skip, because nobody can know it is filler until they have read it. A page that spends 150 words arriving has already told the reader what the rest will be like.","examples":[{"before":"Before we get into the details, some context helps. Communication inside a growing company is something many teams think about. Getting it right changes how work gets done. With that in mind, here is what we found.","after":"Two of our four teams stopped using the shared channel in February. Both said the same thing in the retro: the channel had turned into an announcement feed, so nobody read it. We moved announcements to a weekly digest on 3 March and replies came back within a fortnight.","note":"Forty words of arrival became a date, a cause and a result. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Read from the start down to the first paragraph carrying a figure, a date, a name, a place or a specific checkable claim. Count the words before that point. Then check whether anything in those words is used again later in the piece. Escape hatch: the scene-setting opener in feature journalism, which plants a person, a place or a moment and comes back to it; if the opening detail recurs, it is a lede and not filler. Return the word count before the first specific, quote the first specific, and say whether the opening material recurs."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Named as a quality problem in current first-party search rater guidance and carried by two independent community taxonomies. Nothing in the record suggests the habit is weakening."}],"evidence_grade":"corroborated","false_positive_notes":"Feature journalists open on a scene rather than on a fact, and the first two hundred words of a well-made profile can pass before the subject is even named. Academic writing opens on background because the field expects the literature first. The discriminator is recurrence: a lede plants a person, a place or a moment that the piece returns to, and filler plants nothing that comes back. Read the whole piece before judging its opening, because the two look identical for the first paragraph.","model_attribution":"Undocumented by vendor as a suppressed habit. Assistants tuned to be helpful open by orienting the reader, which is reasonable in a conversation and dead weight on a page.","platform_notes":[{"platform":"google","note":"Section 5.2.2 of the quality rater guidelines names filler as a poor user experience and calls out prominent placement ahead of the main content. The guidelines are authorship-indifferent throughout: effort, originality and added value decide the rating."}],"languages":["en"],"sources":[{"kind":"external","title":"Google Search Quality Rater Guidelines, 11 Sep 2025","url":"https://static.googleusercontent.com/media/guidelines.raterhub.com/en//searchqualityevaluatorguidelines.pdf","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"one-point-dilution","name":"One-point dilution","aka":["content duplication","the same claim restated"],"category":"structural","subcategory":"repetition","description":"One claim, restated across several paragraphs in different clothes. tropes.fyi lists content duplication and one-point dilution as two patterns; they fire on the same text, so they are one entry here. vale-ai-tells carries a document-level duplication check. Nobody has measured the pattern, and the evidence grade says so.","why_it_reads_ai":"Length is the easiest thing to add to a draft and the hardest thing to add honestly. A generator asked for 800 words on a 200-word idea will produce 800 words. Severity is high because the reader pays four times for one thing, and because the padding hides how thin the underlying material was.","examples":[{"before":"Remote work needs clear written communication. When teams are distributed, writing things down is how information travels. Without written records, distributed teams lose track of what was decided. Teams that do not write things down end up having the same conversation twice.","after":"Our decisions used to live in calls, so anyone who missed the call missed the decision. In February we started writing a four-line note after every meeting: what we decided, who owns it, when it lands, what we rejected. The rejected line turned out to matter most, because the same two ideas kept coming back every six weeks.","note":"One claim in four coats became one claim with a mechanism, a date and a surprise. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"List the distinct claims in the passage, counting a claim as distinct only when it introduces a noun, a figure or an actor that the others do not. Divide the word count by the number of distinct claims. Then reorder the paragraphs and reread: check whether any pronoun loses its referent or any comparison loses its first term. Escape hatch: speech transcripts and teaching texts, where restatement is the medium doing its job, and safety or compliance material written so a reader can enter at any point. Return the claim count, the words per claim, and whether reordering broke anything."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Carried by two community rule sets as separate patterns that fire on the same text, and consistent with first-party search guidance treating low-value repetition as a quality problem. No measurement exists, so the grade stays at community-observed while severity reflects what the pattern costs a reader."}],"evidence_grade":"community-observed","false_positive_notes":"Speech transcripts repeat by design, because a listener cannot reread and oral rhetoric was built on restatement long before any of this. Safety and compliance material repeats deliberately so a reader entering at any point still gets the whole instruction. Beginner instructional writing does the same, and so does anything written to be read aloud. The check is the medium and the ordering: if the paragraphs can be shuffled with no reference breaking anywhere, the piece is one point wearing four coats.","model_attribution":"Undocumented by vendor. The pattern follows the word count in the prompt rather than the model, which is why it survives every change of model family.","platform_notes":[{"platform":"google","note":"Search spam policy covers scaled content abuse no matter how the content was created, in Google own wording. The policy is about value to the reader, not about tooling."}],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"listicle-in-prose","name":"Listicle in a trench coat","aka":["prose that is secretly a list","bulleted thinking in paragraph form"],"category":"structural","subcategory":"list-shape","description":"Paragraphs that are list items with connective tissue pasted between them. tropes.fyi names the pattern and is the only taxonomy in our harvest that carries it, so this entry rests on a single source and the grade reflects that rather than hiding it.","why_it_reads_ai":"An argument has an order. A list does not. When the paragraphs of a piece can be shuffled without anything breaking, the piece was a list before it was prose, and the transitions between paragraphs are doing decorative work. The reader is asked to follow a line of thought that was never drawn.","examples":[{"before":"One area worth attention is pricing. Pricing sets expectations before anyone speaks to sales. Another area is onboarding. Onboarding decides whether the first week goes well. A further consideration is support, which shapes renewal.","after":"Pricing sets expectations before anyone speaks to sales, which is why our onboarding email now repeats the plan limits in its first line. People who read that line open half as many support tickets in week one, and they renew at a higher rate than the ones who skip it.","note":"Three interchangeable paragraphs became one chain where each clause depends on the one before it. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Number the paragraphs, reverse their order and reread. If no pronoun loses its referent, no comparison loses its first term, and no paragraph now precedes what it depends on, the piece is a list. Then read the transitions: a transition that says only that another item follows is decoration. Escape hatch: service journalism, reference entries, glossaries and specification documents, which are legitimately list-shaped and usually announce themselves as such. Return whether reversal broke anything, quote one transition, and say whether the piece presents itself as an argument."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Named in one taxonomy and unmeasured anywhere. Ships active because the shape is common in currently published material, and low severity because plenty of good writing is legitimately list-shaped."}],"evidence_grade":"community-observed","false_positive_notes":"Service journalism is legitimately list-shaped, and a piece about ten places to eat has no argument to carry between them. Reference writing, glossaries and specification documents are the same, and shuffling their sections breaks nothing because nothing was meant to depend on order. The discriminator is what the piece claims to be: a list that announces itself is doing what it says, and an essay that turns out to be a shuffled list is not.","model_attribution":"Undocumented by vendor. The shape follows from prompts asking for an article about a topic, where the model has items and no thesis to order them with.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"elegant-variation","name":"Elegant variation","aka":["synonym cycling","referent churn","WP:AIELEVAR"],"category":"structural","subcategory":"synonym-cycling","description":"One referent, many names, rotated so that no word appears twice. Fowler named the habit elegant variation long before any of this, and Wikipedia adopted his term for its own signs-of-AI-writing page under the shortcut WP:AIELEVAR. The mechanism is a decoding property rather than a fact about any product: a repetition penalty lowers the probability of a token that has already appeared, so the next mention comes back as a synonym. Teaching-side accounts describe the same habit arriving in student work.","why_it_reads_ai":"The reader loses the referent. Once the audit becomes the review, then the assessment, then the study, a reader has to decide at every step whether a new thing has entered the paragraph. Repeating the noun is clearer and reads as less accomplished, which is exactly why the habit survives contact with an editor.","examples":[{"before":"The 2024 audit found three gaps. The review also noted a delay in reporting. The assessment recommended a follow-up, and the study is now with the board.","after":"The 2024 audit found three gaps and a two-week reporting delay. The same audit recommended a follow-up, and the audit sits with the board until 30 September.","note":"The repair repeats the noun on purpose. Four names for one document became one name plus the dates. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Pick the main referent of the passage. List every noun phrase used to refer to it. For each alternative after the first, decide whether it carries a distinction the first one does not, such as a change of scope, of time or of authorship. Count the alternatives carrying no distinction. Escape hatches: writers taught to avoid repetition, which includes much second-language prose and Italian-schooled writing specifically, and any text where the alternatives are established terms of art with different meanings. Return the referent, the list of names used for it, and the count that carry no distinction."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Carried by the largest maintained public taxonomy under its own shortcut, by a pattern directory and by a teaching-press account, with a stated decoding mechanism behind it. No published work reports it fading."}],"evidence_grade":"corroborated","false_positive_notes":"Writers schooled to avoid repetition produce this on purpose, and Wikipedia's own page records that Italian schools teach the habit explicitly, citing Italian-language sources for it. Style guides in several languages ask for the same thing. Work separating model families by text statistics operates over many documents at aggregate level and gives nobody a licence to call a single paragraph. Where the alternative names carry a real distinction, the variation is doing work; where they are interchangeable, it is costing the reader the referent.","model_attribution":"Attributed to repetition-penalty decoding, which is a property of how text is sampled rather than of any one vendor. Aggregate classifiers can separate model families from text alone, which is a population result and not a per-document one.","platform_notes":[{"platform":"wikipedia","note":"Shortcut WP:AIELEVAR. The page carries its own caveat that editors who are not native English speakers may avoid repeated words as a matter of schooling, and cites Italian-language sources for it."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Chronicle of Higher Education, ten ways AI is ruining your students writing","url":"https://www.chronicle.com/article/10-ways-ai-is-ruining-your-students-writing","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Sun et al., Idiosyncrasies in Large Language Models, ICML 2025 (arXiv:2502.12150)","url":"https://arxiv.org/abs/2502.12150","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"broken-enumeration","name":"Broken enumeration","aka":["every item number one","numbering that fails"],"category":"structural","subcategory":"list-integrity","description":"A numbered list whose numbering repeats, resets, or contradicts the count the text promised. The Sports Illustrated articles that Futurism reported in November 2023, and that CNN covered when the site pulled them, included a personal-finance piece whose numbered list labelled every item number one. The specimen below is written for this entry. The published text is not reproduced here.","why_it_reads_ai":"Numbering is the one part of a list a reader trusts without checking it. When it breaks, the piece was assembled rather than written, and nobody read it back before it went out. Severity is high because this is not a matter of taste: a proofreader catches it in a single pass, so its presence says something about the pipeline rather than about the prose.","examples":[{"before":"1. Check the chain with a wear gauge.\n1. Replace the brake pads.\n1. Re-tape the bars.","after":"1. Check the chain with a wear gauge. A stretched chain wears the cassette, which costs more to replace than the chain does.\n2. Replace the brake pads once the wear grooves have gone.\n3. Re-tape the bars, because old cork slips in the wet.","note":"The repair fixes the numbering and gives each item the reason it exists. Numbering that a renderer generates has to be checked in the rendered output, not in the source."}],"detection":{"type":"deterministic","pattern":"^[ \\t]*1[.)][ \\t]+\\S[\\s\\S]{0,400}?^[ \\t]*1[.)][ \\t]+\\S","flags":"gm","scope":"document"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Documented in the November 2023 reporting on fabricated bylines, where a numbered list labelled every item number one, and still turning up in published material. Nothing in the record shows it fading, because it is a review failure rather than a writing habit."}],"evidence_grade":"corroborated","false_positive_notes":"Markdown authors write every item as 1. on purpose, because the renderer numbers the list for them, so this rule has to run against rendered text and never against source. Content management systems mangle list markup on import routinely, and pasting between editors can renumber a list with nobody touching the words. Legal drafting numbers clauses as 1.1 and 1.2 under a single 1, which this pattern leaves alone by design. Check the source markup and the publishing path before treating a hit as anything at all about the writing.","model_attribution":"No vendor documents the behaviour. The failure is usually in the pipeline rather than in the model: markup survives generation and does not survive the paste into a publishing system.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Futurism, Sports Illustrated published articles by fake AI-generated writers","url":"https://futurism.com/sports-illustrated-ai-generated-writers","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"CNN, Sports Illustrated deletes articles with fake author names","url":"https://www.cnn.com/2023/11/27/media/sports-illustrated-deletes-articles-fake-author-names-ai-profile-photos/index.html","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"essay-scaffold-in-social-post","name":"Essay scaffolding in a social post","aka":["five-paragraph post","intro-body-conclusion in 150 words"],"category":"structural","subcategory":"genre-transfer","description":"A 120-word post carrying a general opening, a middle of parallel points and a closing that returns to the opening. The architecture is correct and the container is wrong. No external taxonomy in our harvest carries this pattern, so it rests on our own reading of FeedSquad agent drafts and takes the lowest evidence grade in the set. The corpus and the window are named in the sources so the claim can be checked or refused.","why_it_reads_ai":"The essay shape is what a generator has most of, and a short prompt does not displace it. On a feed the scaffolding eats the post: the opening spends the first two lines, the closing spends the last, and the reader gets three sentences of content inside a format built for three thousand words. The reader sees the shape before the sentences and stops.","examples":[{"before":"Hiring is one of the harder parts of running a small company. There are a few reasons for this. The market is competitive. A small team cannot absorb a bad hire. Interviews predict less than people think. For these reasons, hiring deserves careful thought.","after":"We made two hires last year and one of them left after six weeks. The one who stayed had done the take-home in 40 minutes and then sent us a page on why the brief was wrong. We now send the brief first and read the objections before the CV.","note":"The scaffold went out and one first-hand account came in, at roughly the same length. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Count the words. Under roughly 300, check for three parts: a general opening sentence that names the topic without carrying a fact, a middle of parallel points, and a closing sentence that returns to the opening. Mark it only when all three are present and the middle carries no name, figure, date or first-hand detail. Escape hatch: a deliberately structured argument post, which will carry at least one specific inside the middle, and any post whose closing names a next action. Return which of the three parts you found, and quote the closing sentence."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Recorded in FeedSquad agent drafts during the index build on 14 and 15 August 2026. No external taxonomy carries the pattern, so the entry ships at the lowest evidence grade with its corpus named, and severity is low because the shape is competence in the register it came from."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"Students and recent graduates write the structure they were taught, and in the register it came from that structure is competence rather than laziness. Writers who learned composition through a formal curriculum, which includes many second-language writers, fall back on it under time pressure because it is the shape that arrives first. School systems outside English teach their own equivalents, so the habit crosses languages. The check is whether the middle carries any first-hand specific: scaffolding with a real detail inside it is a well-made short argument, and scaffolding with nothing inside it is a shape.","model_attribution":"Undocumented by vendor and unmeasured by anyone. The shape follows from what generic writing looks like in training material, and it survives across model families for that reason.","platform_notes":[{"platform":"linkedin","note":"The March 2026 LinkedIn authenticity post addresses automated posting, third-party tools and engagement pods. It says nothing about post architecture or sentence structure, and no published LinkedIn policy does."}],"languages":["en"],"sources":[{"kind":"feedsquad-observed","title":"FeedSquad agent social drafts, read during the AI Tells Index build","observed":"2026-08-15","corpus":"Short-form drafts produced by FeedSquad publishing agents for feedsquad.com, read by hand on 14 and 15 August 2026 during the AI Tells Index build. The reading was qualitative, so no count and no prevalence figure is published, and the pattern is recorded as observed rather than measured."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"staccato-fragment-run","name":"Staccato fragment run","aka":["short punchy fragments","parallel staccato"],"category":"structural","subcategory":"cadence","description":"Runs of verbless fragments used to set a rhythm. tropes.fyi carries them as short punchy fragments, vale-ai-tells implements a parallel-staccato rule with eight surface forms, and slop-lint lists the reply-register version. This entry replaces a cut candidate whose sentence-length figures sit on the never-publish list for this project, and it claims nothing statistical. The pattern below fires only on two adjacent comparative fragments, which is the narrowest form of the habit.","why_it_reads_ai":"The fragments carry the cadence of a conclusion without stating one. Two of them land in a row and the paragraph sounds decided. Read them back and the claim is missing: a comparative with no baseline to compare against, a noun with no verb attached to it. The rhythm does the work the evidence was supposed to do.","examples":[{"before":"We rebuilt the checkout screen and the difference showed up straight away. Faster onboarding. Fewer support tickets. Nobody has asked what either of those is measured against, which is usually a sign that the answer would be awkward.","after":"We rebuilt the checkout screen in March. Median signup time fell from four minutes to ninety seconds, and tickets tagged checkout dropped by half over the following month.","note":"Faster and fewer than what, over what period, measured how. The fragments sound decided and the sentences around them never supply the baseline. The repair states it twice. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"(?:^|[.!?]\\s+)(?:More|Less|Fewer|Faster|Better|Cleaner|Smarter|Simpler|Stronger|Cheaper|Bigger|Higher|Lower|Longer|Shorter|Deeper|Broader|Leaner|Safer)\\s+[a-z]{3,}(?:\\s+[a-z]{3,})?\\.\\s+(?:More|Less|Fewer|Faster|Better|Cleaner|Smarter|Simpler|Stronger|Cheaper|Bigger|Higher|Lower|Longer|Shorter|Deeper|Broader|Leaner|Safer)\\s+[a-z]{3,}(?:\\s+[a-z]{3,})?\\.","flags":"gm","scope":"document"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Carried by three community sources, one of which implements it as a live rule. No measurement exists and none is claimed here. Severity stays low because the form is a working tool in several human registers."}],"evidence_grade":"community-observed","false_positive_notes":"Direct-response copywriters are taught punch-then-develop rhythm and use it deliberately, and sports writing has run on clipped fragments for a century. Poets and lyricists take the fragment as their basic unit. The rule is deliberately narrow and fires only on two adjacent comparative fragments, so most legitimate fragment writing passes it untouched. A hit in advertising copy is the register working as intended, and the question to ask there is whether the comparative ever names what it is being compared with.","model_attribution":"Undocumented by vendor. The form appears across assistants in short-form output and is heavier where a prompt asked for a punchy or engaging register.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"contraction-avoidance","name":"Contraction avoidance","aka":["no contractions anywhere","uniform formal register"],"category":"structural","subcategory":"register","description":"Not one contraction anywhere in a text that otherwise reads like speech. vale-ai-tells implements the check at document level. Wikipedia's signs-of-AI-writing page lists formal prose in itself among the indicators that do not work, and this entry cites that against itself, which is the reason it ships contested rather than active.","why_it_reads_ai":"Readers use contractions as a warmth marker, and that heuristic is the whole problem. The cues people rely on when judging how a text was made are predictable, which makes them easy to satisfy from either direction. What the rate measures is how a text was styled. A zero rate in a register that is otherwise conversational is worth a second read and nothing more.","examples":[{"before":"I am glad you asked about the pricing change. It is not something we do often, and we do not expect it to affect most customers. We will publish the details on Friday, and I will answer questions in the thread.","after":"I'm glad you asked. The starter plan goes from 19 to 24 euros on 1 September, existing customers keep the old rate until renewal, and I'll be in the thread all Friday afternoon.","note":"The contractions matter less than the two numbers and the date that arrived with them. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"contractions-per-1000-words","direction":"below","threshold_basis":"No threshold ships. No published baseline exists for contraction rate by genre, and a rate of zero is normal in academic, legal and government registers by house rule rather than by habit. Report the rate beside the register. A zero rate in text that otherwise reads as speech is a FeedSquad reading cue that sends the draft back for a pass, and it is not a measurement anyone has validated."},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Ships contested. Wikipedia lists formal prose in itself among its ineffective indicators, and readers treat contractions as a warmth marker, which makes the rate a fact about styling rather than about effort."}],"evidence_grade":"community-observed","false_positive_notes":"Academic, legal and government registers forbid contractions outright, and much business English written outside the United States does the same by house style. Second-language writers often avoid them because contractions were taught as informal and risky. Wikipedia's own page lists formal prose per se among the indicators that do not work, which is the strongest argument against this entry and the reason it carries a contested status. Read the register before reading the rate, and never read the rate on its own.","model_attribution":"Vendor prompts suppress several register habits release by release, and none of the published instructions we read names contractions in either direction. The rate follows the prompt and the surface more than the model family.","platform_notes":[{"platform":"wikipedia","note":"The signs-of-AI-writing page keeps a list of ineffective indicators that includes formal or fancy prose in itself, perfect grammar and bland tone. This entry sits on that list and is published contested because of it."}],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"challenges-then-future-outlook-closer","name":"The challenges-then-outlook closer","aka":["Despite these challenges","Future Outlook section","Challenges and Legacy","WP:FACESCHALLENGES","faces several challenges","the future remains promising","outlook section reflex"],"category":"rhetorical","subcategory":"document-shape","description":"A closing move in two beats. First a paragraph conceding difficulty in general terms. Then a paragraph predicting that the difficulty will be overcome. Neither beat names anything specific enough to be wrong. Wikipedia files it as a rigid formula and states the caveat plainly: the sign is the formula, not the mention of challenges. Russell and colleagues, whose study is the page's citation for it, also found that heavy LLM users identify AI text at roughly 90% accuracy, which is the strongest human-detection figure in the literature and still one error in ten.","why_it_reads_ai":"The section exists because the document needed an ending, not because the writer learned anything. It is the document-shaped version of a summary that adds nothing. A writer who knows the field closes on the next decision and who has to make it.","examples":[{"before":"Despite its rapid growth, the sector faces several challenges, including regulatory uncertainty and questions around long-term sustainability. Nevertheless, with continued innovation and collaboration among stakeholders, the future remains promising.","after":"The certification rules change in March. Two of the four vendors we tested have said they will not certify under the new scheme, which leaves anyone on an annual contract with a decision to make before renewal.","note":"The before paragraph would survive being pasted into an article about any sector at all. The after paragraph would not survive being moved one industry sideways."}],"detection":{"type":"judge","rubric":"Read only the final section or the last two paragraphs. Answer yes if all four of the following hold. (1) The passage concedes difficulty, obstacles, limitations or criticism. (2) The concession is generic: no named actor, date, figure, place or document appears inside it. (3) A later sentence reverses toward optimism or continued relevance, often opening with Despite, Nevertheless, However, Yet or Still. (4) The reversal makes no falsifiable prediction: no future event could show it to be wrong. Additionally answer yes when a heading reads Challenges, Future Outlook, Challenges and Opportunities, or Conclusion and the passage beneath it satisfies (2) and (4). Answer no if either beat names an entity, a number, or a dated commitment. Answer no if the passage states what the writer will do next. Quote the sentence that decided the call."},"severity":"high","status":"active","status_history":[{"date":"2026-08-14","status":"active","rationale":"Named by Wikipedia as a current content sign with its own shortcut and its own citation. No source reports it declining, and unlike a lexical marker it survives paraphrasing, because it is a shape rather than a phrase."}],"evidence_grade":"corroborated","false_positive_notes":"Equity analysts and management consultants are paid to end on a balanced outlook, and the risks-then-outlook shape is house style in annual reports and grant applications. Undergraduate essay instruction teaches the same closing move explicitly. Wikipedia's own caveat applies: a closer that names a specific obstacle and a specific date is not this tell, however balanced it sounds. Do not treat the presence of a Challenges heading as sufficient; the rubric requires the vagueness, not the heading.","model_attribution":"Reported for chat assistants broadly. Wikipedia cites Russell and colleagues for the formula. Nothing in the literature separates model families on this pattern, and the sources that name a product at all name the ChatGPT surface.","platform_notes":[{"platform":"wikipedia","note":"Shortcut WP:FACESCHALLENGES. Wikipedia separately reports that an Awards and recognition section is nearly ubiquitous in AI-generated articles, which is the same habit at heading level."},{"platform":"youtube","note":"YouTube's monetization policy names templated storylines with minimal variation across videos. A closer that fits any topic is the prose form of a template, and enforcement there is channel-wide when many videos qualify."},{"platform":"google","note":"Google's spam policy turns on many pages generated with little value to users, no matter how they are created. A closer that transfers between topics is part of what makes those pages cheap to produce at scale."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing, section on outline-like conclusions about challenges and future prospects","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Russell, Karpinska, Iyyer: People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (ACL 2025)","url":"https://aclanthology.org/2025.acl-long.267/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"YouTube Help: Channel monetization policies (inauthentic content)","url":"https://support.google.com/youtube/answer/1311392","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Google Search Central: Spam policies for Google web search (scaled content abuse)","url":"https://developers.google.com/search/docs/essentials/spam-policies","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-14","updated":"2026-08-15"},{"id":"replacive-contrastive-negation","name":"\"It's not X, it's Y\"","aka":["contrastive negation, replacive subtype","epanorthosis","correctio","negative parallelism","antithesis","contrastive reframe","X rather than Y","The danger isn't X. It is Y.","Most people think X. The reality is Y.","Forget X. Focus on Y.","uncontracted contrast","strawman consensus opener","corrective anaphora"],"category":"rhetorical","subcategory":"contrastive-negation","description":"A clause denies one description of the subject and immediately substitutes a larger one, with no connective between the halves. The denied term was never asserted by anyone. Wikipedia's editors file the family under negative parallelisms and give three templates. Linguists separate the substituting form, called replacive contrastive negation, from the additive form built on not only X but also Y. The additive form is catalogued separately, because the two constructions do different work. This entry covers the replacive one only, and only the asyndetic version with no but between the halves. That narrowing is deliberate. A sentence like the report is not the fastest option, but it is the cheapest is ordinary English and stays out of scope.","why_it_reads_ai":"The construction promises a correction and delivers an upgrade. Classical rhetoric names the genuine move correctio: amending a first thought to make it stronger. Here nothing is amended. The rejected term was never claimed, so the sentence performs the shape of insight at no cost. Models produce it because the shape reads as emphasis to a human rater and the second half accepts any abstraction. The Economist reported models reaching for it repeatedly to carry emphasis. One instance means nothing. Three in a short post means the writer had a cadence and went looking for content to fill it.","examples":[{"before":"Onboarding isn't a checklist, it's a relationship. That distinction sounds slight on the page and it changes very nearly everything about how the first conversations go. Companies arrive at this understanding eventually, though usually after losing a handful of accounts they had counted as safe, and starting from the relationship is mostly a way of skipping that part.","after":"Our onboarding is five steps over eleven days. The only one that moves retention is the day-three call with a person. We cut the other four to two and retention held.","note":"The repair is not a synonym swap. The second version names the change and what happened after."},{"before":"This isn't a pricing change, it's a statement about who we build for. Price has always been the clearest thing a company says about itself, whether or not anyone means it that way. The realisation usually arrives some time after the invoice does, which is why the explanation afterwards is so much harder than the decision was.","after":"We raised the entry tier from 12 to 19 dollars a seat and kept the old price for accounts opened before June. Two thirds stayed. The accounts that left were each using one feature.","note":"Figures in this repair are invented for the specimen."},{"before":"They're not customers, they're partners. That distinction shapes everything about how the relationship actually works, and most companies say it long before they have earned it. The good ones let the word do some work. The rest put it on a slide and move on.","after":"Four of our accounts test releases before we ship them, and two have written docs we now publish. We give those four a direct line to the engineer who owns the area."}],"detection":{"type":"deterministic","pattern":"\\b[A-Za-z][\\w'’-]*(?:\\s+[\\w'’-]+){0,3}(?:\\s+(?:is|are|was|were)(?:\\s+not|n['’]t)|['’](?:re|m)\\s+not)\\b(?![^.;!?]*\\bbut\\b)[^.;!?]{0,80}?[,;]\\s*(?:it|they|this|that|these|those)(?:['’]s|['’]re|\\s+(?:is|are|was|were))\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-14","status":"active","rationale":"Opened as active. Named by Wikipedia's editors as a current sign with three templates, reported by The Economist as a repeated model habit in a 2026 rewrite study, and given a precise linguistic taxonomy by Silvennoinen. The human base rate is real and the entry says so, which is why severity is moderate rather than strong."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Writers of myth-busting and common-misconceptions pieces use this form as their organising device, and Wikipedia's own guide says so in the same section that names the tell. Speechwriters and preachers use it deliberately, where it is a real correctio rather than filler. Sports commentary runs on it. Sales copy has used it for a century, and the software joke about a bug and a feature predates the whole discourse. Writers whose first language is German or Spanish have a dedicated corrective conjunction, sondern and sino, and reach for the English equivalent more readily than an English monolingual would. One occurrence is not a finding. The signal is repetition inside one short document, plus a second half that names nothing the first half did not.","model_attribution":"Not tied to a vendor. Wikipedia lists it as a general LLM sign. The Washington Post counted not just X but Y variants in about six percent of shared ChatGPT chats in July 2025, and that figure covers the additive template rather than this one, so it does not transfer. One research pass found no peer-reviewed source assigning the construction to any model family.","platform_notes":[{"platform":"wikipedia","note":"Catalogued under negative parallelisms, shortcut WP:AIPARALLEL, with three named templates. The same section states that human writers use the form commonly, especially in listicles about myths."},{"platform":"linkedin","note":"LinkedIn's May 2026 policy suppresses out-of-network distribution for content with no unique perspective. It names no sentence construction, and neither does any other LinkedIn policy text."},{"platform":"x","note":"X applies an llm_slop_post label that routes to spam handling for thirty days. The classifier prompts are withheld from the public repo, so nothing published shows this or any other construction being scored."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia:Signs of AI writing, section Negative parallelisms (WP:AIPARALLEL)","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Silvennoinen, Not only apples but also oranges: Contrastive negation and register, VARIENG vol. 19 (2017)","url":"https://varieng.helsinki.fi/series/volumes/19/silvennoinen/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Silva Rhetoricae, Epanorthosis (correctio), ed. Gideon O. Burton, Brigham Young University","url":"https://rhetoric.byu.edu/Figures/E/epanorthosis.htm","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"The Economist, How to spot AI writing (2026-07-30)","url":"https://www.economist.com/culture/2026/07/30/how-to-spot-ai-writing","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Reuters Institute, How AI-generated prose diverges from human writing and why it matters","url":"https://reutersinstitute.politics.ox.ac.uk/news/how-ai-generated-prose-diverges-human-writing-and-why-it-matters","accessed":"2026-08-15","tier":"press"}],"added":"2026-08-14","updated":"2026-08-17"},{"id":"rule-of-three-padding","name":"Rule of three used as padding","aka":["tricolon","triadic list","rule of three","power of three","verb tricolon","three-item coordinated series"],"category":"rhetorical","subcategory":"parallelism","description":"Three parallel items in a coordinated series, used as the default shape for every list in a document. Rhetoric calls the figure tricolon and treats it as an emphasis device reserved for a high point. The figure is not the tell. The tell is the figure arriving every few sentences, with a third item that carries no information the first two lacked. The threshold below is a review trigger rather than a measured cutoff. No study publishes a per-thousand-word tricolon rate for human or model prose, so the number comes from arithmetic: at five per thousand words, a 700-word post trips at four tricolons and stays clear at three.","why_it_reads_ai":"A three-item series reads as complete. Two items read as a pair. Four read as an inventory. Models reach for the three-shape when they have one idea and need the cadence of an argument, and Wikipedia's guide ties the habit to making a superficial analysis look more comprehensive. Density does the work here. A reader does not consciously count the series. The reader feels the padding and stops.","examples":[{"before":"The new dashboard is faster, cleaner, and smarter. It helps teams plan, execute, and review. The result is better visibility, better decisions, and better outcomes.","after":"The dashboard now loads in under a second on a cold cache, down from four seconds. That is the whole change. People stopped opening a second tab to check whether the first one had finished.","note":"Nine items across three sentences in the specimen. Not one of them is a number."},{"before":"Good writing is clear, concise, and compelling. It respects the reader, earns attention, and delivers real value. The best writers understand that structure, rhythm and restraint matter far more than vocabulary. Everything else is style, preference and habit. Get those three right and very little else about a sentence turns out to need deciding at all.","after":"Good writing survives a reader skimming on a phone at the end of a workday. We read every draft on a phone before it ships, and anything that needs a second pass gets cut rather than rewritten."}],"detection":{"type":"statistical","metric":"three-item-coordinated-lists-per-1000-words","threshold":5,"direction":"above","threshold_basis":"No published work compares tricolon rates in human and model prose, which is the main reason this entry is contested rather than active. Five three-item coordinated lists per 1,000 words is a density observation set by FeedSquad, not a measured cutoff, and the tricolon is a deliberate device in speechwriting and legal drafting. Never treat a hit here as a signal on its own."},"severity":"low","status":"contested","status_history":[{"date":"2026-08-14","status":"contested","rationale":"Opened as contested. Wikipedia's guide names the pattern under WP:RO3 and The Economist reports models favouring three-item lists, but neither publishes a rate, and a separate research pass found no primary source assigning the figure to any model family. Classical rhetoric has treated the tricolon as a deliberate device since antiquity, so the base rate in edited human prose is high and unmeasured."}],"evidence_grade":"community-observed","false_positive_notes":"The tricolon is one of the oldest devices in rhetoric and speechwriters reach for it deliberately, so its presence is close to meaningless on its own. Legal and policy drafting also produces genuine three-item enumerations because the underlying list has three members. No published work compares human and model tricolon rates, so treat this as a density observation and never as a standalone signal.","model_attribution":"General, not vendor-specific. Wikipedia lists it as a current sign and The Economist observed it across four chat products in a 2026 rewrite exercise. No peer-reviewed source assigns it to a model family, and one research pass classified the family-level version of the claim as folklore.","platform_notes":[{"platform":"wikipedia","note":"Named as a sign under shortcut WP:RO3, with the stated function of making a superficial analysis appear more comprehensive."},{"platform":"google","note":"Google's spam policy is method-agnostic and turns on value to users regardless of how a page was made. Scaled content abuse also requires many pages, so a single padded post is outside the policy entirely."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia:Signs of AI writing, section Rule of three (WP:RO3)","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Silva Rhetoricae, Tricolon, ed. Gideon O. Burton, Brigham Young University","url":"https://rhetoric.byu.edu/Figures/T/tricolon.htm","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"The Economist, How to spot AI writing (2026-07-30)","url":"https://www.economist.com/culture/2026/07/30/how-to-spot-ai-writing","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Heritage & Greatbatch, Generating applause: a study of rhetoric and response at party political conferences, American Journal of Sociology 92(1) 1986","url":"https://doi.org/10.1086/228465","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-14","updated":"2026-08-17"},{"id":"self-answering-question-opener","name":"Rhetorical question that answers itself","aka":["anthypophora","hypophora","subiectio","rogatio","The result? A number.","ask-then-answer opener","self-answered question"],"category":"rhetorical","subcategory":"opener","description":"A section or paragraph opens with a question aimed at the reader, and the next sentence answers it. Rhetoric calls the ask-and-answer move anthypophora and distinguishes it from erotema, the question left hanging. The device is ancient and legitimate. The pattern under review is the version where the answer adds nothing the question did not already contain, and the question exists to manufacture a transition. This is a judge rule because no regex can tell a real setup from a hollow one.","why_it_reads_ai":"The move buys a paragraph break without an argument, and it flatters the reader by pretending they asked. Evidence cuts against treating it as a machine marker. The largest grammatical comparison of LLM and human prose found that models underuse questions relative to matched human writing. So if this is a tell at all, it is about placement and formula rather than frequency. Treat a hit as a reason to read the section. It is not evidence of authorship.","examples":[{"before":"So what does this mean for your pipeline? It means you need to rethink how you qualify leads, because qualification comes down to judgement in the end, and judgement takes a while to build.","after":"Qualifying on company size stopped predicting anything for us once we moved upmarket. Headcount and budget had decoupled. We now qualify on whether the buyer already has someone doing the job by hand.","note":"Numbers and specifics in the repair are invented for the specimen. The point is that the repair carries a fact the question did not."},{"before":"Why does any of this matter? Because attention is the scarcest resource in business today.","after":"We tracked which posts got read to the end. Posts under 120 words finished at roughly twice the rate of posts over 300, and the drop-off was sharpest in the first two lines.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Task: count self-answering question openers in the supplied text. For each paragraph or section that begins with a question, mark it only when all four conditions hold. (1) The opener is a question addressed to the reader. It is not a quoted question, not a survey item, and not a heading on a page whose format is questions and answers. (2) A sentence within the next two sentences answers that question. (3) The answer introduces no number, name, date, place, or source that the question did not already contain. (4) Deleting the question leaves a paragraph that still opens cleanly. Output the count, then quote each marked question verbatim. Mark nothing when the question stays unanswered for more than two paragraphs. Mark nothing in a document that is explicitly a question-and-answer format. If the text has no question openers, output a count of zero and stop."},"severity":"low","status":"contested","status_history":[{"date":"2026-08-14","status":"contested","rationale":"Opened as contested, and the contest is on the record. Reinhart et al. in PNAS found questions among the features LLMs underuse relative to matched human prose, which is direct counter-evidence to a frequency-based version of this tell. Wikipedia's guide does not name it. The entry ships because the formulaic placement is worth reviewing, not because the evidence supports a verdict."}],"evidence_grade":"corroborated","false_positive_notes":"Anthypophora is standard equipment for lecturers and for anyone writing explanatory prose to a reader who cannot interrupt. Technical documentation uses it as a navigation device. Pages whose whole format is questions and answers are built from it by definition, which is why the rubric excludes them by name. Preachers and trial lawyers deploy it deliberately and can say why. Explanatory journalism uses a question header as a house convention to set the reader's expectation for the paragraph below. Teachers writing for students at a lower reading level are taught to do this. Flag a passage only when the answer repeats the question and adds nothing measurable.","model_attribution":"Unattributed. No vendor documentation and no peer-reviewed source names this pattern as a model habit. The nearest peer-reviewed finding points the other way: instruction-tuned models produced fewer questions than humans across matched registers. Any claim that a particular product does this more than a person is unsupported as of the access date.","platform_notes":[{"platform":"linkedin","note":"The question opener is ordinary LinkedIn craft and predates chat products on that surface, so a hit there is weak evidence. LinkedIn's stated target is content that restates without adding, and bulk automated comments."},{"platform":"wikipedia","note":"Not named in the Wikipedia guide. The closest listed pattern is the outline-like conclusion formula about challenges and future prospects, which is a different shape and has its own entry."}],"languages":["en"],"sources":[{"kind":"external","title":"Silva Rhetoricae, Anthypophora, ed. Gideon O. Burton, Brigham Young University","url":"https://rhetoric.byu.edu/Figures/A/anthypophora.htm","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Reinhart et al., Do LLMs write like humans? Variation in grammatical and rhetorical styles, PNAS 122(8) 2025, PMC mirror","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11874169/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Silva Rhetoricae, Erotema (the question left unanswered), ed. Gideon O. Burton, Brigham Young University","url":"https://rhetoric.byu.edu/Figures/E/erotema.htm","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-14","updated":"2026-08-17"},{"id":"additive-contrastive-negation","name":"Not only X but also Y","aka":["additive contrastive negation","correlative coordination","not just X but Y"],"category":"rhetorical","subcategory":"contrastive-negation","description":"Two coordinated members joined by a correlative pair. The first is negated, the second added on top. Silvennoinen separates this additive subtype of contrastive negation from the replacive subtype, which denies one description and substitutes another; that one has its own entry and its own rule. Grammar calls the frame correlative coordination, and it is ordinary taught English. The catalogued version is narrower: both members name the same thing at two levels of abstraction, so the sentence widens without adding a fact. The rule below skips the placeholder form, where the members are a bare X and a bare Y, because that is how the construction gets named rather than used.","why_it_reads_ai":"The frame produces the feeling of a scope claim at no cost. A writer who has measured two effects names both, and the reader can check either one. A writer who has one sentence and needs weight puts the small version first, negates it, and adds the large version after the correlative. Wikipedia files the family under negative parallelisms. Density decides. One in a page is prose. Four in a short post is a cadence looking for content.","examples":[{"before":"Our scheduler is not only a calendar but also a strategist for your entire content operation. Planning ahead is what makes publishing consistent, and consistency is what makes any of the rest of it worth measuring at all. The right cadence looks different for everyone, which is why the tool adapts to the way you already work rather than the other way around.","after":"The scheduler moved 40 posts for us in July when a launch slipped by a week, and it caught the two that would have gone out during a client embargo.","note":"Figures in this repair are invented for the specimen."},{"before":"The report is not merely a summary but a call to action for every team involved.","after":"The report ends with four asks. Each one names the team that owns it and the date it is due, and all four land before the March board meeting."}],"detection":{"type":"deterministic","pattern":"\\bnot\\s+(?:only|just|merely|simply)\\s+(?![XY]\\b)[^.;!?\\n]{1,80}?\\s+but\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active. Wikipedia lists the negative-parallelism family as a current sign, and Silvennoinen gives the construction a precise grammar and separates it from the replacive form. The human base rate is high, which is why severity stays moderate and the entry says the signal lives in repetition."}],"evidence_grade":"corroborated","false_positive_notes":"Academic writers use the correlative frame to make a genuine scope claim, and it is standard taught English with a grammar of its own; Silvennoinen documents it as ordinary register variation rather than a defect. Grant applications and abstracts run on it because the genre rewards stating both a narrow and a broad contribution. Second-language writers taught English through formal writing courses reach for it more often than a monolingual would, since the correlative is explicitly drilled. A reviewer separates the two cases by testing the second member: if it names something the first did not, and something a reader could check, the sentence is doing work.","model_attribution":"Not tied to a vendor. The Washington Post counted not-just-X-but-Y variants in roughly six percent of shared ChatGPT chats in July 2025, a figure our research pass recorded second hand and did not verify at source, so it is reported here as context and not as a rate. No peer-reviewed source assigns the construction to a model family.","platform_notes":[{"platform":"wikipedia","note":"Catalogued under negative parallelisms with the shortcut WP:AIPARALLEL. The same section states that human writers use the form commonly."},{"platform":"linkedin","note":"The frame is house style in professional-network copy and predates chat products there, so a single hit on that surface carries little weight."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Silvennoinen, Contrastive negation in English, VARIENG 19 (2017)","url":"https://varieng.helsinki.fi/series/volumes/19/silvennoinen/","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"triple-denial-reveal","name":"Not X. Not Y. Just Z.","aka":["triple negation reveal","two denials then a reveal"],"category":"rhetorical","subcategory":"staccato-reveal","description":"Three fragments in a row. The first two deny, the third supplies a minimal answer. tropes.fyi lists the shape by name and slop-lint catalogues the staccato reply register it belongs to. The shipped replacive-contrast entry does not reach it: that rule requires a negated copula and a comma, and a fragment triplet carries neither. The boundary is deliberate, so the two rules stay disjoint and a specimen never fires both.","why_it_reads_ai":"Both denials are free, because nobody proposed either term. The third fragment then lands as a revelation while carrying one word of content. It is the rhythm of a conclusion with none of the work behind it. On reply surfaces the mold appears whenever a prompt asks for something punchy, since the template produces a finished-sounding line from any noun.","examples":[{"before":"Not a tool. Not a platform. Just leverage. That is the whole idea, and it takes a while for anyone to stop asking what the software does and start asking what it removes.","after":"It is a queue with a scheduler on top. The scheduler retries a failed post twice, which is the only reason we stopped losing Friday launches."},{"before":"Not luck. Not timing. Just the work.","after":"We published every Tuesday for nine months. The two posts that carried the year were both written the morning after a support call, which is the only pattern we found.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"(?:^|[.!?]\\s+)Not\\s+[^.!?\\n]{2,40}\\.\\s+Not\\s+[^.!?\\n]{2,40}\\.\\s+(?:Just|Only|Simply)\\s+[^.!?\\n]{1,40}\\.","flags":"gm","scope":"document"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active. Named directly by tropes.fyi and covered by slop-lint, whose molds were screened against a human baseline before shipping. No source reports the shape declining, and it survives paraphrase because it is a mold rather than a phrase."}],"evidence_grade":"community-observed","false_positive_notes":"Advertising copywriters built this template and still teach it. Agency taglines, film trailer voice-over, sports broadcast and political slogans all use the denial-denial-reveal triplet on purpose, and a copywriter can usually say which brand line it descends from. Poets and lyricists use fragment triplets for rhythm. A reviewer separates craft from filler by asking whether the third fragment names something checkable, and whether the same mold recurs across a body of posts by the same account. One triplet is a device. The same triplet every week is a generator.","model_attribution":"Unattributed. No vendor documentation and no peer-reviewed source names the triplet as a model habit. Both sources that carry it are practitioner repositories, which is what the community-observed grade records.","platform_notes":[{"platform":"x","note":"Short fragment triplets fit the reply register that X routes through its llm_slop_post label. The classifier prompts behind that label are withheld from the public repository, so nothing published shows this shape being scored."},{"platform":"linkedin","note":"The triplet is a staple of the one-line-per-sentence post format that human growth marketers popularised on the surface in 2017, well before chat products."}],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"anaphora-abuse","name":"Anaphora abuse","aka":["stacked anaphora","repeated openers for effect"],"category":"rhetorical","subcategory":"parallelism","description":"The same opening words on three or more successive sentences, built as parallel construction. Burton records anaphora as repetition at the beginning of successive clauses, a figure taught continuously since the Rhetorica ad Herennium. The figure is not the pattern. The pattern is the figure arriving with no argument underneath it, where the three members differ only in their last noun. Separate from the statistical opener-repetition measure, which counts function-word entropy across a document and requires no deliberate parallelism at all.","why_it_reads_ai":"Repetition reads as conviction and costs one clause to produce. Heritage and Greatbatch measured parallel three-part structures drawing applause in real political speech, so the shape demonstrably works on people, which is exactly why a preference-trained model reaches for it. That evidence also cuts the other way, and the entry ships contested because of it. Rates in ordinary edited prose have never been measured, for humans or for models.","examples":[{"before":"We write for the people who read every word. We write for the people who skim and come back later. We write for the people who never reply at all. It is an easy thing to say about a newsletter and a much harder thing to keep true once the list is larger than the room.","after":"About one reader in five replies to something each quarter. The other four never will, so the newsletter is written to be useful without a reply.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"(?:^|[.!?][\"'’)\\]]?\\s+)([A-Z][a-z]{1,12})\\s+([a-z][a-z'’]{0,12})\\b[^.!?\\n]{0,120}[.!?]\\s+\\1\\s+\\2\\b[^.!?\\n]{0,120}[.!?]\\s+\\1\\s+\\2\\b","flags":"gm","scope":"document"},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Opened as contested. tropes.fyi and vale-ai-tells both name stacked anaphora, but Burton records an unbroken teaching tradition and Heritage and Greatbatch measured the parallel shape winning applause in human political speech in 1986. Nobody has published a rate for either population, so the entry ships as a review trigger with the counter-evidence stated."}],"evidence_grade":"corroborated","false_positive_notes":"Sermon, eulogy and stump-speech writers use anaphora as a load-bearing figure with an unbroken teaching tradition, and Heritage and Greatbatch measured parallel structures drawing applause at party conferences in 1986, decades before any of this. Children of that tradition are everywhere: closing arguments, campaign copy, liturgy, and hip-hop lyrics all repeat openers on purpose. Second-language writers taught rhetoric formally also reach for it. A reviewer checks whether the repeated members differ in more than their final noun, and whether the passage would lose an argument if the repetition were flattened. If flattening loses nothing, the figure was carrying the paragraph alone.","model_attribution":"Unattributed at the family level. tropes.fyi and vale-ai-tells describe it as a general assistant habit and neither publishes a measurement. Our research pass classified any claim that one product does this more than another as folklore.","platform_notes":[{"platform":"linkedin","note":"Repeated openers are a convention of the one-line-per-sentence post format on that surface, which a human growth marketer popularised in 2017."}],"languages":["en"],"sources":[{"kind":"external","title":"Silva Rhetoricae (Burton, BYU), anaphora","url":"https://rhetoric.byu.edu/Figures/A/anaphora.htm","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Heritage & Greatbatch, Generating applause: a study of rhetoric and response at party political conferences, American Journal of Sociology 92(1) 1986","url":"https://doi.org/10.1086/228465","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"false-suspense-transition","name":"False-suspense transition","aka":["Here's the kicker","But here's where it gets interesting","narrative pivot"],"category":"rhetorical","subcategory":"pivot","description":"A line announcing that something surprising follows. tropes.fyi files the family under its most quoted member and vale-ai-tells ships a narrative-pivot rule with 53 tokens. The pivot alone decides nothing. The check is mechanical: read the sentence after the pivot and look for a name, a number or an event that has not already appeared in the piece. If none is there, the pivot was the payload.","why_it_reads_ai":"The line is free to write. It promises a reveal, which buys a paragraph break and a beat of attention, and a promise costs nothing to make. The shape around a good story is easy to imitate; the reveal at the centre needs a fact. Long-form feature writers earn the same pivot every week and pay for it in the next sentence, which is why the test has to be on what follows rather than on the words themselves.","examples":[{"before":"Here's the kicker: nobody realises this is happening while it is happening. The signals sit there the whole time and nothing gets looked at until something breaks loudly enough to be noticed.","after":"Two of the six teams we onboarded in June had a second, older workspace nobody had closed. Both were still being billed for it.","note":"Figures in this repair are invented for the specimen."},{"before":"But here's where it gets interesting. The implications are bigger than they first appear.","after":"The rate limit applies at queue time, not at publish time, so anything scheduled before the July cutoff kept the old allowance. That is why the queue drained faster in June."}],"detection":{"type":"deterministic","pattern":"\\b[Hh]ere(?:['’]s| is)\\s+(?:the\\s+(?:kicker|twist|catch|best\\s+part|wild\\s+part|crazy\\s+part)|where\\s+it\\s+gets\\s+(?:interesting|wild|good|weird))\\b","flags":"g","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active. Named by tropes.fyi and implemented as a rule family in vale-ai-tells. Neither source publishes a rate, which is what the community-observed grade records, and no source reports the shape declining."}],"evidence_grade":"community-observed","false_positive_notes":"Long-form feature writers, podcast hosts and direct-response copywriters use earned pivots as a matter of craft, and the device predates chat products by a century of magazine writing. Broadcast scripts use it to hold an audience across an ad break. The reviewer check is mechanical and does not require taste: read the sentence immediately after the pivot and look for a name, a number, a date or an event that has not already appeared. When one is there, the pivot was earned and the entry does not apply. When the next sentence only restates the promise in bigger words, the pivot was the whole payload.","model_attribution":"Unattributed. Both sources are practitioner repositories describing assistant output in general terms, and neither separates model families.","platform_notes":[{"platform":"youtube","note":"The YouTube spam policy names malicious clickbait in titles, thumbnails and descriptions. A pivot that promises a reveal the video never delivers is the prose form of that, though the policy turns on the mismatch and not on the phrasing."}],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"magic-adverbs","name":"Magic adverbs","aka":["quietly reshaping","silently transforming","subtly"],"category":"rhetorical","subcategory":"unearned-drama","description":"One adverb carries the entire claim. tropes.fyi names the family after its most common member. The sentence asserts that a large change happened without anyone noticing, which conveniently removes the obligation to say who noticed it, when, or how it was measured. The repair is never a smaller adverb. It is a date and a name.","why_it_reads_ai":"An unwitnessed change cannot be checked. The adverb supplies the drama and deletes the evidence in the same word, which makes it the cheapest available way to sound like reporting. Feature ledes use the move deliberately and then settle the bill in the next paragraph with a scene, a source and a time. The low-effort version never settles it. A separate semantic entry covers the claim-level version of this problem, where the inflation lives in the sentence rather than in one word; this rule stays on the adverb.","examples":[{"before":"A handful of small studios are quietly reshaping how software gets sold. The shift has been under way for some time and nobody has quite noticed, which is more or less how shifts of this kind work.","after":"Three studios we buy from moved to usage pricing between January and April. Two of them told us in the same week, after a customer asked why the invoice had changed.","note":"Figures in this repair are invented for the specimen."},{"before":"This shift is subtly transforming the way distributed teams work together.","after":"Our standup moved to a written thread in February. Attendance stopped being a question, and the two people in other time zones now answer first."}],"detection":{"type":"deterministic","pattern":"\\b(?:quietly|silently|subtly|invisibly|steadily|slowly)\\s+(?:reshap|transform|redefin|rewrit|upend|disrupt|reinvent|remak|revolutionis|revolutioniz)(?:e|es|ed|ing)\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active. tropes.fyi names the adverb family directly and vale-ai-tells ships figurative and inevitability rules that catch neighbouring forms. Neither publishes a rate, so the grade stays community-observed and severity stays low."}],"evidence_grade":"community-observed","false_positive_notes":"Feature ledes carry drama on purpose, and a reporter who has spent a month on a story has earned the word. Trade press, obituaries and profile writing all use the unwitnessed-change frame as a hook before naming the witness two sentences later. Fiction uses it without any obligation at all. The reviewer check is whether the sentence, or the one after it, names who noticed the change and when. If the piece can produce a person and a date, the adverb is doing ordinary work. If the change has no observer anywhere in the text, the adverb is the only evidence offered.","model_attribution":"Unattributed. Neither source separates model families, and no vendor documentation names the adverb family.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"false-range","name":"False range","aka":["from X to Y with no spectrum","endpoints that define nothing"],"category":"rhetorical","subcategory":"false-structure","description":"A fronted from-to phrase that looks like a spectrum and bounds nothing. tropes.fyi lists false ranges by name. The test is a midpoint: if no third thing sits between the two endpoints, the construction is decoration. Real ranges have interiors, which is why the rule below excludes numeric endpoints and fires only when the clause after the comma makes a claim about everyone.","why_it_reads_ai":"The frame promises coverage. Two nouns stand in for a whole field, and the sentence then makes a claim about all of it without having examined any part. It is comprehensiveness at the price of a comma, and it survives rewriting, so a word-level check never sees it.","examples":[{"before":"From startups to enterprises, every team needs a content strategy. The specifics differ, obviously, though the discipline underneath does not change much as a company grows, and pretending otherwise is how unsustainable plans get made.","after":"Teams under ten people plan a week ahead. Teams over two hundred plan a quarter ahead and route everything through legal. The companies in between are the ones with no process at all.","note":"Figures in this repair are invented for the specimen."},{"before":"From onboarding to offboarding, everything runs through a single inbox.","after":"Account creation and access removal both run through the same shared inbox. The manager checklist in between is still a spreadsheet nobody owns."}],"detection":{"type":"deterministic","pattern":"(?:^|[.!?][\"'’)\\]]?\\s+|\\n)From\\s+(?![\\d£$€])[^,.;!?\\n]{2,30}\\s+to\\s+(?![\\d£$€])[^,.;!?\\n]{2,30},\\s+(?:every|all|each|everyone|everything|anyone|nobody|no one|the entire|the whole)\\b","flags":"gm","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active. tropes.fyi names the pattern and the anti-slop-writing list carries it as a structural rule. That second list is an evasion prompt and is marked as one, so tropes.fyi is the independent support and the grade stays community-observed."}],"evidence_grade":"community-observed","false_positive_notes":"Survey writing, geography, price bands, date ranges and salary bands all use from-to constructions where the interior is real and the endpoints were chosen to bound it. A train timetable, a demographic breakdown and a product tier table are built from them. The rule below therefore refuses numeric endpoints and requires the following clause to make a universal claim, which keeps ordinary ranges out. A reviewer applying the entry by hand asks for a midpoint: name one thing that sits between the two endpoints. If the writer can, the range is real and the entry does not apply.","model_attribution":"Unattributed. The anti-slop-writing list associates the shape with assistant output generally and publishes no measurement, which is why it is corroboration and never sole support.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"anti-slop-writing system prompt and pattern list","url":"https://github.com/adenaufal/anti-slop-writing","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"asserted-obviousness","name":"Asserted obviousness","aka":["The truth is simple","needless to say","it goes without saying","cannot be overstated"],"category":"rhetorical","subcategory":"assertion","description":"The claim is declared self-evident in place of being argued. tropes.fyi names the shape after one of its members, vale-ai-tells ships an absolute-assertions rule that catches the same family, and slop-lint carries the reply-register forms. The rule below fires on four fixed phrases and nothing else, because the general move is far older and far more common than any of this.","why_it_reads_ai":"Asserting obviousness is the cheapest substitute for evidence available in English, and it never fails a fluency check. A writer with no source can still make a claim sound settled. A writer who has a source usually prefers to give it, since a source persuades better than an assurance does.","examples":[{"before":"The truth is simple: consistency beats talent every time. Nobody disputes it, almost nobody acts on it, and the gap between those two facts is where the interesting part sits.","after":"We published weekly for a year and monthly the year before. The weekly year brought four times the inbound, and the two best posts in it were both written in a hurry.","note":"Figures in this repair are invented for the specimen."},{"before":"Needless to say, the value of a good onboarding call cannot be overstated.","after":"Accounts that take the day-three call renew at about twice the rate of accounts that skip it. That is the only onboarding number we track.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:needless\\s+to\\s+say|it\\s+goes\\s+without\\s+saying|the\\s+truth\\s+is\\s+simple|can(?:not|['’]t)\\s+be\\s+overstated)\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active with low severity. Three practitioner lists carry the family, and none publishes a rate. The move itself is ancient, so the rule is restricted to four fixed phrases and the entry says why."}],"evidence_grade":"corroborated","false_positive_notes":"Polemicists have used the move since long before print, and it is standard in opinion writing addressed to an audience assumed to agree already. Editorials, sermons, campaign literature and after-dinner speeches all lean on it, and so does technical writing when a step really is common ground for its readers. British administrative English keeps several of these phrases as neutral connectives. A reviewer checks what follows the assertion: if the next sentence supplies a source, a number or an example, the phrase was throat-clearing rather than a substitute for argument.","model_attribution":"Unattributed. All three sources are practitioner lists describing assistant output in general, and none separates one product from another.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"aphorism-molds","name":"Manufactured aphorism molds","aka":["X is what keeps A from becoming B","is measured by how","as a hypothesis, not a promise"],"category":"rhetorical","subcategory":"reply-register","description":"Reply-sized wisdom cast from a small set of reusable frames. slop-lint catalogues the molds by surface string and states its method: each one was screened against roughly 60,000 words of human baseline for zero false positives, and required two independent sightings before it shipped. That method is the entire basis for the grade here, and it is one repository, so the entry stays community-observed. The frames produce sentences with the shape of earned insight that can be filled with any two nouns.","why_it_reads_ai":"A reply that has to sound wise in twenty words has two options. Say a specific thing, which requires knowing one, or reach for a frame that always produces a finished sentence. Frames win on cost every time. Recurrence is the signal: one aphorism is a writer having a line, the same mold three times in a month is a generator running.","examples":[{"before":"Trust is what keeps a team from becoming a queue. Every company says it values trust, and almost none of them can point to something they did last quarter to earn any. The teams that get this right rarely talk about it. The ones that talk about it constantly are usually the ones who lost it a while ago.","after":"We stopped writing tickets for anything under ten minutes in March. Queue time halved, and two people said they finally knew what everyone else was working on.","note":"Figures in this repair are invented for the specimen."},{"before":"Culture is measured by how quickly bad news travels. Every leadership team believes its own culture is open, and most of them have never once tested the claim. Organisations that handle bad news well tend to handle everything else well too. It is a simple idea and a hard one to live with.","after":"The outage on 12 May reached the whole company in nineteen minutes because the on-call channel is public. The one in January took two days and three forwarded emails.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:is\\s+what\\s+keeps\\s+[^.;!?\\n]{1,40}\\s+from\\s+becoming\\b|is\\s+measured\\s+by\\s+how\\s+(?:well|quickly|fast|often|much|little|far)\\b|as\\s+a\\s+hypothesis,\\s+not\\s+a\\s+promise\\b|lowers?\\s+[^.;!?\\n]{1,30}\\s+without\\s+lowering\\b)","flags":"gi","scope":"sentence"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active at high severity. slop-lint screened each mold against roughly 60,000 words of human baseline and required two independent sightings before shipping it, which is a stronger method than any other list in this category offers. It is still a single repository, so the grade stays community-observed."}],"evidence_grade":"community-observed","false_positive_notes":"Professional aphorists exist, and so do quote-card writers, keynote speakers, copywriters and the authors of management books, all of whom are paid to produce exactly these sentences. Proverbs in every language use the same frames, and a writer raised on them will reach for one without thinking. The discriminator is recurrence rather than presence: check whether the same mold appears across a body of posts from one account, and whether the two nouns in the frame could be swapped for any other pair without changing what the sentence claims. A line that only works with its own nouns was written by someone who meant it.","model_attribution":"Unattributed by family. slop-lint records the molds as assistant reply-register output and names no product, and its README documents the discovery method rather than a per-model rate.","platform_notes":[{"platform":"x","note":"The reply register is where slop-lint found these. X routes reply spam and its llm_slop_post label to the same handling, and reply-guy behaviour is the archetypal target named in the public algorithm repository."},{"platform":"linkedin","note":"Quote-card posts and comment replies are the native home of the molds on that surface. LinkedIn shipped a user-facing report control for suspected slop in July 2026 whose signal feeds ranking, and no LinkedIn document names any sentence shape."}],"languages":["en"],"sources":[{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"mic-drop-closer","name":"Mic-drop closer","aka":["epiphonema","standalone final line","closing one-liner"],"category":"rhetorical","subcategory":"closer","description":"A final line, set as its own paragraph, engineered to land. Rhetoric calls the pithy closing sentence epiphonema and the slot it occupies the peroratio, and both are legitimate and old. The catalogued version restates the piece in a shorter and grander sentence that introduces nothing. This entry is positional and applies only to the last line; the aphorism-mold entry fires on surface strings anywhere in a text, so the two never cover the same sentence.","why_it_reads_ai":"The closer is the easiest sentence in any piece to generate and the hardest to justify. It has no obligation to be true, only to sound final, and finality is a matter of rhythm rather than content. vale-ai-tells ships a 63-token list for the family, which is a fair measure of how standardised the vocabulary has become.","examples":[{"before":"We rebuilt the publish queue in June and the retry logic in July.\n\nThe future belongs to the teams who ship.","after":"We rebuilt the publish queue in June and the retry logic in July.\n\nThe retry change is the one that mattered. A failed post now goes out inside ten minutes instead of never.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Read only the final paragraph of the supplied text. Answer yes when all four hold. (1) The final paragraph is a single sentence standing alone. (2) It makes a general claim about people, work, time or the future rather than a claim about the subject of the piece. (3) It contains no proper noun, number, date, quoted term or named object that appears earlier in the body. (4) Deleting it removes no information from the piece. Answer no when the closer states what the writer will do next, names a person, product or date, answers a question the body left open, or is dialogue. Answer not applicable when the text has fewer than two paragraphs. Return the verdict, then quote the final sentence verbatim."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active. Three practitioner sources carry the family, one of them as a rule with a 63-token vocabulary list. None publishes a rate, and the closing device itself is ancient, so the grade is community-observed and the rubric requires four conditions rather than one."}],"evidence_grade":"community-observed","false_positive_notes":"Columnists, speechwriters and essayists end on a line for good reasons, and the peroratio has been a taught structural slot for two thousand years. Sports writing, obituaries and campaign copy all close on a summarising sentence by convention. The check is whether the closer names something specific from the body: a person, a number, a decision, a date. A closer that gathers up the actual argument is doing the job the figure was invented for. A closer that would fit unchanged at the end of a different article about a different subject is the version this entry describes.","model_attribution":"Unattributed. The three sources describe assistant output in general and none separates model families or publishes a per-product rate.","platform_notes":[{"platform":"linkedin","note":"The standalone final line is a formatting convention of the surface, where posts are read in a narrow column and the last line sits above the follow button. The convention predates chat products there."}],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"hedging-preamble","name":"Hedging preamble","aka":["It's worth noting","It's important to note","worth noting that"],"category":"rhetorical","subcategory":"hedging","description":"A softening scaffold placed in front of a plain statement. Lakoff named the device a hedge in 1972, and Hyland remains the standard reference for hedging in academic prose, where it is a taught obligation rather than a tic. The catalogued version prepends the scaffold to a sentence nobody would have challenged. Contested by construction: this family sits inside the Academic Formulas List, so the base rate in careful human writing is high and unmeasured.","why_it_reads_ai":"Assistants are trained to avoid overclaiming, and a preamble is the most visible cheap way to comply. The scaffold also buys several words before the sentence has to commit to anything. The result reads as caution with no exposure behind it, because the statement that follows is usually not contentious in the first place.","examples":[{"before":"It's worth noting that the deadline may shift depending on a number of external factors. Timelines of this kind are rarely as fixed as they look on a plan, and most teams build in some room without ever saying so. What matters is keeping everyone informed as things develop. Surprises are much easier to absorb early.","after":"The deadline moves if the certification body misses its own 30 April response window. It missed that window last year and the whole schedule went back a month.","note":"Figures in this repair are invented for the specimen."},{"before":"It is important to note that results can vary considerably between teams.","after":"Two of our five teams saw nothing change at all. Both were already publishing daily before the switch, which is most of the explanation.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:it['’]s|it\\s+is)\\s+(?:worth\\s+(?:noting|mentioning|pointing\\s+out)|important\\s+to\\s+(?:note|remember|mention))\\s+that\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Opened as contested. Four independent lists carry the family, including Wikipedia and a professional editing service, but the same formulas appear in the Academic Formulas List as taught academic register. Hedging is a documented obligation in academic, medical and legal writing, so the entry ships with the contest on the record."}],"evidence_grade":"corroborated","false_positive_notes":"Academic hedging is a documented and explicitly taught register, and Hyland describes it as a requirement of research writing rather than a weakness. Medical, legal and regulatory writers hedge because their liability requires it, and a clinician who writes that a result may vary is stating the truth about the evidence. Second-language academic writers use lexical bundles from the taught formula lists at higher rates than native speakers, which is exactly the population that detection tools already misjudge. A reviewer checks the sentence underneath the scaffold: if the claim could be wrong, the hedge is doing real work; if the claim is uncontroversial, the scaffold is filler.","model_attribution":"Reported for chat assistants broadly. Anthropic publishes dated system prompts that suppress several neighbouring words, which is first-party evidence of a default pull toward this register, though those prompts name words rather than this frame. No source separates model families on the preamble itself.","platform_notes":[{"platform":"wikipedia","note":"The guide lists softening scaffolds among the signs while its own manual of style treats unattributed hedging as a separate editorial problem with its own shortcut."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Proofed, AI editing checklist","url":"https://proofed.com/knowledge-hub/ai-editing-checklist-how-to-spot-and-fix-ai-writing-patterns/","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tellsign word and phrase lists","url":"https://github.com/ctkrug/tellsign","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"vague-attribution","name":"Vague attribution","aka":["weasel wording","experts argue","industry reports suggest","source inflation"],"category":"rhetorical","subcategory":"weasel","description":"The authority is asserted and never named. Wikipedia calls this weasel wording and keeps a shortcut for the AI-writing case. The same entry covers source inflation, where one document becomes reports and one analyst becomes analysts, because the text looks identical on the page and the check is identical too: ask who, and see whether the piece can answer.","why_it_reads_ai":"A writer who cannot retrieve a source can still produce the sentence shape that a source would license. The plural does all the work, and it raises the cost of checking past what most readers will pay. This is the highest-severity entry in the category for that reason: the sentence makes a claim about the world and supplies no way to test it.","examples":[{"before":"Experts argue that the shift will accelerate through the rest of the decade. The pace is harder to call than the direction, and reasonable people disagree about both without much changing as a result. Organisations that prepare early are generally glad they did, though what preparing early consists of is never quite specified. By the time a change of this size is obvious to everyone, whatever advantage there was has usually gone somewhere else.","after":"Two of the four vendors we buy from have said they will drop support for the old format after 2027. The other two have not answered the question twice now.","note":"Figures in this repair are invented for the specimen."},{"before":"Industry reports suggest that daily publishing is now the norm. The pressure to keep up is real and felt hardest by the teams with the smallest budgets, and whether the pace is sustainable gets asked a great deal less often than whether it is being met.","after":"A trade body survey of 312 marketers in June put daily publishing at 18 percent. It sampled agencies only, which is most of the reason the number looks low.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:(?:many|some|most)\\s+)?(?:experts|analysts|critics|observers|industry\\s+(?:reports|analysts|insiders|watchers))\\s+(?:argue|suggest|say|believe|agree|note|claim|warn|point\\s+out|have\\s+noted)\\b","flags":"gi","scope":"sentence"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active at high severity. Wikipedia names it with a dedicated shortcut, tropes.fyi lists vague attributions, and vale-ai-tells ships a rule with 61 tokens for the family. The severity reflects the consequence rather than the frequency: the sentence makes a factual claim and removes the means of checking it."}],"evidence_grade":"corroborated","false_positive_notes":"Journalists protecting sources write around names as a professional obligation, and a reporter who writes that two people familiar with the deal said something is following a rule, not dodging one. Trade press summarises a field where naming every practitioner would be unreadable. Analyst notes are often under embargo. The check is whether the piece says why the source is unnamed, or names the class precisely enough to be found: a trade body, a named survey, a court filing. Where the text can answer the question who, the attribution is thin rather than empty. Where nothing in the piece could ever answer it, the sentence is decoration in the shape of evidence.","model_attribution":"Reported for chat assistants broadly and named by Wikipedia among current signs. No source separates model families, and the shape is old enough that its manual of style entry predates chat products by many years.","platform_notes":[{"platform":"wikipedia","note":"The weasel-wording shortcut is one of the oldest editorial rules on the project, and the AI-writing guide reuses it rather than inventing a new one."},{"platform":"google","note":"The Search spam policy turns on value to users no matter how a page was created. Unattributed claims are part of what makes low-value pages cheap to produce at scale, though no policy text names the construction."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"false-vulnerability","name":"False vulnerability","aka":["fake-humility opener","I don't usually post this, but","performed reluctance"],"category":"rhetorical","subcategory":"performed-intimacy","description":"A confession used as a hook. tropes.fyi lists the pattern by name and vale-ai-tells ships neighbouring rules for the register. The opener claims reluctance, a failure, or an unusual departure from habit, and the piece then proceeds exactly as a planned piece would, arriving at a lesson or an offer within a few sentences. The reluctance is packaging.","why_it_reads_ai":"Disclosure buys attention, and a claimed disclosure buys the same attention without the disclosure. This is a judge rule because the surface strings vary endlessly while the structure does not: the admitted failure carries no date, no amount, no named person and no consequence that outlived the paragraph, and the text recovers immediately. Human growth marketers built the move years before assistants could imitate it.","examples":[{"before":"I almost did not share this. Three years ago I nearly walked away from all of it. What I learned about resilience changed how I work.","after":"I closed the agency in March 2023 with two months of runway and told four people they no longer had a job. I still owe one of them a reference.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Read the first three sentences and the last three sentences of the supplied text. Answer yes when all four hold. (1) An opening sentence claims reluctance to post, an unusual departure from habit, or a confession. (2) The admitted failure names no date, no amount, no person and no consequence that continued past the paragraph. (3) Within three sentences the text turns to advice, a lesson, a product or an invitation. (4) Deleting the opening sentence leaves the remainder intact and coherent. Answer no when the confession carries a checkable specific, when the text stays with the failure to the end, when the writer states a cost that is still being paid, or when the piece is a reply to someone else. Return the verdict and quote the opening sentence verbatim."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active. tropes.fyi names the pattern and vale-ai-tells covers the surrounding register. Neither publishes a rate. Our own drafts show the opener recurring from a single prompt template, which is recorded as an observation about our generations and not as a measurement of anyone else."}],"evidence_grade":"community-observed","false_positive_notes":"People who genuinely hesitate before posting write this sentence because it is true, and memoirists, grief writers and anyone disclosing an illness or a redundancy open the same way for the same reason. The move also has a documented human commercial history: growth marketers on professional networks taught it explicitly years before assistants could imitate it, so a hit says nothing about who typed it. The separating question is cost. A real disclosure names a date, an amount, a person or a consequence the writer is still carrying, and it does not resolve into advice within three sentences. Treat the entry as a reason to read the piece, never as a judgement about the writer.","model_attribution":"Unattributed. Neither source separates model families. The pattern appears in our own generated drafts across several prompt templates, which is an observation about our tooling rather than about any vendor.","platform_notes":[{"platform":"linkedin","note":"This is a native shape on the surface and has been since well before chat products. LinkedIn policy targets automated posting and engagement pods rather than any writing style, so nothing in the policy record applies to it."}],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"feedsquad-observed","title":"FeedSquad draft review, social openers","observed":"2026-08-15","corpus":"Generated LinkedIn and X drafts inside FeedSquad, read by hand during the August 2026 index build. The opener recurred across drafts produced from one prompt template. No count is published here because the sample was assembled by hand and is not a rate."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"manufactured-relatability","name":"Manufactured relatability","aka":["We've all been there","claimed shared experience"],"category":"rhetorical","subcategory":"performed-intimacy","description":"A claim about the reader, asserted as shared fact. slop-lint carries the reply-register forms and vale-ai-tells ships a rule for the family. Separate from the false-vulnerability entry, which performs something about the writer. This one performs something about the audience, and it is checkable in a way the other is not: either the piece names the audience it claims to know, or it does not.","why_it_reads_ai":"The sentence manufactures a group and puts the reader inside it before any argument starts. It costs one clause and it flatters. Advice columnists write to a readership that genuinely shares the experience and can usually name it in the same paragraph, which is the difference worth looking for.","examples":[{"before":"We've all been there. The deadline moves, the brief changes, and suddenly it is Friday afternoon, which everyone in this line of work recognises including the people who insist that they do not.","after":"Half our client briefs changed after kickoff last quarter. The two that did not were both from clients who had run the same campaign the year before.","note":"Figures in this repair are invented for the specimen."},{"before":"If you're anything like me, you have tried and abandoned four different systems.","after":"I abandoned three systems in 2024 and kept the fourth, because it survived a week when I was ill and never opened a laptop."}],"detection":{"type":"deterministic","pattern":"\\b(?:we(?:['’]ve|\\s+have)\\s+all\\s+been\\s+there|we(?:['’]ve|\\s+have)\\s+all\\s+(?:felt|done|seen|made)\\s+(?:it|this|that)|if\\s+you(?:['’]re|\\s+are)\\s+(?:anything\\s+)?like\\s+me|you\\s+know\\s+the\\s+(?:feeling|drill))\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active at low severity. Two practitioner lists carry the family and neither publishes a rate. The phrases are also ordinary spoken English, which is why the rule is restricted to four fixed openers."}],"evidence_grade":"community-observed","false_positive_notes":"Advice columnists, teachers, support agents and anyone writing to a defined community are addressing readers who really do share the experience, and naming it is the point of the genre. Comedians build sets on it. Newsletters with a single well-known audience use the phrase as shorthand for something their readers wrote in about last week. The check is whether the piece names the audience it claims to know, or shows any evidence of having met it: a reply quoted, a survey, a support ticket, a room. Where the writer can point to the group, the sentence is reporting. Where the group exists only in that sentence, it was invented to hold the reader.","model_attribution":"Unattributed. slop-lint records the forms as reply-register output without naming a product, and vale-ai-tells does not separate model families.","platform_notes":[{"platform":"threads","note":"Meta has defined and demoted engagement bait since December 2017 and its 2025 and 2026 originality posts never mention AI. A relatability opener is not engagement bait by that definition unless it asks for a reaction."}],"languages":["en"],"sources":[{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"curiosity-gap-withholding","name":"Curiosity-gap withholding","aka":["The one thing nobody tells you","withheld payload headline"],"category":"rhetorical","subcategory":"headline","description":"A headline or subject line names a category of information and withholds the item itself. The peer-reviewed anchor studies headline concreteness in information selection and records that curiosity-gap style spread through digital publishing in the early 2010s, well before any assistant existed. This entry absorbs the false-exclusivity family, where the withheld item is also claimed to be secret. Contested, because the human origin is documented and the machine version adds nothing new.","why_it_reads_ai":"Withholding is free and specifics are expensive. A headline that names its payload has to have one. The version worth flagging is the one where the body never supplies the promised item either, which is a property of the whole piece rather than of the headline, and is why this is a judge rule.","examples":[{"before":"The one thing nobody tells you about hiring your first employee","after":"Your first employee will ask for a written role description before you have written one. Ours did, in week two, and we had nothing to send.","note":"Figures in this repair are invented for the specimen."},{"before":"What most founders miss about pricing, and how to fix it","after":"We moved from per-seat to per-workspace pricing in April. Two accounts doubled their spend, one halved it, and support tickets about seat counts went to zero.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Look only at the headline, subject line or first line of the supplied text. Answer yes when all three hold. (1) It names a category of information, such as a thing, a mistake, a reason, a number or a secret, without naming the item itself. (2) The withheld item is presented as unknown to the reader or hidden from them. (3) Reading the body does not produce the promised item as something a reader could write down in one sentence. Answer no when the body names the item within its first two paragraphs, when the headline itself names the item, or when the format is a quiz, a puzzle or a fiction serial where withholding is the genre. Return the verdict, quote the headline, and state the item the body supplies if it supplies one."},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Opened as contested. The peer-reviewed anchor documents curiosity-gap headlines spreading through human digital publishing in the early 2010s, so the shape carries no authorship signal at all. It ships because a headline whose payload never appears is worth a review pass, and the entry says plainly that the origin is human."}],"evidence_grade":"corroborated","false_positive_notes":"Human viral publishing invented this and any tabloid front page still qualifies, which is the reason the entry ships contested rather than active. Newsletter writers are taught curiosity-gap subject lines as craft, and a subject line that gives away the whole issue is a bad subject line by the standards of the trade. Fiction, quizzes and serialised reporting withhold on purpose. The reviewer test is on the body, not the headline: can a reader write down the promised item in one sentence after reading the piece? If yes, the headline was a hook doing its job. If no, the promise was the product.","model_attribution":"Unattributed, and the pattern predates assistants by a decade of human publishing. Neither practitioner source separates model families, and the peer-reviewed source studies human headlines only.","platform_notes":[{"platform":"youtube","note":"The spam policy names malicious clickbait in titles, thumbnails and descriptions. The policy turns on the mismatch between promise and content, not on the headline style."},{"platform":"meta","note":"Engagement bait has been formally defined and demoted since December 2017. The withheld-payload headline is adjacent to that definition but is not the same thing, since the policy targets asks for reactions."}],"languages":["en"],"sources":[{"kind":"external","title":"Scientific Reports, clickbait and curiosity gap headlines","url":"https://www.nature.com/articles/s41598-024-81575-9","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"analogy-reflex","name":"Think-of-it-as analogy reflex","aka":["Think of it as","imagine X as Y"],"category":"rhetorical","subcategory":"analogy","description":"An explanation reaches for a comparison before it has stated the thing. tropes.fyi lists the pattern and vale-ai-tells ships a rule for the family. The analogy usually lasts one sentence and never returns, which is the tell. A working analogy gets carried, extended, and eventually broken on purpose when the writer reaches the point where it stops holding.","why_it_reads_ai":"Comparison is the cheapest form of explanation and it always produces a sentence. It also lets the writer skip the mechanism entirely, since the reader supplies the meaning from the familiar side of the comparison. Science communicators and teachers build whole explanations from analogy and carry them for pages.","examples":[{"before":"Think of it as a nervous system for your whole content operation. Everything connects to everything else, and that is really the whole promise once the feature list is set aside.","after":"It watches the publish queue and reposts anything the platform rejected, which is the only part anyone notices. The rest is a scheduler."},{"before":"Imagine it as a filing cabinet that files itself.","after":"Uploads land in one bucket and a nightly job tags them by client from the filename. Anything it cannot parse goes to a folder someone checks on Mondays."}],"detection":{"type":"deterministic","pattern":"\\b(?:think\\s+of\\s+(?:it|this|them)\\s+as|imagine\\s+(?:it|this|them)\\s+as|picture\\s+(?:it|this)\\s+as)\\s+(?:a|an|the)\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active at low severity. Two practitioner lists name the pattern and neither publishes a rate. The rule is a single narrow phrase family, which keeps the cost of a false hit close to zero."}],"evidence_grade":"community-observed","false_positive_notes":"Science communicators, teachers and technical writers build explanations from analogy because it is the fastest route into an unfamiliar system, and physics teaching in particular is largely a sequence of analogies with their limits marked. Documentation uses the phrase to orient a reader before the reference material starts. The reviewer check is whether the analogy is carried past its first sentence: a writer who means it will extend the comparison, test it, and say where it breaks. A comparison that is dropped immediately was decoration, and the mechanism it stood in for is still missing.","model_attribution":"Unattributed. Neither source separates model families or names a vendor.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"imagine-a-world-opener","name":"Imagine-a-world opener","aka":["Imagine a world where","Picture this"],"category":"rhetorical","subcategory":"opener","description":"The hypothetical-scenario opener. tropes.fyi lists it and vale-ai-tells ships an opening-cliches rule with 92 tokens covering it and its neighbours. Burned rather than active: the phrase has been quoted in so many guides and jokes that it now reads as parody, and the writers who used it in earnest have mostly stopped.","why_it_reads_ai":"A hypothetical costs nothing to invent and delays the first checkable fact by a paragraph. Publicity is what burned it. The same mechanism has been measured elsewhere: Geng and Trotta recorded a publicised marker collapsing in academic writing soon after it was named, while unpublicised markers kept rising. A directory like this one accelerates that decay for everything it publishes, which is why status is a field here rather than a footnote.","examples":[{"before":"Imagine a world where your content writes itself. The calendar fills, the drafts appear, and the rest comes down to deciding what is worth keeping and what was never worth writing.","after":"Our drafts come out of the same prompt every Monday and two of us rewrite them by Wednesday. The rewrite is the job."},{"before":"Picture this: it is Monday morning and your inbox is already full.","after":"On 3 February the shared inbox held 214 unread client messages at nine in the morning, because the weekend autoresponder had failed silently on the Friday.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"(?:^|[.!?][\"'’)\\]]?\\s+|\\n)(?:Imagine\\s+a\\s+world\\s+(?:where|in\\s+which)|Picture\\s+this[:.,]|Imagine\\s+for\\s+a\\s+moment)","flags":"gm","scope":"document"},"severity":"low","status":"burned","status_history":[{"date":"2026-08-15","status":"burned","rationale":"Opened as burned. The opener is quoted in every popular guide to spotting machine prose and in the jokes about them, so human writers now trip it knowingly while the earnest use has thinned out. Geng and Trotta measured exactly this decay for a publicised lexical marker. The rule stays in the dataset and stops firing, because retired signal is still information."}],"evidence_grade":"community-observed","false_positive_notes":"Conference speakers and advertising copywriters invented the move and it still works in a room, where the speaker is present and the audience has agreed to be led somewhere. Fiction, speculative journalism and policy scenario writing all open on a hypothetical as a matter of form. Because the phrase is now famous, plenty of human writers use it with a wink. That combination is why the status is burned: a hit says something about the sentence and close to nothing about who wrote it, and the rule is de-armed for that reason rather than deleted.","model_attribution":"Unattributed. One of the three sources is an evasion prompt listing phrases to avoid, which is inverted evidence of what models produced when it was written, and it is marked as such. The other two are practitioner lists that name no vendor.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"keskinonur gist, words and phrases to avoid for ChatGPT","url":"https://gist.github.com/keskinonur/4b2d9b7f3311332cf60c91cb45efb362","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"scene-setting-opener","name":"Generic scene-setting opener","aka":["In today's fast-paced world","In the ever-evolving landscape of","In today's digital age"],"category":"rhetorical","subcategory":"opener","description":"The zoomed-out first line. It describes an era instead of a subject, and it fits any topic at all. vale-ai-tells covers it inside an opening-cliches rule with 92 tokens. Burned: a decade of content marketing taught the opener deliberately, and by 2026 it is quoted as a joke more often than it is written in earnest. This entry owns the full landscape phrase in its aka list, so no other entry claims a bare landscape token.","why_it_reads_ai":"The opener is topic-independent, which is what makes it cheap to generate and useless to read. A first line that survives being moved to another article was never about this one. Search-trained writers were taught to open this way on purpose, so the shape carries a human history at least as long as the machine one.","examples":[{"before":"In today's fast-paced world, staying visible online has never been more challenging. Audiences have more to choose from and less time to spend on any of it, which puts genuine pressure on anyone trying to reach them, and the brands still standing in a few years will be the brands that made peace with it early.","after":"Reach on our page fell by half between March and June while posting frequency stayed flat. The drop tracked a change in how the feed treats outbound links.","note":"Figures in this repair are invented for the specimen."},{"before":"In the ever-evolving landscape of digital marketing, brands must adapt or fade.","after":"Two of our three paid channels stopped clearing cost in the second quarter. We moved that budget into the newsletter, which is the only channel with a list we own.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\bin\\s+(?:today['’]s\\s+(?:fast[-\\s]paced|digital|modern|ever[-\\s]changing|hyper[-\\s]connected|competitive)\\s+(?:world|age|era|landscape|environment|marketplace)|the\\s+ever[-\\s](?:evolving|changing|shifting)\\s+(?:landscape|world|realm|field)\\s+of)\\b","flags":"gi","scope":"sentence"},"severity":"medium","status":"burned","status_history":[{"date":"2026-08-15","status":"burned","rationale":"Opened as burned. Content marketers taught this opener explicitly for a decade before chat products existed, and it is now the single most quoted example in popular guides, so human writers trip it constantly. The rule stays listed and stops firing, since a retired signal is information about the discourse."}],"evidence_grade":"community-observed","false_positive_notes":"A decade of human content marketing wrote this opener first, and search-trained writers were taught it explicitly as a way to establish topical context in the first line. Textbooks, press releases and student essays all use the era-opener as a taught convention, and second-language writers pick it up from the same instruction. Because the phrase is famous, it also appears in parody. A reviewer should ignore the phrase entirely and look at the second sentence: if the piece has still not named its subject by then, the problem is the missing subject rather than the opener.","model_attribution":"Unattributed. Two of the three sources are evasion prompts listing phrases to avoid, marked as such and usable only as inverted evidence of what models produced when they were written. The practitioner rule set is the load-bearing citation here and names no vendor.","platform_notes":[{"platform":"google","note":"The Search quality rater guidelines rate scaled, unoriginal, low-effort content at the bottom regardless of how it was made. An opener that transfers between topics is one of the things that makes such pages cheap to produce."}],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"keskinonur gist, words and phrases to avoid for ChatGPT","url":"https://gist.github.com/keskinonur/4b2d9b7f3311332cf60c91cb45efb362","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"},{"kind":"external","title":"anti-slop-writing system prompt and pattern list","url":"https://github.com/adenaufal/anti-slop-writing","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"moralizing-closer","name":"Moralizing closer","aka":["unsolicited lesson","what this teaches us"],"category":"rhetorical","subcategory":"closer","description":"A lesson appended to material that did not produce it. The story was specific, the moral is general, and the two are joined by a sentence beginning with what this teaches or what it all means. The is-this-ai-slop list and tropes.fyi both carry the family. Parables put the moral last because the genre requires it, and in a parable the moral names the thing in the story.","why_it_reads_ai":"A lesson can be generated from a summary. It requires no access to the events, only to their shape, which is why it can be produced by anything that has read the first paragraph. That is also the mechanical check: whether the moral names anything from the story it closes.","examples":[{"before":"The migration took eleven days and cost us two weekends.\n\nWhat this teaches us is that preparation always pays off in the end.","after":"The migration took eleven days and cost us two weekends.\n\nThe part that ran long was the account merge, because we had never written down which of the two systems owned the email field. That decision goes in the plan next time."}],"detection":{"type":"judge","rubric":"Read the final paragraph of the supplied text. Answer yes when all three hold. (1) It states a lesson, principle or rule addressed to the reader. (2) It names nothing from the body: no person, place, number, date, system or decision that appeared earlier shows up in it. (3) It would fit unchanged at the end of a different piece about a different subject. Answer no when the closer names a specific from the body, when the piece is a parable, fable or teaching story whose genre puts the moral last, or when the lesson states a change the writer is making and says what it is. Return the verdict and quote the closing sentence verbatim."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active at low severity. Two practitioner lists carry the family and neither publishes a rate. The moralising closer is also an old and legitimate genre convention, which the rubric excludes by name rather than by taste."}],"evidence_grade":"community-observed","false_positive_notes":"Parables, fables, sermons and writing for children put the moral last because the form requires it, and religious and pedagogical traditions have done so for millennia. Case studies in business and medical teaching end on a stated lesson by convention, and so do incident reviews, where the lesson is the deliverable. The check is whether the moral names anything from the story it closes: a person, a system, a number, a decision. A lesson built from the events is the point of the genre. A lesson that would fit at the end of any other story was written without reading this one.","model_attribution":"Unattributed. Both sources are practitioner lists describing assistant output in general terms and neither separates model families.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"is-this-ai-slop word and phrase list","url":"https://github.com/didrod205/is-this-ai-slop","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"superlative-stacking","name":"Superlative stacking","aka":["absolutely perfect","hands down the best"],"category":"rhetorical","subcategory":"superlative","description":"Superlatives and absolute intensifiers piled inside a short passage. Ott, Choi, Cardie and Hancock measured more superlatives in deceptive hotel reviews than in truthful ones in 2011, more than a decade before any of this. That finding is about human deception. It carries no claim about who or what wrote a text, and it is included here as a low-effort signal on exactly those terms.","why_it_reads_ai":"Praise with no interior is cheap to produce at any length. The same 2011 paper frames deceptive writing as imaginative rather than informative and reports that people inventing an experience have trouble encoding concrete spatial detail. That is the useful reading of a stack: look at what the passage fails to contain rather than at the adjectives it contains. The same authors found human judges performing at roughly chance on the task, which is the strongest reason to treat any stack as a prompt to read rather than a conclusion.","examples":[{"before":"Absolutely perfect from start to finish. Hands down the best experience we have ever had, and the staff were incredibly amazing throughout.","after":"We arrived at 23:00 and the kitchen reopened to make one plate of pasta. The room was above the bins, which we would mention to anyone booking in July."}],"detection":{"type":"statistical","metric":"superlative-and-absolute-intensifier-density-per-1000-words","threshold":8,"direction":"above","threshold_basis":"Ott and colleagues report that deceptive reviews contain more superlatives than truthful ones and publish no per-thousand-word rate for either class. No other source publishes one. Eight per 1,000 words is a FeedSquad review trigger set by arithmetic rather than measurement: a 250-word review crosses it on the third stacked superlative. Treat a crossing as a reason to read the passage for concrete detail. It is not a verdict, and it is not evidence about authorship."},"severity":"medium","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Opened as contested. The peer-reviewed evidence is a 2011 study of human deceptive reviews, which is a different claim from a machine signal, and the same paper found human judges at roughly chance on the task. The entry ships with that boundary stated and with a threshold declared as a review trigger."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Enthusiastic real customers write exactly like this, and so do fans, awards copy, restaurant blurbs and anyone describing a holiday they loved. Ott and colleagues measured the pattern as a deception signal in human text in 2011, which is a claim about effort and invention rather than about tooling, and the same paper reports human judges at roughly chance when they try to call individual reviews. A reviewer should look for one physical specific instead: a floor number, a dish, a time, a name, a street. A passage that stacks superlatives and also contains such details is a happy customer. A passage with nine superlatives and nothing a person could have seen is worth reading twice.","model_attribution":"Not a model signal. The peer-reviewed source predates chat products by over a decade and studies human writers. vale-ai-tells lists the phrases as assistant output as well, with no measurement, so the honest reading is that the pattern marks low-effort praise regardless of who produced it.","platform_notes":[{"platform":"amazon","note":"Fake-review detection there is described as behavioural and graph-based, with no stylistic criterion published. Style is the amateur method on this surface; volume, coordination and payment signals are what the platform describes using."}],"languages":["en"],"sources":[{"kind":"external","title":"Ott et al., Finding Deceptive Opinion Spam by Any Stretch of the Imagination, ACL 2011","url":"https://aclanthology.org/P11-1032/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"real-question-pivot","name":"The real question pivot","aka":["The real question is","the harder truth","the uncomfortable reality"],"category":"rhetorical","subcategory":"pivot","description":"The stated topic is set aside for a claimed deeper one. slop-lint catalogues the framing family and the real-question pivot inside it, and vale-ai-tells ships a matching rule. One reframe per essay is craft, so density is the whole signal here: the rule fires only when two pivots appear within about 600 characters of each other.","why_it_reads_ai":"The move claims depth without paying for it. Dismissing the question you were asked is easier than answering it, and the substituted question is always more abstract, so it needs no evidence either. Two in a paragraph means the writer never intended to answer anything.","examples":[{"before":"The real question is whether any of this scales. But the harder truth is that scale was never the constraint, and everyone in the room usually knows it well before anyone is willing to say so.","after":"We ran the same workflow at twelve clients and at ninety. It broke at forty, when one person could no longer read every draft before it went out.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"(?:\\bthe\\s+(?:real|deeper|harder|bigger|uncomfortable|honest|quiet)\\s+(?:question|truth|answer|reality|tension|issue)\\s+(?:is|here\\s+is)\\b[\\s\\S]{0,600}?){2}","flags":"gi","scope":"document"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Opened as active. slop-lint carries the framing family from a screened human baseline and vale-ai-tells ships a matching rule. Neither publishes a rate, so the density requirement of two pivots in one passage is a FeedSquad decision stated in the pattern rather than a measured cutoff."}],"evidence_grade":"community-observed","false_positive_notes":"Essayists, columnists and philosophers make one genuine reframe per piece, and the whole point of an essay is often that the obvious question was the wrong one. Socratic teaching is built from the move. Interviewers use it to open a subject up. The check is arithmetic rather than aesthetic: count the pivots. One is a thesis. Two in a paragraph means neither question was answered, and by the second one the reader has been told twice that the real subject is elsewhere without ever being taken there.","model_attribution":"Unattributed. slop-lint records the family from reply-register observation and names no product; vale-ai-tells does not separate model families.","platform_notes":[{"platform":"x","note":"The reply register is where slop-lint found the family. X handles its llm_slop_post label as spam with a thirty-day window, and the classifier prompts behind it are withheld, so nothing published shows this framing being scored."}],"languages":["en"],"sources":[{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"bold-lead-list-items","name":"Bolded lead-in on every list item","aka":["bold lead-ins","inline-header vertical lists","key takeaways bullets","WP:AILIST","inline-header vertical list","bolded run-in head","key-takeaways list","key takeaways list","label-and-explain bullets","bold-first bullets"],"category":"formatting","subcategory":"list-shape","description":"Every item in a vertical list opens with a bolded phrase followed by a colon and an explanation. The bolded fragment acts as an inline header, so the list reads as a glossary rather than as a sequence of points.","why_it_reads_ai":"The shape is a default in assistant output: models bold the lead term of each enumerated item unless told otherwise, and Anthropic and OpenAI both publish guidance instructing their models to stop doing it in prose contexts. Three or more consecutive items in that shape is the signal, not any single bolded phrase.","examples":[{"before":"**Consistency:** Post regularly so your audience knows when to expect you. **Authenticity:** Share real experiences rather than generic advice. **Engagement:** Reply to comments to build community.","after":"We post on Tuesdays and Thursdays because that is when the support inbox is quietest and someone can actually answer the replies. When we tried five days a week, comments sat unanswered for two days and the threads died.","note":"The three labels in the before block are interchangeable with any other three. The after paragraph names the constraint that produced the schedule, which no label could carry."},{"before":"**Step 1:** Open the settings panel. **Step 2:** Choose your export format. **Step 3:** Click Export.","after":"Open the settings panel and choose an export format. CSV keeps the raw timestamps; the PDF option rounds them to the nearest minute, which matters if you are reconciling against billing.","note":"This pair is here as a warning. Numbered procedure steps are legitimate technical writing, and the repair is only an improvement because the original steps carried no information the buttons did not already carry."},{"before":"- **Clarity:** say what you mean.\n- **Brevity:** cut what you can.\n- **Consistency:** keep the same voice throughout.","after":"We cut the opening paragraph from every post after watching where readers stopped. The target length went from 400 words to 180. Nothing in the shorter version needed the paragraph we removed.","note":"The specimen has three headings and no facts. The repair drops the headings and keeps what happened."}],"detection":{"type":"deterministic","pattern":"(?:^[ \\t]*(?:[-*+•]|\\d{1,2}[.)])[ \\t]+\\*\\*[^*\\n]{1,60}?(?:[:：]\\*\\*|\\*\\*[ \\t]*[:：—–-])[^\\n]*\\n){2}^[ \\t]*(?:[-*+•]|\\d{1,2}[.)])[ \\t]+\\*\\*[^*\\n]{1,60}?(?:[:：]\\*\\*|\\*\\*[ \\t]*[:：—–-])","flags":"gm","scope":"document"},"severity":"low","status":"contested","status_history":[{"date":"2026-08-14","status":"contested","rationale":"Supported at model level by the ICML 2025 idiosyncrasies study and by Anthropic's dated system prompts, which instruct against exactly this shape. Undercut by OpenAI's documented API default of no Markdown, which relocates the tell from a model family to a product surface. Technical documentation has used definition-style bullets since long before 2022, so the base rate in the affected genres is high."}],"evidence_grade":"corroborated","false_positive_notes":"Technical writers use definition lists, and the major documentation style guides sanction a bolded term followed by its description. Release notes and API references are built from the shape. Recipe writers use it, and anyone drafting in a Markdown editor with a formatting toolbar produces it without deciding to. Wikipedia's list of ineffective indicators cuts the other way too: when a writer's pre-2022 work already shows the same formatting habits, that continuity is evidence against AI rather than for it.","model_attribution":"ChatGPT surface, hedged. The idiosyncrasies study describes bolding inside enumerations for ChatGPT and minimal formatting for Claude, measured on dated API snapshots that no longer correspond to shipping models. OpenAI documents no Markdown by default in the GPT-5 API, so the same weights behave differently through a different door.","platform_notes":[{"platform":"linkedin","note":"LinkedIn's stated target is posts that restate without adding. Formatting is not mentioned in the policy, and the penalty is reduced distribution outside a person's network rather than removal."},{"platform":"x","note":"The published X ranking code carries a post-level slop label and an account-level spam label, neither tied to formatting. Author diversity decay, which reduces the score of each additional post from the same account, is the closer analogue to a volume penalty."},{"platform":"wikipedia","note":"Shortcut WP:AILIST. Wikipedia notes the list markers are often literal characters rather than wiki markup, which is a stronger signal than the bolding itself."}],"languages":["en"],"sources":[{"kind":"external","title":"Sun, Yin, Xu, Kolter, Liu: Idiosyncrasies in Large Language Models (ICML 2025)","url":"https://arxiv.org/abs/2502.12150","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Anthropic: published Claude system prompts, release notes (formatting instructions by model and date)","url":"https://platform.claude.com/docs/en/release-notes/system-prompts","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"OpenAI Cookbook: GPT-5 prompting guide (API does not format final answers in Markdown by default)","url":"https://developers.openai.com/cookbook/examples/gpt-5/gpt-5_prompting_guide","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"Wikipedia: Signs of AI writing, sections on inline-header vertical lists and ineffective indicators","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Lorenzetti: Keeping conversations real on LinkedIn (LinkedIn policy announcement, 2026-05-20)","url":"https://www.linkedin.com/pulse/keeping-conversations-real-linkedin-laura-lorenzetti-9821e","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"xai-org/x-algorithm: published ranking and abuse-enforcement code (llm_slop labels, author diversity)","url":"https://github.com/xai-org/x-algorithm","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-14","updated":"2026-08-15"},{"id":"emoji-as-section-marker","name":"Emoji used as document structure","aka":["emoji bullets","emoji headers","emoji as formatting","WP:AIEMOJI","emoji as bullets","emoji-prefixed benefit list"],"category":"formatting","subcategory":"decorative-symbols","description":"Emoji placed at the front of headings or list items, carrying the document's structure rather than any meaning. Wikipedia's editors list the habit under style and note it turned up mostly in talk-page comments and edit summaries. A single emoji is not the pattern. The pattern is emoji doing the job of a bullet across a whole document. The threshold is a convention: below three in ten, an emoji reads as an accent, and at or above it the emoji is carrying the outline. No published measurement sets this number, and the entry does not pretend otherwise.","why_it_reads_ai":"Chat surfaces reward visible structure, and an emoji is a cheap section marker that survives a plain-text field where markdown does not. The Washington Post analysed 328,744 shared ChatGPT messages and found about seventy percent carried at least one emoji by July 2025. That figure measures presence in chat replies rather than structural use in published prose, so it supports the habit's existence and cannot set the threshold. Anthropic instructs Claude to skip emoji unless the person used one first, which is a vendor writing down a tendency it wants suppressed.","examples":[{"before":"🚀 Launch week is here!\n✅ New dashboard\n🔥 Faster exports\n💡 Smarter alerts","after":"Launch week. Exports that used to take four minutes now finish in under one. The dashboard changed too, but the export time is the only thing anyone has written to us about.","note":"Four emoji across four lines in the specimen. Every one of them is standing in for a bullet."},{"before":"📌 Key takeaway: consistency beats intensity.\n📈 Growth compounds.\n🎯 Focus on one channel.","after":"We posted three times a week for four months instead of daily for three weeks. The daily run produced more total reach in month one and nothing after it. The slower run was still growing in month four."}],"detection":{"type":"statistical","metric":"share-of-headings-and-list-items-opening-with-an-emoji","threshold":0.3,"direction":"above","threshold_basis":"No published measurement of emoji-as-structure rates in human or model text exists, so this threshold has no external basis. It is a FeedSquad review trigger: below roughly a third of headings and list items, emoji read as occasional decoration rather than as the document structure itself. Treat it as a prompt to look, never as a finding."},"severity":"low","status":"contested","status_history":[{"date":"2026-08-14","status":"contested","rationale":"Opened as contested. Wikipedia lists the habit and says it is rarer now than it was, which points toward decay. The Washington Post measurement of emoji presence in chat messages points the other way, but measures a different thing. The human base rate on social platforms is enormous and long-standing, so the pattern cannot support a verdict about authorship on those surfaces."}],"evidence_grade":"community-observed","false_positive_notes":"Newsletter and community writers have used emoji as visual anchors since well before assistants could write, and the habit is a house style in plenty of developer changelogs and release notes. Writers who use screen readers deliberately avoid it, so absence of emoji says nothing either. Judge the pattern only where one emoji prefixes every section in a document that has no other visual structure.","model_attribution":"Unattributed at family level. Wikipedia lists it as a general sign and says it has become rarer. One research pass found no primary source assigning an emoji habit to any OpenAI, Google, Meta, or DeepSeek model. The only vendor-level documentation is Anthropic's, and it is a suppression instruction rather than a measurement of output.","platform_notes":[{"platform":"linkedin","note":"Emoji bullets are ordinary LinkedIn craft, so this rule is close to useless there as an authorship signal. LinkedIn's May 2026 policy suppresses out-of-network distribution for content with no unique perspective and leaves the post visible to connections."},{"platform":"instagram","note":"The caption field offers no formatting at all, so emoji separators are a workaround for a product limitation rather than a signal about who wrote the caption."},{"platform":"wikipedia","note":"Listed under style with shortcut WP:AIEMOJI. The guide records that emoji appeared almost always in talk-page comments and edit summaries, and that they are rarer now."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia:Signs of AI writing, section Emoji as formatting (WP:AIEMOJI)","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Merrill, Chen and Kumer, What are the clues that ChatGPT wrote something? We analyzed its style. Washington Post (2025-11-13)","url":"https://www.washingtonpost.com/technology/interactive/2025/how-detect-chatgpt-em-dash/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Anthropic, published Claude system prompts and release notes (emoji instruction)","url":"https://platform.claude.com/docs/en/release-notes/system-prompts","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-14","updated":"2026-08-15"},{"id":"raw-markdown-leak","name":"Raw markdown in a surface that does not render it","aka":["unrendered markdown","asterisk bold in plain text","literal markdown syntax","WP:MARKDOWN","code fence leaked into an article"],"category":"formatting","subcategory":"markdown-leak","description":"Markdown syntax appears verbatim in a surface that renders none of it. Asterisks, hash headings, bracket links and backticks survive the paste into LinkedIn, X, or an email body, where they display as literal characters.","why_it_reads_ai":"Models emit markdown by default because their training and their chat surfaces both render it. A person composing in the platform composer has nothing to convert. The artifact locates where the text was written, which is upstream of the platform it landed on.","examples":[{"before":"Read the full breakdown here: [why retention really matters](https://example.com/post)\n\n## What we learned about churn\n\nRetention is one of those subjects everybody agrees is important and almost nobody looks at properly. Churn is a lagging measure of a hundred small decisions, most of them made long before anyone thinks to open the dashboard, and by the time the shape of it is obvious the window to do anything about it has usually closed. Better reporting is rarely what fixes that. It is an easy thing to say and a hard thing to keep up.","after":"Our Q3 retention teardown is on the blog, linked below. The short version: churn moved because we changed the trial length, not because of the pricing test we assumed was responsible.","note":"The first version was composed in a markdown editor and pasted into a composer that renders none of it. Readers see the brackets and hashes. The tell is the unrendered syntax, not the empty copy around it. The copy is empty so that the specimen reads as what the artifact implies: a draft nobody reread before publishing."},{"before":"Three things changed this month. The **billing migration** finally shipped, support volume settled down, and the team grew a little. None of it was dramatic, and all of it took longer than anyone expected it to. That is usually how the useful months go.","after":"Three things changed this month. The billing migration shipped on the 14th after two false starts, support volume dropped by roughly a third in the fortnight after, and Priya joined as our second engineer.","note":"Asterisks around a single phrase mid-sentence, left over from a composer that rendered them."}],"detection":{"type":"deterministic","pattern":"(?:^|\\s)(?:#{1,6}\\s+\\S|\\[[^\\]]{1,60}\\]\\([^)]{1,120}\\)|`{1,3}[^`\\n]{1,60}`{1,3}|\\*\\*(?!\\s)[^*\\n]{0,59}[^*:\\s\\n]\\*\\*)","flags":"gm","scope":"document"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-14","status":"active","rationale":"High precision as an artifact and honest about what it proves. It identifies where the text was composed rather than who composed it, which is why severity is moderate and not strong. Regex verified on 2026-08-14 against three positive and five negative strings, including a Python exponent and a thematic break."}],"evidence_grade":"primary-doc","false_positive_notes":"Anyone who drafts in a markdown editor produces this by hand. Software engineers write markdown as a default register and paste out of it without thinking, and note-taking tools store markdown natively, so their export path is copy and paste. Technical writers, people who live in issue trackers, and anyone who drafts in a static-site repo all hit this the same way. Some publishing tools strip the markers and some do not, so one author can produce the artifact on one platform and not on another with identical source text. The pattern locates the composer. It says nothing about the author.","model_attribution":"Documented at surface level, not family level. The ICML 2025 idiosyncrasy work reports that ChatGPT emphasises key points inside enumerations with bold while Claude often returns responses with no bold and no headers, and OpenAI documents that GPT-5 in the API emits no Markdown by default to preserve compatibility. Attribute the habit to a chat surface with a renderer and a system prompt, not to a model family.","platform_notes":[{"platform":"linkedin","note":"The post composer stores plain text, so asterisks publish as asterisks. LinkedIn retired its own generative rewrite tool in July 2026 and replaced it with a proofreader described as not changing the writer's voice, which is the opposite direction of travel from pasted chat output."},{"platform":"x","note":"Post bodies are plain text. Literal asterisks survive to the timeline and read as leftover markup."},{"platform":"wikipedia","note":"The editor guide treats markdown in an article as a sign in itself, because wikitext is thinly represented in model training data. It names the inline-header vertical list specifically: list marker, bold inline header, colon, descriptive text."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia:Signs of AI writing (sections: Inline-header vertical lists; Use of Markdown)","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"OpenAI Cookbook: GPT-5 prompting guide (API default emits no Markdown)","url":"https://developers.openai.com/cookbook/examples/gpt-5/gpt-5_prompting_guide","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"Sun, Yin, Xu, Kolter, Liu. Idiosyncrasies in Large Language Models. ICML 2025","url":"https://arxiv.org/abs/2502.12150","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-14","updated":"2026-08-17"},{"id":"em-dash-density","name":"Em-dash density","aka":["dash addiction","WP:AIDASH","em dashes per thousand words"],"category":"formatting","subcategory":"punctuation","description":"Em dashes arrive at a rate well above what the genre and the surface usually carry. The count is the unit here, not the mark. Spacing is a separate entry and this one measures density only.","why_it_reads_ai":"Model output uses the mark more often than non-professional human writing of the same genre, and puts it where a comma or a full stop would do the same work. The population evidence is solid and narrow. Czuma measured em dash presence in medRxiv Discussion sections rising from 4.23 percent of documents before ChatGPT to 11.58 percent after, odds ratio 2.96 with a confidence interval of 2.77 to 3.17, reaching 20.3 percent by 2025. That author states the measure is not a per-paper detector. A count on one document tells a reviewer where to look and nothing about who typed it.","examples":[{"before":"The rollout went well—better than we planned—and support tickets dropped. Our onboarding flow—short and honest about limits—did most of the work. The lesson—start smaller—holds for any team.","after":"The rollout went well and support tickets dropped by about a third in the first fortnight. The onboarding flow is four screens long, and screen two says that imports over 50MB will fail. That sentence killed most of our old tickets.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"em-dashes-per-1000-words","threshold":10,"direction":"above","threshold_basis":"Ten em dashes per 1,000 words is a review trigger, not a finding, and four things about its provenance ship with it. Freeburg is a single-author preprint whose human baseline is eight essays. The abstract runs from 0.0 for Llama to 9.1 for GPT-4.1 under suppression, so the trigger sits above the top of the abstract-verified model range. The unconstrained model top figure and the human mean are body-level table values this project could not confirm at page level, and a Crossref search for the same measurement in peer-reviewed venues returned nothing, which is how a preprint comes to stand alone here. The false-positive rate is unknown and no baseline exists for marketing or social copy. Czuma supplies the verified population shift, 4.23 percent to 11.58 percent of medRxiv Discussion sections with an odds ratio of 2.96, and states it is not a per-paper detector."},"severity":"medium","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Opened as contested. The population shift is measured and holds up: Czuma reports em dash presence in medRxiv Discussion sections rising from 4.23 to 11.58 percent with an odds ratio of 2.96, and its author says the measure does not work per paper. Against that, the published human range of 0.33 to 17.12 per 1,000 words contains every model mean measured so far, two of twelve models emit none at all, and OpenAI shipped suppression in GPT-5.1 in November 2025. A per-document verdict is not available from any of it."}],"evidence_grade":"corroborated","false_positive_notes":"Essayists and literary writers have leaned on the mark for two centuries, and the published human range of 0.33 to 17.12 per 1,000 words contains every model mean anyone has measured. Two of the twelve models Freeburg tested emit none at all, so a low count says nothing either. Typographers, anyone drafting in a tool that auto-converts a double hyphen, and writers trained on Chicago style all produce high counts by preference. A reviewer who wants to use this number needs the writer's own earlier work as the baseline, because no baseline exists for marketing or social copy anywhere in the literature.","model_attribution":"Reported per model, and the ordering moves. Freeburg's twelve-model table puts a Claude model and GPT-4.1 near the top and Llama at zero. The Economist reported in July 2026 that among contemporary models only Claude exceeded professional writers and that ChatGPT used fewer; this project read that claim second-hand through Wikipedia and labels it reported. OpenAI shipped em dash suppression in GPT-5.1 in November 2025, so any per-family claim needs a date attached to it.","platform_notes":[{"platform":"linkedin","note":"The claim that LinkedIn penalises em dashes has no first-party source. No LinkedIn policy or newsroom post this project read mentions punctuation at all."},{"platform":"wikipedia","note":"Shortcut WP:AIDASH. The guide says the sign is most useful alongside other indicators and that it turns up far more on discussion pages than in article text."}],"languages":["en"],"sources":[{"kind":"external","title":"Czuma, Em-ergence of the em-dash: a population-level rise in em-dash frequency in medRxiv preprints, measured on Discussion sections (arXiv:2606.29540)","url":"https://arxiv.org/abs/2606.29540","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Freeburg, em dash rates across twelve models, single-author preprint (arXiv:2603.27006)","url":"https://arxiv.org/abs/2603.27006","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"The Economist, how to spot AI writing (30 Jul 2026)","url":"https://www.economist.com/culture/2026/07/30/how-to-spot-ai-writing","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Washington Post interactive, how to detect ChatGPT em dashes","url":"https://www.washingtonpost.com/technology/interactive/2025/how-detect-chatgpt-em-dash/","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"space-surrounded-em-dash","name":"Space-surrounded em dash","aka":["spaced em dash","word space dash space word"],"category":"formatting","subcategory":"punctuation","description":"Em dashes set with a space on each side, repeatedly, in a surface whose house style closes them up. Wikipedia treats this as the better-specified form of the dash tell. Density is the unit here too, so one spaced dash is a typing choice.","why_it_reads_ai":"Most people who reach for the mark learned a house rule with it and apply that rule every time. Model output tends to space the mark whatever surface it lands on, which is why the habit shows up in plain-text fields where nothing enforced it. The signal lives in the mismatch between the spacing and the convention of the place, so a reviewer who cannot name the convention has nothing to measure against.","examples":[{"before":"We shipped the new importer late — later than any of us wanted — and nobody complained. Shipping late is something every team learns to live with — usually more than once — and it stops feeling like a crisis somewhere around the third time. The work landed, and landing is most of what anyone remembers a year later.","after":"We shipped the new importer three months late, on 14 July, and nobody complained. The delay came down to the CSV parser, which we rewrote twice after it choked on the semicolon-delimited exports our largest account sends every Monday.","note":"Figures in this repair are invented for the specimen. The tell is the spacing around the mark, not the empty prose it sits in. The prose is empty so a reader is not left thinking the entry objects to a well-made sentence."}],"detection":{"type":"deterministic","pattern":"\\S[ \\u00A0]\\u2014[ \\u00A0]\\S[\\s\\S]{0,400}?\\S[ \\u00A0]\\u2014[ \\u00A0]\\S","flags":"g","scope":"document"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active because it is the one dash check that survives the density argument. Wikipedia records that model dashes are usually surrounded by spaces, against the typographic guidance most human users of the mark already know. The pattern asks for two spaced dashes rather than one, so a single typing choice does not register. Regex verified on 2026-08-15 against its own specimen and against three negatives, one of them a spaced en dash in the British convention."}],"evidence_grade":"community-observed","false_positive_notes":"Associated Press style spaces its dashes by house rule, and much British publishing sets a spaced en dash where American publishing closes up an em dash. Plenty of newsrooms and newsletter templates apply their own spacing on the way to the page, so the author never chose it. Anyone drafting in an editor that auto-spaces produces the artifact without a decision. A reviewer separates the two by finding the publication style sheet first and then checking whether the same writer spaced dashes before 2022. Consistent spacing across years is a house habit; spacing that appears only in recent drafts is worth a second look and nothing stronger.","model_attribution":"Undocumented at family level. Wikipedia records the spacing habit without assigning it to a vendor, and no measurement of dash spacing by model exists in anything this project read.","platform_notes":[{"platform":"linkedin","note":"The composer stores plain text and applies no typographic conversion, so whatever spacing the writer pasted is what publishes."},{"platform":"x","note":"Post bodies are plain text. Spacing survives to the timeline exactly as it was typed, which makes the surface a clean place to observe the habit and a poor place to judge anyone by it."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"double-hyphen-dash","name":"Double-hyphen dash","aka":["ASCII double hyphen in prose"],"category":"formatting","subcategory":"punctuation","description":"Two ASCII hyphens standing in for a dash in rendered text. The form survives when the composer does not convert it and nobody read the published version.","why_it_reads_ai":"It is the same pivot habit as the dash tell with the character unavailable. Two taxonomies record it separately, tropes.fyi as a named pattern and vale-ai-tells as a rule with its own repair message. Both write it narrowly, because a double hyphen at the front of a command-line flag is ordinary and correct.","examples":[{"before":"The migration took two weeks--twice what we budgeted--and the lessons were worth the delay. Every migration teaches a team something about its own systems that no amount of planning surfaces in advance. What matters is that we came out of it better prepared for the next one.","after":"The migration took two weeks against a one-week budget, and we never tested the rollback plan. That is why the Tuesday outage ran to 40 minutes instead of the five we told the board it would.","note":"Figures in this repair are invented for the specimen. The tell is the pair of hyphens, not the emptiness around them. The earlier version of this specimen admitted a real failure with a real consequence, which made the entry look like it was policing punctuation in good writing."}],"detection":{"type":"deterministic","pattern":"(?:[A-Za-z]--[A-Za-z]|[A-Za-z] -- [A-Za-z])","flags":"g","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active and cheap. Two independent taxonomies list the form, and the pattern is anchored to a letter on both sides so that command-line flags, which begin with a space and two hyphens, never register. Regex verified on 2026-08-15 against its own specimen and against three negatives including a flag and a chained flag argument."}],"evidence_grade":"community-observed","false_positive_notes":"Writers who learned on typewriters and plain-text email type two hyphens for a dash by muscle memory and have done so since long before any of this. People who draft in Markdown or reStructuredText type it because their toolchain converts it on render. Technical documentation is full of it as a flag prefix, and a sentence about a flag will carry it correctly. Some editors convert on the fly and some do not, so one writer produces the artifact in one tool and not in another with identical source text. A reviewer should treat it as a note about the composer and the review pass, never about the author.","model_attribution":"Unassigned. Neither taxonomy that lists the form attributes it to a model family, and no measurement of hyphen substitution by vendor exists in anything this project read.","platform_notes":[{"platform":"x","note":"Plain-text bodies with no conversion step, so the two hyphens publish exactly as typed and read as leftover markup."}],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"title-case-headings","name":"Title Case headings","aka":["WP:AITITLECASE","Every Word Capitalized"],"category":"formatting","subcategory":"headings","description":"Headings capitalise every main word in a publication whose house style is sentence case. The tell exists only against a named convention, which is why this entry stays contested.","why_it_reads_ai":"Wikipedia records that chatbots strongly tend to capitalise all main words in section headings, and cites the 2025 work on people who use these tools heavily and recognise their output. The habit comes out of documentation, slide decks and listicles in the training material, where headline capitalisation is the norm. Set against a publication that writes headings in sentence case, the mismatch is visible on the page. Set against a United States editorial style sheet, there is nothing to see.","examples":[{"before":"Impact Of Technology On Regional Growth\n\nDigital tooling has changed how the region thinks about growth, and much of that change took place before anyone had settled on a name for it.\n\nSustainable Development And Local Policy\n\nOfficers will bring a further report to committee in due course. The authority remains committed to an approach that balances growth with the needs of existing communities, and further engagement with stakeholders is anticipated in the usual way. A number of options remain under consideration at this stage.","after":"Regional growth after the 2024 reporting change\n\nCouncils began publishing quarterly figures in March 2024. Two of the eleven councils missed the first deadline, and both blamed the same spreadsheet template, which rounded part-year revenue to the nearest million.","note":"Figures in this repair are invented for the specimen. The tell is the capitalisation of the heading lines, which is why the body under them carries nothing a reader could check: an index card flattens the line breaks, so the prose has to read as machine-made on its own."}],"detection":{"type":"deterministic","pattern":"^[ \\t]*[A-Z][A-Za-z0-9]*(?:[ \\t]+(?:[A-Z][A-Za-z0-9]*|of|the|and|in|for|to|a|an|or|with|on)){4,}[ \\t]*$","flags":"gm","scope":"document"},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Contested by construction. Wikipedia records the tendency and cites the 2025 study of heavy tool users, so the pattern is real in that corpus. It is also mandated by American editorial style, produced automatically by content management systems, and baked into slide templates, so the base rate in the affected genres is enormous. The pattern requires a standalone line of five or more words with no terminal punctuation, so running prose never registers. Regex verified on 2026-08-15 against its own specimen and four negatives."}],"evidence_grade":"corroborated","false_positive_notes":"American editorial style mandates headline capitalisation, and the Chicago Manual of Style, wire-service headline conventions and most United States magazines all require it. Marketing teams inherit it from slide templates and content management systems that title-case automatically, so nobody chose it. Non-native English writers taught from American style guides apply it consistently. A reviewer has to find the house style sheet before treating a capitalised heading as anything: this is a signal only where sentence case is the stated rule and where the same writer used sentence case in earlier work. Everywhere else it is a preference.","model_attribution":"Unassigned at family level. Wikipedia describes the tendency as general to chatbots and cites the 2025 study of heavy tool users for it. No paper this project read breaks heading capitalisation down by vendor.","platform_notes":[{"platform":"wikipedia","note":"Shortcut WP:AITITLECASE. The manual of style calls for sentence case in headings, which is what makes the mismatch legible there and illegible almost everywhere else."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Russell, Karpinska, Iyyer: People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (ACL 2025)","url":"https://aclanthology.org/2025.acl-long.267/","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"boldface-in-running-prose","name":"Boldface in running prose","aka":["WP:AIBOLD","key-term bolding","emphasis carpet-bombing"],"category":"formatting","subcategory":"emphasis","description":"Chosen terms bolded through ordinary paragraphs with no list anywhere near them. Every appearance of a phrase gets the same treatment, so the emphasis stops marking anything. Bolded openers inside list items belong to a separate entry and the specimens are kept apart.","why_it_reads_ai":"Wikipedia files the habit under WP:AIBOLD and traces it to readmes, fan wikis, how-tos, sales pages, slide decks and listicles in the training material, all of which bold heavily. Anthropic's dated system prompts instruct Claude away from over-formatting, which is a vendor writing down the pull it wants suppressed. Count rather than judge. One bolded term in a long piece is editing.","examples":[{"before":"The **billing migration** was the quarter that taught us the most. **Billing** touches everything we do, so the **migration** was never going to be simple, and the **team** knew that going in. What we came away with is that preparation matters more than pace, which is the sort of lesson most teams end up learning twice.","after":"The billing migration was the quarter's largest project. It touches every subscription we hold, so it had to run without downtime. We rehearsed it twice against a copy of production and hit the same ordering bug both times, which is why the live run took 11 minutes rather than the 40 we had planned for.","note":"Figures in this repair are invented for the specimen. The tell is the repeated emphasis, not the emptiness underneath it. The specimen is written in asterisks because a JSON field cannot carry rendered bold, which is why the raw-markdown-leak rule also matches it."}],"detection":{"type":"statistical","metric":"bold-spans-per-1000-words-outside-lists","threshold":6,"direction":"above","threshold_basis":"No published human baseline for bold density in prose exists in anything this project read. The one measurement located is a suppression experiment on models rather than a human comparison: Freeburg reports markdown features dropping to roughly zero for eleven of twelve models when they are instructed to write prose, while em dashes persist. Six bold spans per 1,000 words outside lists is therefore a FeedSquad review trigger and not a published cutoff. A reading above it means read the paragraph."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active. The vendor evidence is first-party and dated: Anthropic publishes system prompts that instruct against over-formatting, which documents the tendency without measuring output. Wikipedia records the same habit in article space and names the training genres it comes from. The threshold is ours and says so, because no human baseline for bold density has been published."}],"evidence_grade":"primary-doc","false_positive_notes":"Textbooks, revision guides and exam material bold keywords on purpose, and so do plenty of developer documentation house styles. Product marketers bold the one term a skimmer has to catch. Newsletter and content management templates apply bold to a first clause automatically, so the writer never chose it. A reviewer separates the two by asking what removing every bold span would cost the reader: deliberate key-term bolding marks terms the piece goes on to define, while the habit marks the same phrase every time it appears and defines nothing.","model_attribution":"Anthropic documents an instruction against over-formatting in dated system prompts, which is first-party evidence of the pull rather than a measurement of what ships. OpenAI documents that GPT-5 in the API emits no Markdown by default, so one set of weights bolds through one door and not another. Attribute the habit to a surface and a date, never to a family on its own.","platform_notes":[{"platform":"linkedin","note":"The composer has no bold at all, so asterisks publish as asterisks. That artifact is the raw markdown entry rather than this one."},{"platform":"wikipedia","note":"Shortcut WP:AIBOLD. The manual of style already restricts boldface in article text, so the habit reads as a style violation there before it reads as anything else."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Anthropic dated system prompts, release notes","url":"https://platform.claude.com/docs/en/release-notes/system-prompts","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"anti-slop-writing system prompt and pattern list","url":"https://github.com/adenaufal/anti-slop-writing","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"curly-quotes-in-plain-text","name":"Curly quotes in plain-text registers","aka":["smart quotes","WP:AICURLY","typographic apostrophes"],"category":"formatting","subcategory":"typography","description":"A text carries the full typographic set, curly quotation marks together with curly apostrophes, in a register whose composer produces straight ones. Wikipedia records the habit for two model families and states that two others typically do not produce it.","why_it_reads_ai":"Chat surfaces render typographic characters, and text copied out of one keeps those characters when it lands in a field that would never have made them. The artifact locates the composer. It is heavily confounded by word processors and phone keyboards, which is why this entry ships as contested and why the pattern asks for the full set rather than a single mark.","examples":[{"before":"We asked our customers what “good support” actually means to them, and the answers weren’t surprising so much as familiar. People want to feel heard, and they want to feel it quickly. Every company says it already does this. The ones that genuinely do rarely need to say so.","after":"We asked forty customers what \"good support\" meant. Twenty-three said the same thing: talking to the same person twice. We changed the routing rules in week two so a reopened ticket goes back to whoever closed it, and reopen volume fell by half in August.","note":"Figures in this repair are invented for the specimen. The tell is the full typographic set in a plain-text field, not the emptiness. The specimen was hollowed because the earlier version carried a real research finding, which made the entry read as a complaint about apostrophes in decent writing."}],"detection":{"type":"deterministic","pattern":"[\\u201C\\u201D][^\\n]{0,200}[\\u2018\\u2019]|[\\u2018\\u2019][^\\n]{0,200}[\\u201C\\u201D]","flags":"g","scope":"document"},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Contested from the start. Wikipedia lists the habit and immediately lists the confounds, naming Microsoft Word, the macOS and iOS defaults, Chicago style and citation tools, and stating that two widely used model families typically do not produce curly marks at all. The pattern was narrowed to require a curly quotation mark and a curly apostrophe together, so a lone autocorrected apostrophe does not register. Regex verified on 2026-08-15 against its own specimen and three negatives."}],"evidence_grade":"corroborated","false_positive_notes":"Microsoft Word, the macOS and iOS default keyboards and grammar tools such as LanguageTool convert straight marks to curly ones without being asked, so anyone drafting in them produces the artifact by default. Directional marks are correct under Chicago style and standard in typeset books and major newspapers, and citation tools copy them straight out of page titles. Some fonts render a curly apostrophe as a straight one, so the distinction can be invisible to the very reader making the judgement. Wikipedia's own caution is the one to keep: curly marks alone establish nothing, and two widely used model families do not produce them.","model_attribution":"Wikipedia attributes the habit to ChatGPT and DeepSeek specifically and states that Gemini and Claude models typically do not use curly quotes. slop-lint reports roughly nine curly marks per thousand words for GPT and Grok output and none for Claude, which is one repository's measurement and should be read as one.","platform_notes":[{"platform":"linkedin","note":"The composer stores whatever characters were pasted, so a draft written on a phone and a draft pasted out of a chat window are indistinguishable on this signal alone."},{"platform":"x","note":"Plain-text bodies, no conversion. The marks survive, which makes the surface easy to observe and useless for a verdict about any single account."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"decorative-horizontal-rules","name":"Decorative horizontal rules","aka":["thematic break spam","rule between every section"],"category":"formatting","subcategory":"dividers","description":"A rule drawn between every section of a piece short enough not to need one. The count of rules against the count of sections is the tell, not the presence of a rule.","why_it_reads_ai":"Wikipedia records chatbots inserting a thematic break between each section of a text and notes that the habit comes out of Markdown output, where a break is cheap and looks like structure. In a short piece the rules outnumber the ideas, and the reader gets a set of boxes instead of an argument.","examples":[{"before":"What we changed this month\n\n----\n\nWe moved the export job to a queue.\n\n----\n\nSupport tickets about timeouts stopped.\n\n----\n\nNext month we look at imports.","after":"We moved the export job to a queue after the 90-second timeout started firing on accounts with more than 20,000 rows. Support tickets about timeouts went to zero the week after. Imports are next, and they sit behind the same 90-second ceiling.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"^[ \\t]*(?:-{3,}|_{3,}|\\*{3,}|={3,})[ \\t]*$[\\s\\S]{0,800}?^[ \\t]*(?:-{3,}|_{3,}|\\*{3,}|={3,})[ \\t]*$","flags":"gm","scope":"document"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active and cheap to check. Wikipedia documents the habit with dated article examples in which a break sits under every section. The pattern requires two rules within 800 characters, so a single genuine scene break never registers. Regex verified on 2026-08-15 against its own specimen and three negatives, one of them a text carrying exactly one break."}],"evidence_grade":"community-observed","false_positive_notes":"Long essays and manuscripts use a rule for a real scene break, and one divider between the two halves of a piece is ordinary craft. Newsletter and email templates draw rules under headers and above footers automatically, and the writer never sees the markup. Documentation generators emit them between generated sections. A reviewer counts rules against sections: one break in a 3,000-word essay is punctuation, and a rule under every 80-word block is decoration standing in for structure the piece does not have. Judge the ratio and nothing else.","model_attribution":"Unassigned. Wikipedia describes the habit as common in Markdown output without naming a vendor, and no measurement of divider use by model family exists in anything this project read.","platform_notes":[{"platform":"wikipedia","note":"The guide keeps dated examples in which a break sits between every section of a draft, which is how the ratio became legible rather than the presence of one rule."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"tiny-tables","name":"Unnecessary tiny tables","aka":["WP:AITABLE","two-row table where prose would do"],"category":"formatting","subcategory":"tables","description":"A table built for two or three rows of information that was never tabular. The rows share no dimension, so the grid adds borders and nothing else.","why_it_reads_ai":"Wikipedia's editors record chatbots producing small tables where prose or an infobox would carry the same content, and the dated examples they keep show the giveaway: column headings reading Metric and Figure, holding items with no metric in common. A table is structure that can be produced without deciding anything. Prose forces a decision about what connects the rows.","examples":[{"before":"Our support numbers\n\n| Metric | Figure |\n| Tickets 2025 | about 4,100 |\n| Median first reply | 3 hours |\n| Languages covered | four |","after":"We handled about 4,100 tickets in 2025 with a median first reply of three hours. Four languages are covered. Finnish is the one where the median doubles, because one person answers it and she works Tuesdays to Thursdays.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Scan the text for tables with three or fewer data rows. For each one, check whether the rows share a single dimension a reader would want to compare across, such as the same measurement taken for different subjects, or one subject measured at different times. Escape hatches that mean no hit: a comparison of two named options across the same attributes; a schedule; a price list; a specification sheet; any table the surrounding prose refers to by column name. If the rows share no dimension and no sentence near the table points at it, return a hit and quote the heading row. Otherwise return no hit."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active but thin. The only support is one section of Wikipedia's style list, which says that in rare cases some models build unnecessary small tables and keeps dated examples. Severity is set to low to match single-source evidence, and the rubric carries four escape hatches so that correct small tables pass."}],"evidence_grade":"community-observed","false_positive_notes":"Comparison journalism and buyer's guides use small tables correctly, and two rows are enough when two rows are the whole comparison. Accessibility guidance sometimes prefers a short table to a sentence stuffed with parenthetical figures, and screen-reader users often prefer the table. Specification sheets, price lists and schedules are tabular by nature at any size. This entry rests on one section of one community style list, which is thin evidence, and the low severity is set to match rather than dressed up.","model_attribution":"Unassigned. Wikipedia's own wording is that in rare cases some models create unnecessary small tables. No study this project read measures table use by model family.","platform_notes":[{"platform":"wikipedia","note":"Shortcut WP:AITABLE. The guide keeps dated draft examples, and its own framing is that the behaviour is rare, which is worth repeating anywhere this entry is used."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"vendor-markup-artifact","name":"Vendor markup artifacts","aka":["oaicite","turn0search0","contentReference","grok_card","lenticular bracket leak","utm_source=chatgpt.com"],"category":"formatting","subcategory":"tool-leak","description":"A tool's internal citation token published as text. The marker was never meant to leave the product window, and it names the product that made it. What a hit establishes is that the draft went out without a read-through. Raw markdown syntax in a surface that does not render it is a separate entry, because a person drafting in a Markdown editor produces that by hand.","why_it_reads_ai":"These tokens come from a citation layer that renders as a link inside the product and as raw text everywhere else. Wikipedia's editors catalogue them per vendor, listing the ChatGPT reference and search markers, the Grok card tags and the DeepSeek bracket markers, and they treat the presence of one as grounds for speedy deletion under criterion G15, which covers pages published with no human review. Nobody types these by hand.","examples":[{"before":"Our numbers this quarter are in line with the sector average :contentReference[oaicite:4]{index=4}, which is broadly what we had modelled at this point in the cycle. Performance from here will be a function of execution rather than of market conditions, and execution is the variable we have the most control over. Full methodology is here: https://example.com/method?utm_source=chatgpt.com","after":"Our Q2 revenue grew 4 percent against a sector average of 6 percent, which is the first quarter we have trailed it. The methodology is on the blog, and the sector figure comes from the trade association survey of 210 firms published in May.","note":"Figures in this repair are invented for the specimen. The tell is the citation token, not the emptiness around it. The sentence it props up says nothing checkable, which is the state a draft is in when nobody read it to the end."},{"before":"Vacancy rates on the high street were about 6.2 percent last year【85†L261-269】, broadly in line with what observers had expected. Retail has been changing for a long time, in ways that are easier to describe after the fact than before it. Whether that is a recovery or a plateau depends largely on who is asked.","after":"Vacancy rates on the high street were 6.2 percent last year against a national average of 16 percent. The council publishes the count every March, and the 2026 figure is the first since 2019 that did not rise.","note":"Figures in this repair are invented for the specimen. The bracket form is the DeepSeek variant Wikipedia records."}],"detection":{"type":"deterministic","pattern":"(?::contentReference\\[oaicite:|oai_citation|\\bciteturn\\d|\\bturn\\d+search\\d|grok_render_citation_card_json|<grok_card|utm_source=(?:chatgpt\\.com|openai|copilot\\.com)|referrer=grok\\.com|\\u3010\\d+\\u2020L?\\d)","flags":"g","scope":"document"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active and high severity. Wikipedia catalogues the token families per vendor with dated article examples, and criterion G15 makes their presence grounds for speedy deletion because the page reached publication with no human review. The pattern is anchored to the full artifact rather than the bare word, so an article about these tokens does not register. Regex verified on 2026-08-15 against both specimens and three negatives, one of which mentions a token by name in ordinary prose."}],"evidence_grade":"corroborated","false_positive_notes":"Researchers, editors and pages like this one quote these tokens on purpose, so quoted and code-fenced examples have to be excluded before a hit counts. The tracking parameter is the softest member of the family: someone who copies a link out of a chat window and pastes it into a draft carries the parameter across without generating a word of the text, and analytics teams add their own source parameters that look much the same. Read a hit as a statement about the review process rather than about the writer. It says the draft reached publication without anyone reading it to the end.","model_attribution":"Per vendor, which is what makes this family legible. Wikipedia catalogues the ChatGPT reference and search markers, the Grok card tags and the DeepSeek bracket markers separately, and records the tracking parameter each product appends to links. The token names the tool. It says nothing about whether a person wrote the sentence around it.","platform_notes":[{"platform":"wikipedia","note":"Criterion G15 lists these markers among the signs that make a page eligible for speedy deletion, on the stated ground that it was generated and published with no human review."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"unfilled-placeholder","name":"Unfilled placeholders","aka":["[Your Name]","INSERT_SOURCE_URL","access-date=2025-XX-XX"],"category":"formatting","subcategory":"template-leak","description":"A template slot published with the slot still in it. The bracket text names the field nobody filled.","why_it_reads_ai":"Wikipedia records model output inserting placeholder dates into citation fields, most often the access date, and placeholder strings into source and publisher fields. Sign-offs arrive with the name bracket intact. This is not a claim about writing quality at all. It records that the draft was published without a read-through, which is the failure criterion G15 is written around.","examples":[{"before":"Thanks again for the call last week. I have attached the deck and the pricing sheet.\n\nBest regards,\n[Your Name]","after":"Thanks again for Tuesday's call. The deck is attached, and the pricing sheet now carries the 40-seat tier we talked about, on page 3.\n\nBest regards,\nAnni","note":"Figures in this repair are invented for the specimen."},{"before":"Source: INSERT_SOURCE_URL_30, publisher SOURCE_PUBLISHER, accessed 2025-XX-XX.","after":"Source: the Finnish Tax Administration guidance on VAT for digital services, published 12 March 2025 and read on 14 August 2026."}],"detection":{"type":"deterministic","pattern":"(?:\\[(?:Your|Insert|Client|Company|Product|Brand)[ _][A-Za-z ]{2,20}\\]|\\bINSERT_[A-Z][A-Z0-9_]{2,30}\\b|\\bSOURCE_(?:PUBLISHER|URL|TITLE)\\b|\\b\\d{4}-(?:XX|xx)-(?:XX|xx)\\b|\\b\\d{4}-\\d{2}-(?:XX|xx)\\b)","flags":"g","scope":"document"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active and high severity. Wikipedia keeps dated examples of citations published with an unresolved access date and with source and publisher slots left as literal strings, and criterion G15 treats the same class of artifact as grounds for speedy deletion. The pattern is anchored to bracketed slot names, screaming-case tokens and unresolved date fields, so ordinary lowercase identifiers do not register. Regex verified on 2026-08-15 against both specimens and three negatives."}],"evidence_grade":"corroborated","false_positive_notes":"Mail merge and content management templates produce the identical artifact with no model involved: a broken merge field, a snippet pasted out of a boilerplate library, a citation manager that could not resolve a date. Writers working from a house template leave the brackets in when a deadline moves. Every one of those cases deserves the same flag for the same reason, so treat a hit as a publishing-process failure rather than a claim about who drafted the sentence. Writing about templates quotes these strings on purpose, and quoted examples have to be excluded before a hit counts.","model_attribution":"Wikipedia records the citation-field forms in model output, naming the access-date parameter as the most common slot and the source and publisher fields as rarer ones. No vendor documents the behaviour, and no measurement by model family exists in anything this project read.","platform_notes":[{"platform":"wikipedia","note":"Criterion G15 covers pages published with no human review, and an unresolved citation field is the cheapest evidence that no review happened."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"summary-box-reflex","name":"TL;DR box reflex","aka":["TL;DR block","summary box on a short piece","what you'll learn preamble"],"category":"formatting","subcategory":"summary-block","description":"A summary block sitting on top of a piece too short to need one. The block restates the body in nearly the same words, so a reader who stops there has read the whole thing twice.","why_it_reads_ai":"Chat interfaces reward a visible answer at the top, and the habit follows the text out of the window. vale-ai-tells carries heading rules for the announcement family, including headings that promise what a reader will learn before the piece has said anything. The judgement is about proportion. On a 2,000-word piece a summary is a service. On a 200-word post it is the post.","examples":[{"before":"TL;DR: We moved our newsletter to a new provider and it went fine.\n\nWe moved our newsletter to a new provider last month. It went fine. Deliverability held and the API took an afternoon to wire up.","after":"We moved our newsletter to a new provider last month. The migration took an afternoon, and the only thing that broke was the unsubscribe link in our oldest template, which had been hard-coded in 2023. Open rates did not move either way across the first three sends.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Count the words in the body, excluding the summary block. If the body runs to 400 words or more, return no hit. Below that, list every fact stated in the summary block, then check each one against the body. Escape hatches that mean no hit: the block carries a figure, a date, a name or a caveat the body never repeats; the surface is documentation whose summary is a navigation aid; the publishing template requires a summary field. If every fact in the block already appears in the body and the block adds nothing, return a hit and quote the block. Otherwise return no hit."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active with a low grade and low severity, which is the honest pairing. One machine-checkable rule set carries heading rules for the announcement family, and the rest of the support is our own review queue, named as a source rather than dressed up as measurement. The rubric is gated on body length so that long-form summaries never register."}],"evidence_grade":"community-observed","false_positive_notes":"Long-form journalism and technical documentation put a summary at the top for readers who will not reach the end, and that is a service to them rather than a habit. Academic abstracts, executive summaries and legal head-notes are required by the form and cannot be dropped. Many content management templates render a summary automatically out of the field the writer filled in for search results, so it never passed through a writing decision. Judge proportion and novelty together: a block carrying a figure or a caveat the body never repeats is doing work, and one paraphrasing a 200-word post is filling a slot.","model_attribution":"Undocumented. No vendor names a summary-block instruction in anything this project read, and no study measures the habit. The evidence here is one machine-checkable rule set plus our own review queue, which is why the grade sits where it does.","platform_notes":[{"platform":"linkedin","note":"The first two lines are the only part shown before the more link, so a summary block spends the whole preview restating what follows it."}],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"feedsquad-observed","title":"FeedSquad review queue, summary blocks on short agent drafts","observed":"2026-08-15","corpus":"Agent-drafted LinkedIn and blog posts held for human approval in the FeedSquad review queue, weeks 30 to 33 of 2026. No sample count is claimed; the observation is qualitative and is the reason this entry carries the lowest evidence grade with an external source attached."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"decorative-symbol-string","name":"Decorative symbol strings","aka":["arrow bullets","star string dividers","repeated exclamation emphasis"],"category":"formatting","subcategory":"decoration","description":"Arrows, star strings and runs of punctuation carrying structure or emphasis that the sentences never earned.","why_it_reads_ai":"Microsoft's advertising blog advises writers to keep punctuation plain and names arrows, star strings and long runs of punctuation as things that break machine parsing. That is advice with no data behind it and no ranking claim attached, so read it as one company's guidance rather than as a measurement. The reason the ornament reads as low effort is older than any of it. A symbol is faster to insert than a sentence that would justify the emphasis.","examples":[{"before":"Three changes this week ★★★\n→ faster exports\n→ fewer timeouts\n→ a new billing page\n\nGo and look!!!","after":"Three changes this week. Exports on the largest accounts finished in 50 seconds against four minutes in June. Timeouts stopped once the job moved to a queue. The billing page now shows the next invoice date, which was the most common support question last quarter.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"(?:(?:^|[ \\t])(?:[\\u2192\\u21D2\\u27A1\\u2794\\u25B6\\u25BA]|[\\u2605\\u2606]{2,}))|[!?]{3,}","flags":"gm","scope":"document"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active with low severity, because the human base rate on social surfaces is enormous. Microsoft names the same decorations in guidance written for human copywriters, which is a vendor recording the pattern without measuring it. The character class is narrow: arrow glyphs, runs of two or more stars, and three or more stacked exclamation or question marks. Regex verified on 2026-08-15 against its own specimen and three negatives."}],"evidence_grade":"primary-doc","false_positive_notes":"Social marketers invented most of this ornament and still reach for it because the composer offers no formatting at all, so on Instagram and X an arrow is a workaround for a missing bullet list rather than a habit. Arrows are exact notation in writing about state machines, chemistry and sound change, where the glyph is the term. Star strings are how review sites render a rating. Judge by whether removing the symbol would cost the reader information: a rating loses its meaning without the stars, and a run of exclamation marks loses nothing at all.","model_attribution":"Unassigned. Microsoft's guidance is written for human copywriters and names no model, and no source this project read attributes arrows or star strings to a model family.","platform_notes":[{"platform":"instagram","note":"The caption field offers no formatting, so a symbol standing in for a bullet is a workaround for a product limitation before it is anything else."},{"platform":"bing","note":"Microsoft frames the advice around machine parsing for AI answers. It carries no ranking claim and no data, and this project cites it as advice."}],"languages":["en"],"sources":[{"kind":"external","title":"Microsoft Advertising, optimizing content for inclusion in AI search answers","url":"https://about.ads.microsoft.com/en/blog/post/october-2025/optimizing-your-content-for-inclusion-in-ai-search-answers","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"emphatic-italics","name":"Emphatic italics","aka":["italicised stress","italics-heavy emphasis"],"category":"formatting","subcategory":"emphasis","description":"Italics used for spoken stress across running prose. The mark does the job a sharper sentence would have done.","why_it_reads_ai":"The ICML 2025 idiosyncrasies work classifies five model families from text alone at high accuracy and reports that they differ in how they use bold, headers, enumerations and italics. That is an aggregate classifier result on a dated API snapshot, and it does not license calling a single document. Per-model italic rates could not be confirmed at page level by this project, so the ranking here stays qualitative. Two candidate entries were merged into this one during the build, because one measured finding had been minted twice.","examples":[{"before":"The migration *did* run clean. What surprised us was that it ran clean for the *right* reasons, which is the kind of thing you only see clearly afterwards. Most teams never stop to ask *why* something worked, and that is usually the difference between a good quarter and a good year.","after":"The migration ran clean on all 340 accounts, including the two that had needed a manual fix in every previous run. The difference was one ordering change in the batch job, which we found by replaying the June failure against a copy of production.","note":"Figures in this repair are invented for the specimen. The tell is the italic stress landing on words the sentence never earned, not the emptiness. The earlier specimen carried a well-observed incident, which made the italics look like ordinary voice."}],"detection":{"type":"statistical","metric":"italic-spans-per-1000-words","threshold":8,"direction":"above","threshold_basis":"No human baseline for italic density in prose has been published in anything this project read. The ICML 2025 work reports formatting differences between chat and instruct variants and notes that italics vary less than bold and headers, without per-model rates this project could confirm at page level. Eight italic spans per 1,000 words is a FeedSquad review trigger and not a published cutoff. A reading above it means read the paragraph, nothing more."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active and deliberately weak. The peer-reviewed support is a classifier study showing that model families differ in formatting habits, which is aggregate evidence from a dated snapshot rather than a per-document test. A second candidate entry naming one family by its italics was merged in here, because both rested on that one finding. The threshold is ours and says so."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Conversational essayists stress with italics as a house habit, and the practice has a magazine tradition much older than any of this. Fiction writers italicise interior thought by convention. Academic and technical writing italicises terms on first use, foreign words and titles of works, none of which is emphasis at all. A reviewer separates the two by checking what the italics mark: a convention marks the same category of word every time it appears, while the habit marks whichever word the sentence failed to stress on its own.","model_attribution":"Reported at family level and not usable per document. The ICML 2025 study classifies five families from text alone and describes distinct formatting habits including italics, from a dated API snapshot that no longer corresponds to shipping models. An association between italics and one family circulates from the same paper; this project could not confirm the per-model figures at page level and does not print them.","platform_notes":[{"platform":"linkedin","note":"The composer has no italics, so asterisks publish as asterisks and this signal is unobservable there without a rich-text source."},{"platform":"x","note":"Plain-text bodies again. Italic stress can only reach the timeline as Unicode substitution characters, which is a different habit with a different base rate."}],"languages":["en"],"sources":[{"kind":"external","title":"Sun et al., Idiosyncrasies in Large Language Models, ICML 2025 (arXiv:2502.12150)","url":"https://arxiv.org/abs/2502.12150","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"specificity-vacuum","name":"Specificity vacuum","aka":["generic content","superficial analysis","detail-free advice","content with no nouns in it","empty profundity","spatial detail vacuum","one-size-fits-all claim","averaged claim","content any model could produce"],"category":"semantic","subcategory":"absence-of-fact","description":"The text asserts things about the world and names nothing in it. No identified person, product or place. No quantity. Swap the subject for a different one and the paragraphs still read as written. That substitution is the practical test.","why_it_reads_ai":"A model writing without retrieval has no specifics to spend. It has the shape of the genre and the average of the corpus, so it produces the shape and leaves the slots empty. Detectors pick this up sideways. Low perplexity means every next word was the expected one, and proper nouns and figures are exactly the words a model would not have predicted.","examples":[{"before":"Consistency is the foundation of any successful content strategy. Brands that post regularly build trust with their audience over time, and trust is what turns followers into customers. The key is finding a cadence you can sustain and then committing to it. Quality matters as much as quantity, so focus on delivering real value in every post.","after":"We posted twice a week for eleven weeks, then dropped to once. Replies per post roughly doubled while total replies stayed flat, so the extra post was buying nothing. The piece that did best that quarter was a screenshot of an invoice our billing code got wrong, with the fix under it. It took twenty minutes. The polished essays took a day each and went nowhere.","note":"The repair is not better wording. It is a fact the writer had and the first draft did not use."},{"before":"Effective onboarding is critical for reducing churn. Companies that invest in the early user experience see significantly better retention outcomes, and a thoughtful approach to activation pays dividends across the entire customer lifecycle.","after":"Churn in our first month was 31 percent until we cut the signup form from nine fields to three. It dropped to 19 percent over the next two months. Nothing else changed in that window, which is the strongest claim we can honestly make about it.","note":"The after version is falsifiable. Someone can ask what the other two months looked like. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Task: decide whether the text makes claims about the world while naming nothing in it.\n\nStep 1. Count identifying proper nouns: a specific person, product, company, place, work or institution. Do NOT count the author's own name, the platform's name, generic category nouns (marketers, founders, teams), or brand names used only as examples of a category.\nStep 2. Count world-quantities: prices, counts, durations, percentages, sizes, dates. Do NOT count list numbering, round rhetorical figures ('a million reasons'), or numerals inside quoted material.\nStep 3. Count time anchors: any clause that fixes a claim to an occasion (a month, a quarter, a named year, 'the week we launched').\nStep 4. Count falsifiable claims: sentences a reader could check and prove wrong.\n\nDecision. Flag only if ALL of these hold: (a) the text is 150 words or longer; (b) it gives advice or asserts what works; (c) Step 1 total is 0; (d) Step 2 total is 0; (e) Step 4 total is 0 or 1.\nWeak flag if Steps 1 to 3 sum to 1 or fewer per 200 words while the text claims outcomes.\nDo not flag: text under 150 words; fiction; poetry; devotional or aphoristic writing; a personal reflection that makes no claim about the world; text that states its specifics were removed for confidentiality.\n\nOutput: the four counts, the single strongest specific found (quoted) or 'none', and one of FLAG / WEAK / PASS."},"severity":"high","status":"active","status_history":[{"date":"2026-08-14","status":"active","rationale":"Named in some form by every platform policy reviewed and by the Wikipedia editor guide. It survives the decay argument that retires lexical tells, because it describes missing substance rather than surface style."}],"evidence_grade":"corroborated","false_positive_notes":"Confidentiality produces identical prose. A consultant under NDA and an investor-relations writer at a listed company both strip names and figures on legal advice, and a clinician or social worker strips them to protect a person. Beginners write this way too, because they have not done the work yet, and calling a beginner a machine is the most common way this entry gets misused. Devotional and aphoristic writing has no specifics by design and never claimed to. The measured harm case sits next door: Liang et al. found seven detectors misclassified 61 percent of human TOEFL essays, and the mechanism was constrained word choice, which correlates with the same emptiness this entry describes. Require length and an advice-giving posture before flagging.","model_attribution":"Not attributable to a model family. The excess-vocabulary and idiosyncrasy literature measures word choice at population scale, and no published work assigns detail-poverty to one vendor.","platform_notes":[{"platform":"linkedin","note":"LinkedIn's May 2026 policy defines its target as content that sounds polished but lacks unique perspective or substance, and lists generic AI-generated content without a clear perspective as one of three named targets. The penalty is out-of-network distribution, not removal."},{"platform":"youtube","note":"YouTube's inauthentic-content policy names AI-generated content built from generic templates only when it lacks the creator's original insight or perspective. Every qualifier in that sentence does the work."},{"platform":"google","note":"Google's helpful-content self-assessment asks whether the content provides original information, reporting, research or analysis, and its spam policy states the test applies no matter how the content was created."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia:Signs of AI writing (sections: Superficial analyses; Vague attributions and overgeneralization of opinions)","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Keeping conversations real on LinkedIn (Laura Lorenzetti, VP and Executive Editor, LinkedIn Global Editorial, 2026-05-20)","url":"https://www.linkedin.com/pulse/keeping-conversations-real-linkedin-laura-lorenzetti-9821e","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Google Search Central: Creating helpful, reliable, people-first content","url":"https://developers.google.com/search/docs/fundamentals/creating-helpful-content","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Liang, Yuksekgonul, Mao, Wu, Zou. GPT detectors are biased against non-native English writers. Patterns 4(7), 2023","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC10382961/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Ott, Choi, Cardie, Hancock. Finding Deceptive Opinion Spam by Any Stretch of the Imagination. ACL-HLT 2011, pp. 309-319","url":"https://aclanthology.org/P11-1032/","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Kommers et al., Why Slop Matters (arXiv:2601.06060), accepted to ACM AI Letters 23 Dec 2025","url":"https://arxiv.org/abs/2601.06060","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Google Search Quality Rater Guidelines (dated 11 September 2025)","url":"https://static.googleusercontent.com/media/guidelines.raterhub.com/en//searchqualityevaluatorguidelines.pdf","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-14","updated":"2026-08-15"},{"id":"no-first-person-incident","name":"No first-person incident","aka":["authorless advice","perspective-free post","experience-shaped prose with no experience in it","no first-hand specifics"],"category":"semantic","subcategory":"absence-of-incident","description":"The piece gives advice and reports no occasion on which the author did anything. First person can still be present as opinion or as address to the reader. What is missing is an event with a time and an outcome. Nothing in the text could only have been written by the person whose name is on it.","why_it_reads_ai":"A model has no history to draw on, so it cannot supply an occasion. It supplies the posture instead. This is the pattern platforms describe when they say content lacks perspective. It also resists editing, because adding an incident requires having had one.","examples":[{"before":"Hiring your first employee is one of the most important decisions a founder will make. Take your time, define the role clearly, and trust your instincts. Culture fit matters more than credentials, and the right person will grow with the company as it scales.","after":"We hired our first employee in February and wrote the job description in an afternoon. That was the mistake. She spent her first six weeks doing support, because support was what we actually needed, and nothing in the ad had said so. She left in August. The second time, we logged what the two of us did for a week and posted the log as the description.","note":"Same advice underneath. The difference is that the second version can be wrong about something."},{"before":"Many teams underestimate how much technical debt slows them down. It is worth auditing your codebase regularly and prioritising the areas that cause the most friction for developers day to day.","after":"We spent a Friday counting where our build time went and found 40 percent of it in a test suite for a feature we had removed in the spring. Deleting it took an hour. I had been complaining about build times for four months without ever measuring them, which is the part worth copying.","note":"The incident carries a time, an action and an outcome. Opinion alone would not. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Task: decide whether the text contains a first-person incident.\n\nAn incident needs three parts, all present in the same passage: (1) an actor the author was, marked by I or we; (2) an event located in time, either by date, season, sequence marker or a named occasion; (3) an outcome or consequence that followed.\n\nCount as an incident: 'We moved the pricing page behind a login in March and demo requests halved.'\nDo NOT count as an incident: opinion ('I believe pricing pages should be simple'); a habit with no occasion ('I always start with the headline'); a hypothetical ('imagine you are launching'); an intention ('we plan to test this'); address to the reader ('you have probably noticed').\nAlso count, separately: attributed second-hand incidents, meaning a named person, a cited study or a quoted source that supplies the event instead of the author.\n\nDecision. Flag only if ALL hold: (a) the text is 150 words or longer; (b) it gives advice or asserts what works; (c) first-person incident count is 0; (d) attributed second-hand incident count is 0.\nDo not flag: encyclopedic or reference writing, which forbids first person by convention; third-person news reporting; product or API documentation; recipe, spec or legal text; anything whose byline is not the writer, including ghostwritten and translated work.\n\nOutput: incident count, second-hand incident count, the strongest candidate sentence quoted, and FLAG or PASS."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-14","status":"active","rationale":"Three platform policies name the absence of the creator's own insight as the enforcement trigger. Scope is deliberately narrow: it applies to first-person advice under a named byline, and the false-positive note lists the genres where absence of first person is the house rule."}],"evidence_grade":"corroborated","false_positive_notes":"Whole genres forbid first person, and their writers are the people this entry can hurt. Encyclopedia style bans it outright. Most newspaper reporting outside the column pages does too. Technical documentation is written from nowhere by convention, and so are specs, recipes and legal drafting. Ghostwriters produce advice out of someone else's experience for a living, which means the byline is not the writer and the absence is contractual. Translators carry incidents that were never theirs. Junior writers with a real job and no war stories yet also fail this test honestly. Apply it to personal-brand posts under a named byline. Do not apply it anywhere else.","model_attribution":"Not attributable to a family. The absence follows from the model having no autobiography, which is true of every assistant surveyed. No published work measures first-person incident density per vendor.","platform_notes":[{"platform":"youtube","note":"The channel monetization policy names AI-generated content from generic templates that gives the impression of mass production without adding the creator's original, authentic insights or perspective. The last clause is the operative one."},{"platform":"linkedin","note":"Generic AI-generated content lacking a clear perspective is one of the three targets in LinkedIn's May 2026 policy. LinkedIn also states plainly that AI and slop are not the same thing, so the trigger is the missing perspective rather than the tool."},{"platform":"instagram","note":"Meta's 2026 original-creator policy counts a remix or overlay as original only when the creator adds fresh information, analysis or substantial improvements. Narrating someone else's material without adding value is treated as unoriginal. Meta's policy text does not mention AI at all."}],"languages":["en"],"sources":[{"kind":"external","title":"YouTube channel monetization policies (inauthentic content; page states last updated 2025-07-15)","url":"https://support.google.com/youtube/answer/1311392","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Keeping conversations real on LinkedIn (Laura Lorenzetti, VP and Executive Editor, LinkedIn Global Editorial, 2026-05-20)","url":"https://www.linkedin.com/pulse/keeping-conversations-real-linkedin-laura-lorenzetti-9821e","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Rewarding Original Creators on Facebook (Meta Newsroom, 2026-03-13)","url":"https://about.fb.com/news/2026/03/rewarding-original-creators-on-facebook/","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Wikipedia:Signs of AI writing (section: Signs of human writing)","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Google Search Quality Rater Guidelines (dated 11 September 2025)","url":"https://static.googleusercontent.com/media/guidelines.raterhub.com/en//searchqualityevaluatorguidelines.pdf","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Google Search Central: Optimizing your website for generative AI features on Google Search","url":"https://developers.google.com/search/docs/fundamentals/ai-optimization-guide","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-14","updated":"2026-08-15"},{"id":"real-entity-fabricated-work","name":"Real entity, fabricated work product","aka":["fake book by a real author","invented credential at a real institution","fabricated quote from a real person","I never said that"],"category":"semantic","subcategory":"fabrication","description":"A person, institution or company that a reader can look up, attached to a book, quotation, study or credential that does not exist. The entity checks out. The work product hung on it does not. Half the sentence survives verification, which is what makes the other half so hard to see.","why_it_reads_ai":"A name and a plausible title for that name come out of the same distribution, so they arrive together with nothing binding them. Nothing in the text records whether the book was ever opened. The best documented case is a newspaper auditing itself: the Chicago Sun-Times review of the syndicated section it printed in May 2025 found recommended titles credited to living authors who had never written them, an expert placed at a university that had not employed her, and quoted remarks from a named blogger who told reporters he had said nothing of the kind. The review noted that every story carrying named sources carried errors, which inverts the normal editorial heuristic.","examples":[{"before":"Dr. Helena Marsh of the Northfield Institute for Consumer Trust found in her 2023 study Signals That Sell that shoppers abandon a checkout within four seconds of seeing an unexpected fee.","after":"We logged 1,412 checkouts in June. Thirty-eight percent of the abandonments happened on the screen where shipping was added. That number is ours, from our own funnel, and anyone who wants to argue with it can ask how we counted.","note":"The researcher, the institute and the study in the first version are invented for this specimen. The repair drops the borrowed standing and uses a figure the writer can defend. Figures in this repair are invented for the specimen."},{"before":"Our method was reviewed by a former lead assessor for the National Standards Board of Applied Analytics, who confirmed it meets the current bar.","after":"Our advisor reviewed the method on 4 July and has agreed to be named in the appendix, with her review notes published beside it. There is no board and no bar. There is one person, named, whose objections you can read.","note":"The board in the first version does not exist. That is the whole of the problem, and no amount of rewording fixes it."}],"detection":{"type":"judge","rubric":"Task: find places where a checkable entity carries an uncheckable work product.\n\nStep 1. List every named person, institution, company, journal, award or court in the text.\nStep 2. For each, list what the text attributes to it: a title, a quotation, a credential, an affiliation, a study, a finding, a ruling.\nStep 3. Mark the entity VERIFIED or UNVERIFIED by looking it up.\nStep 4. Mark each attribution VERIFIED, UNRESOLVED or ABSENT. UNRESOLVED means the entity's own site, catalogue or register returns nothing under that title, that person or that year.\n\nEscape hatches, which return PASS for that item: the text marks the attribution as paraphrase from memory; the work is stated as forthcoming, unpublished, internal or under embargo; the source is a private communication the text identifies as one; the entity is presented as fictional; the reviewer cannot reach the catalogue at all, which is unknown rather than absent.\n\nDecision. FLAG when at least one attribution is UNRESOLVED and the entity carrying it is VERIFIED. That exact pairing is the pattern. Return WEAK when the entity is also UNVERIFIED, because that is a different failure and a different entry. Return PASS otherwise.\n\nOutput: each entity, its attribution, both marks, and one of FLAG / WEAK / PASS. State nothing about who wrote the text."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"A newspaper published its own audit of the failure in May 2025, naming fake titles credited to real authors, a fabricated expert at a real university and invented quotes from a real person. Snopes and NPR corroborate the same section. Self-incriminating evidence, which is the strongest kind available here."}],"evidence_grade":"corroborated","false_positive_notes":"Ordinary citation error produces the same surface. A transposed year, a title remembered a word wrong, an author confused with a co-author, a book reissued under a different title in another market: all of these leave a real name attached to a reference that will not resolve. Catalogues are also incomplete, and a great deal of pre-1990 trade publishing was never indexed at all. Verification has to come before any accusation, and the correct first move is to ask the writer for the source rather than to reach a conclusion about how the text was produced. A writer who supplies a scan or a shelf mark has answered the question. A writer who supplies another unresolvable reference has not.","model_attribution":"No published work assigns this to a vendor. The documented cases run across several years and several assistants, and every one of them reached print through a pipeline where nobody checked a citation.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Chicago Sun-Times, special section with fake book list plagued with additional errors","url":"https://chicago.suntimes.com/news/2025/05/29/special-section-king-fake-book-list-errors-sun-times-review","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Snopes, Chicago Sun-Times AI reading list fact check","url":"https://www.snopes.com/fact-check/chicago-sun-times-ai-reading-list/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"NPR, fake summer reading list","url":"https://npr.org/2025/05/20/nx-s1-5405022/fake-summer-reading-list-ai","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"well-formed-citation-no-referent","name":"Well-formed citation, no referent","aka":["fake case cite","invalid ISBN checksum","unresolvable DOI"],"category":"semantic","subcategory":"citation-artifact","description":"The citation has correct form and no referent. Case name, reporter volume and pin cite in the right order, resolving to no case. A DOI with a valid prefix that returns nothing. An ISBN whose check digit fails arithmetic.","why_it_reads_ai":"Citation form is one of the most patterned strings in written English, so it is cheap to produce. The referent is a fact about the world that the form does not carry, so it is not produced with it. Wikipedia built the failure into a speedy-deletion criterion, G15, which names unresolvable DOIs and invalid ISBN checksums among the signs a page could only have come from a language model. Courts have been dealing with the same thing in filings since 2023; a public database tracks the cases, though its counts disagree between summaries and the site refused our fetchers, so this entry cites its existence and no number.","examples":[{"before":"A 2019 review in the Journal of Applied Retail Analytics (doi:10.9999/jara.2019.04.007) found that shelf placement outperformed price promotion in 61 percent of trials.","after":"We could not find a study comparing shelf placement with price promotion, so here is ours instead. Two stores, six weeks, the same product at eye level in one and at knee level in the other. Eye level sold 22 percent more units. Two stores is not a trial and we are not calling it a finding.","note":"The journal, the identifier and the percentage in the first version are invented for this specimen. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Task: check whether each citation resolves to something that exists.\n\nStep 1. Extract every citation: author, title, venue, year, identifier, page or pin cite.\nStep 2. Run the cheap arithmetic checks first. ISBN-10 and ISBN-13 check digits are computable without a network. So are volume and page ranges that fall outside a journal's published run.\nStep 3. Resolve each identifier. A DOI that returns no record, a case citation absent from the reporter, a URL that never existed rather than one that has moved.\nStep 4. Search title plus author independently of the identifier. A real work with a mistyped identifier will surface here. A fabricated one will not.\n\nEscape hatches, which return UNKNOWN rather than FLAG: paywalls and login walls; databases the reviewer cannot reach; works in languages or scripts the reviewer does not index; pre-digital material; preprints later withdrawn; personal communications and interviews, which have no public referent by nature and should be marked as such in the text.\n\nDecision. FLAG when a citation fails Step 2 arithmetic, or fails both Step 3 and Step 4. Mark UNKNOWN when only Step 3 fails and Step 4 could not run. Otherwise PASS.\n\nOutput: one line per citation with its verdict, plus the count of FLAG results. A dead link and a fabricated citation are different findings and must be reported differently."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Wikipedia's speedy-deletion criterion G15 names invalid identifiers as a sign, and court coverage records filings with citations that resolve to nothing. The pattern has an arithmetic component, which is rare in this dataset and makes it unusually cheap to check."}],"evidence_grade":"corroborated","false_positive_notes":"Link rot kills real citations every day. Journals move, publishers merge, DOIs are registered late or never, and a working paper cited in 2021 can be unreachable in 2026 with nothing wrong about it. Legal and academic citation styles also vary enough that a correct reference can look malformed to a reviewer trained on a different one, and transliterated author names defeat naive search. The distinction that matters is between a referent that once existed and one that never did: an archive copy, a library catalogue record or a citing paper settles it. Reviewers working outside their own field should record UNKNOWN and ask, because absence of a search result is not absence of a work.","model_attribution":"Documented across assistants rather than in one family. The behaviour follows from generating a reference rather than retrieving one, which is a property of the setup rather than of a vendor.","platform_notes":[{"platform":"wikipedia","note":"Criterion G15 for speedy deletion lists citations with invalid identifiers among the signs a page could only plausibly have been generated by a language model. It is a deletion criterion for pages, not a verdict about an editor."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Charlotin, AI hallucination cases database","url":"https://www.damiencharlotin.com/hallucinations/","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Bloomberg Law, AI faked cases become core issue for judges","url":"https://news.bloomberglaw.com/legal-ops-and-tech/ai-faked-cases-become-core-issue-irritating-overworked-judges","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"invalid-citation-temporality","name":"Citation with invalid temporality","aka":["source dated before the event it reports"],"category":"semantic","subcategory":"citation-artifact","description":"A cited source is dated before the event it is cited for. The 2019 article does not report the 2024 recall. Nobody wrote it about that.","why_it_reads_ai":"This is the sharpest mechanical check in the semantic set, because it needs no judgment about quality and no domain knowledge. Two dates and a subtraction settle it. Wikipedia lists the pattern under criterion G15, alongside the other citation-form failures, where it sits as the variant that a reviewer can run without leaving the page.","examples":[{"before":"The September 2025 recall was widely covered at the time, including in the industry survey cited below, which sets out the same failure mode in detail. [Nordic Component Review, March 2021]","after":"The recall notice is dated 12 September 2025 and we are linking the notice itself. A 2021 industry survey described a similar failure mode four years earlier, which is worth reading and is not coverage of this recall.","note":"The publication and dates in the first version are invented for this specimen. The repair keeps the older source and stops it standing in for reporting it could not contain."}],"detection":{"type":"judge","rubric":"Task: compare each citation's date with the date of what it is cited for.\n\nStep 1. For every citation, record the source date. Use the publication date on the source itself, not the date given in the citing text.\nStep 2. Record the date of the event, finding or state of affairs the citation supports. If the citing text gives no date, look for one in the surrounding paragraph.\nStep 3. Subtract. Mark IMPOSSIBLE when the source predates the event.\n\nEscape hatches, which return PASS: articles updated after publication that keep the original date, so check for a revision history before concluding; wire copy republished under a new timestamp; scheduled or pre-announced events, where reporting legitimately precedes them; forecasts, projections and previews; recurring events where the citation covers an earlier occurrence and the text says so; undated pages, which return UNKNOWN.\n\nDecision. FLAG on one IMPOSSIBLE pairing where no escape applies. One is enough, because the citation cannot be doing the work claimed for it.\n\nOutput: the citation, both dates, the gap in days, and FLAG / PASS / UNKNOWN."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"One community source names it, and the check itself is arithmetic rather than interpretive, so the entry ships on a single citation with that stated. Nothing else in this dataset can be verified as cheaply."}],"evidence_grade":"community-observed","false_positive_notes":"Dates on the web are unreliable in both directions. Publishers backdate updated articles to the original publication, wire services restamp syndicated copy, and content management systems display the date of the last template change rather than the last edit. A source can also legitimately predate an event when it is reporting a plan, a filing, a scheduled launch or a court date. Historians and archivists work with undated material constantly and would fail a naive version of this check every day. The reviewer needs the archive record or the revision history before calling anything impossible, and where neither is reachable the honest verdict is unknown.","model_attribution":"Undocumented per vendor. It follows from a citation being assembled rather than looked up, and the same failure appears in human work produced from a reference list nobody opened.","platform_notes":[{"platform":"wikipedia","note":"G15 groups this with the other citation-form failures. Editors treat it as one of the fastest checks available during cleanup triage."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"confident-arithmetic-error","name":"Confident arithmetic error","aka":["fluent but wrong numbers","APR and APY confusion"],"category":"semantic","subcategory":"factual-failure","description":"Calm, grammatical, authoritative prose whose numbers do not survive a calculator. The register never wavers. Nothing in the sentence signals that the arithmetic went wrong, because nothing in the sentence was ever checked.","why_it_reads_ai":"Fluency and arithmetic are produced by different faculties in a language model, and only one of them is being optimised in the sentence. The worked example that made this famous is a set of personal-finance explainers published in late 2022 and early 2023, in which a total was reported as a gain: ten thousand at three percent yields three hundred, and the text gave the balance instead. CNN reported 77 published stories written with an internally designed AI engine, an audit that followed the first factual error, and corrections on a number of them with a small number described as substantial. The outlet did not say how many of the 77 were corrected, so no count appears here. The confidence is what does the damage, because a hedged wrong number invites checking and an assured one does not.","examples":[{"before":"Put 10,000 euros into an account paying 3 percent a year and after twelve months you have earned 10,300 euros in interest.","after":"Put 10,000 euros into an account paying 3 percent a year and after twelve months you have earned 300 euros. The balance is 10,300. The gain is 300. Confusing those two is the most common error in this kind of explainer and it is worth naming.","note":"The arithmetic in the repair is real and checkable, which is the point of the repair."},{"before":"A 25,000 euro loan at 4 percent costs a flat 1,000 euros in interest per year for the life of the loan, so five years of borrowing costs 5,000 euros.","after":"A 25,000 euro loan at 4 percent costs about 1,000 euros in the first year and less every year after that, because interest is charged on what is left. Repaid in five equal annual instalments of roughly 5,620 euros, the total interest lands near 3,100 euros rather than 5,000.","note":"The repair was computed rather than estimated. Anyone can rerun it with an amortisation formula and should."}],"detection":{"type":"judge","rubric":"Task: recompute every number in the text.\n\nStep 1. Extract each numeric claim with its unit, its rate type and its period.\nStep 2. Recompute. Check these families in order, since they cover most observed failures:\n  (a) total versus gain, where a balance is reported as interest earned or a revenue figure as profit;\n  (b) rate types, where an annual percentage rate is treated as a yield, or a simple rate as a compounding one;\n  (c) percentages versus percentage points;\n  (d) parts against their sum, where listed components do not add to the stated total;\n  (e) date arithmetic, where a duration disagrees with its endpoints;\n  (f) unit and scale, where thousands and millions or metric and imperial are mixed.\nStep 3. Record the stated value, the computed value and the difference.\n\nEscape hatches: figures inside a quotation, which belong to the speaker; numbers the text marks as rounded, illustrative or hypothetical, where the rounding accounts for the gap; currency conversions with no stated date; deliberately simplified worked examples that say they are simplified.\n\nDecision. FLAG on any error large enough to change the reader's decision, and note the rest as minor. Report the recomputation, not an opinion about the writer.\n\nOutput: a table of stated value, computed value, and verdict per claim."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Documented in a news outlet's own correction record and covered by two independent outlets, one of which reported the correction count. The failure recurs wherever numeric copy is published without a numerate reader in the loop."}],"evidence_grade":"corroborated","false_positive_notes":"People are bad at arithmetic and always have been. Financial journalists confuse a rate with a yield, subeditors drop a zero, and spreadsheets export percentages as decimals into copy that nobody recomputes. Fatigue produces the same sentence as anything else. What separates a slip from the pattern is distribution: one wrong figure in a piece is a correction, while a run of pieces in which the numbers are consistently wrong in the same direction and the prose is consistently assured points at a process with no numerate reader in it. That is a finding about a pipeline, not about a person, and it should be reported that way.","model_attribution":"The documented case involved a proprietary in-house system rather than a named consumer assistant, which is worth stating because the tell is often attributed to a vendor it cannot be traced to.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Futurism, CNET's article-writing AI is already publishing very dumb errors","url":"https://futurism.com/cnet-ai-errors","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"CNN, plagued with errors, a news outlet's decision to write stories with AI backfires","url":"https://www.cnn.com/2023/01/25/tech/cnet-ai-tool-news-stories","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"domain-blind-sentence","name":"Domain-blind sentence","aka":["advice no practitioner would write","parseable but absurd instruction"],"category":"semantic","subcategory":"factual-failure","description":"A sentence that parses, scans and means nothing workable. Anyone who has done the thing once reads it and stops. The failure is not in the grammar and not in the facts as stated. It is that the steps cannot be carried out in the order given.","why_it_reads_ai":"Generated text is fluent about a subject without having done it, and practice is where the ordering constraints live. The tell surfaced publicly through a set of sports and finance explainers published under fabricated bylines in 2023, where the giveaway was a piece of advice about a sport that assumed equipment the sport requires. A single such sentence tells you nobody with the relevant experience read the draft, which is a fact about the publishing process and a useful one.","examples":[{"before":"Sourdough is a good first bread for beginners, because the starter can be mixed the same morning you plan to bake and needs no special handling.","after":"Sourdough is a hard first bread, because the starter takes about a week to become reliable. Start it on a Monday, feed it every day, and expect the first loaf that rises properly around the second weekend. Ours took eleven days and two failures before it worked.","note":"Nothing in the first version is ungrammatical or self-contradictory. It is wrong about an ordering constraint that only shows up in practice."},{"before":"Hold the sprint retrospective before the sprint begins, so the team knows in advance what went wrong and can plan around it.","after":"Hold the retrospective in the last hour of the sprint, while people still remember Tuesday. We tried moving it to the following Monday and the notes got vaguer every time, so we moved it back.","note":"The first version inverts an ordering the practice depends on. It reads fine and cannot be done."}],"detection":{"type":"judge","rubric":"Task: decide whether the instructions could be carried out by someone doing this work.\n\nStep 1. Extract every imperative, recipe step, causal claim and piece of advice.\nStep 2. Test each against these failure classes:\n  (a) ordering, where a step requires the output of a step that comes later;\n  (b) prerequisite, where a required material, tool, permission or condition is never obtained;\n  (c) magnitude, where a stated time, quantity or cost is off by an order of magnitude against ordinary practice;\n  (d) category, where a procedure is applied to a material or system it does not apply to;\n  (e) safety, where following the instruction would injure someone or destroy the thing being worked on.\nStep 3. For each hit, write the one sentence a practitioner would say in reply.\n\nEscape hatches: deliberate novelty, where the text argues for the unusual method and gives a reason; regional or historical practice that differs from the reviewer's; specialist equipment that changes the constraint; text marked as untested, speculative or fictional; a reviewer with no knowledge of the domain, who should return UNKNOWN rather than guess.\n\nDecision. FLAG on one hit in classes (a), (b) or (e), which are hard failures. Two or more hits in (c) or (d) also FLAG. Otherwise PASS.\n\nOutput: the sentence, the failure class, the practitioner's reply, and the verdict."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"The 2023 fake-byline investigation and its trade-press follow-up both identify the impossible instruction as the sentence that broke the story. The pattern survives model improvement because it comes from having no practice rather than from weak language."}],"evidence_grade":"corroborated","false_positive_notes":"Freelancers write outside their expertise for a living, and a competent generalist on an unfamiliar beat produces exactly this sentence without any tool involved. Domains also disagree with themselves: bakers, mechanics and clinicians all hold live arguments about ordering that a reviewer from one school will read as an error from another, and regional practice differs enough that an instruction absurd in one country is standard in the next. Translation strips the qualifiers that made a step conditional. The fix in every case is a subject-matter read, which is cheaper than an argument about authorship and produces a better piece either way. A reviewer without the domain should say so rather than score it.","model_attribution":"No vendor can be named. The published cases used an unnamed in-house generator. No study measures practical absurdity per vendor, and the failure is about missing practice rather than about a model family.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Futurism, Sports Illustrated published articles by fake AI-generated writers","url":"https://futurism.com/sports-illustrated-ai-generated-writers","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Poynter, Sports Illustrated and the AI byline problem","url":"https://www.poynter.org/commentary/2023/sports-illustrated-artificial-intelligence-writers-futurism/","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"register-collapse-on-reuse","name":"Register collapse on borrowed phrasing","aka":["valence inversion","source reused with its tone flipped"],"category":"semantic","subcategory":"reuse-failure","description":"Source material is carried across with its emotional register inverted. A line written about hardship arrives as a recommendation. The words survive the move and their meaning does not.","why_it_reads_ai":"Reuse at scale copies strings and drops the context that told a reader how to take them. The canonical case is a travel guide from August 2023 that listed a food bank among the things to see in a city and repeated a line from the charity's own site about arriving hungry. The charity had written it about need. A human editor reading for sense catches this in one pass, which is why it is such a clean signal about the process: nobody read it. Microsoft disputed that a language model produced the guide and attributed it to human error, so this entry describes an unreviewed pipeline rather than a model, and that denial ships with it.","examples":[{"before":"Our third stop is the memorial garden, a peaceful green space the city opened after the 2016 fire. Bring a picnic and make an afternoon of it.","after":"The memorial garden was opened after the 2016 fire and the names are set into the north wall. It is open to the public every day. The city asks visitors to keep it quiet around the anniversary in June, when families come.","note":"The first version borrows a description written for one purpose and files it under another. The repair keeps the fact and restores what the place is for."}],"detection":{"type":"judge","rubric":"Task: decide whether any passage carries a subject associated with harm, loss or need in a register meant for leisure, opportunity or entertainment.\n\nStep 1. List the subjects the text describes: places, services, events, organisations.\nStep 2. For each, name the register the subject ordinarily belongs to. Grief, emergency, poverty, illness, criminal proceedings, and industrial accident all carry one. Restaurants, museums and shops carry another.\nStep 3. Name the register the sentence uses. Recommendation, invitation, ranking and enthusiasm mark the leisure register.\nStep 4. Mark every mismatch, and quote the specific words that carry it.\n\nEscape hatches: memorial and disaster tourism written with acknowledgement, which is an established genre; satire and black comedy where the tone is the point; sources whose own register is celebratory, such as a festival marking a historical event; charities and services describing their own work in the language they choose for it; reclaimed language used by the affected community.\n\nDecision. FLAG on one unacknowledged mismatch involving a subject in the harm register. Do not flag on discomfort alone; the test is whether the text has silently changed what the subject is for.\n\nOutput: the subject, the two registers, the quoted words, and FLAG or PASS."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"One incident, one press source, and a vendor denial that a language model was involved. That denial is why the grade is community-observed rather than corroborated, and the entry keeps the tell while declining to claim its cause."}],"evidence_grade":"community-observed","false_positive_notes":"Rushed human aggregation makes this error constantly, and always has. Travel desks recycle press material, listings are compiled from databases by people with no time to read them, and translation flattens the markers that told a reader how a sentence was meant. Memorial and disaster tourism is a real genre with its own conventions, and writing about a site of suffering as a place worth visiting is legitimate when the piece acknowledges what it is. The vendor in the documented case denied that any language model was involved, which is a useful reminder that this tell identifies an unreviewed process rather than a tool. The only reliable question is whether anyone read the sentence in its new home.","model_attribution":"Disputed at the source. The vendor said the guide came from algorithmic techniques with human review rather than from a language model. This entry treats the tell as a signal about review, not about the generator.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"CBC, Microsoft travel guide recommends Ottawa food bank","url":"https://www.cbc.ca/news/canada/ottawa/artificial-intelligence-microsoft-travel-ottawa-food-bank-1.6940356","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"satire-laundered-as-fact","name":"Satire laundered as fact","aka":["joke-to-advice pipeline","forum humour restated as instruction"],"category":"semantic","subcategory":"retrieval-failure","description":"A joke is retrieved, stripped of the cues that marked it as one, and restated as instruction. The failure happens at retrieval, one layer below the writing.","why_it_reads_ai":"Irony lives in context, in the thread, the venue and the reply that follows. A retrieval step that reads a sentence without them keeps the proposition and loses the frame. The best known case is a search feature that in May 2024 answered practical questions with advice traceable to joke posts, which the vendor then restricted after public reporting. The same failure appears wherever forum text is mined for answers, and it is the moment the word slop moved from niche use into general circulation.","examples":[{"before":"A common fix for a slow laptop is to freeze the drive overnight, which forum users report restores read speeds for good.","after":"A failing drive sometimes reads for a few more minutes when it is cold, which is where the freezer advice comes from. It is a last-ditch move for copying files off a dying disk, it often finishes the disk off, and it fixes nothing. Copy what you need and replace the drive.","note":"The repair keeps the grain of truth the joke grew on and states what it costs. Synonym swaps would have kept the bad instruction."}],"detection":{"type":"judge","rubric":"Task: decide whether any declarative instruction was likely written as a joke somewhere else.\n\nStep 1. Extract every instruction and factual claim that a reader might act on.\nStep 2. Score each against these joke markers: the action is physically absurd; it involves ingesting, burning or destroying something not meant for it; the stated effect contradicts the mechanism given; the phrasing has the rhythm of a punchline; the claim is attributed to unnamed forum or comment-section users; the advice would be funny if it failed.\nStep 3. Where a marker fires, look for the original. Joke sources are usually findable by searching the distinctive phrase.\n\nEscape hatches: the text names the source as humour or as folklore under test; the domain genuinely uses a counterintuitive method and the text explains the mechanism; satire, where the whole piece is the joke; historical practice that sounds absurd now and is documented.\n\nDecision. FLAG when an instruction carries two or more joke markers, or one marker plus an unnamed crowd attribution. Report the suspected source when found.\n\nOutput: the instruction, the markers that fired, the suspected origin, and FLAG or PASS."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"One press account of a search feature answering with material traceable to joke posts, followed by vendor restrictions. One source, so the grade stays community-observed, and the entry says which incident it rests on."}],"evidence_grade":"community-observed","false_positive_notes":"Human aggregators have laundered satire into news since the early 2000s, and every large newsroom has run a correction for repeating a joke site. The failure is register-blindness at retrieval, which says nothing about who assembled the text afterwards. Some counterintuitive advice is also correct: cold does briefly revive a failing drive, salt does go in some sweet recipes, and a practitioner who explains the mechanism is not repeating a joke. Domains with a strong deadpan culture, including much of open-source support, produce sincere advice that reads as parody to outsiders. The check is whether the original context can be found, and where it cannot the honest answer is that the claim is unsourced rather than that it is a joke.","model_attribution":"The documented case involved one vendor's search feature, and the failure has since been reported across retrieval products generally. It belongs to retrieval design rather than to a model family.","platform_notes":[{"platform":"google","note":"The May 2024 incident led to restrictions on which queries the feature would answer. The company characterised the failures as arising on very uncommon queries, which is a claim about frequency rather than about the mechanism."}],"languages":["en"],"sources":[{"kind":"external","title":"Forbes, Google restricts AI search tool after nonsensical answers","url":"https://www.forbes.com/sites/roberthart/2024/05/31/google-restricts-ai-search-tool-after-nonsensical-answers-told-people-to-eat-rocks-and-put-glue-on-pizza/","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"pseudo-lexical-label","name":"Pseudo-lexical labels","aka":["diagram non-words","text-shaped strings that are not words"],"category":"semantic","subcategory":"non-language","description":"Strings shaped like technical vocabulary that are not language. They have the right length, the right suffixes and the right position in a diagram, and they are not words. This index is otherwise text-only, and this entry is its single declared exception, because the strings are legible text carried inside an image.","why_it_reads_ai":"An image generator producing letterforms is drawing the texture of technical writing rather than spelling anything. The published case is a paper retracted three days after publication in February 2024, whose figures carried labels that looked like signalling-pathway vocabulary and were not words in any language. The authors had disclosed the tool in the text. It passed review anyway, which is the part worth remembering: the labels were visible to anyone who looked at the figure.","examples":[{"before":"Figure 2 labels the three stages of the pipeline as Preproccenting Layer, Feature Extructation and Ouput Nomalisation, and the caption calls the diagram self-explanatory.","after":"Figure 2 labels the three stages as tokenisation, feature extraction and normalisation. The caption names who drew it and from which run of the data, and the alt text repeats the three labels so a screen reader gets the same words a sighted reader does.","note":"The strings in the first version are invented for this specimen, in the shape the failure takes. The repair also fixes the accessibility hole that let the labels go unread."}],"detection":{"type":"judge","rubric":"Task: read every string inside figures, diagrams, charts, screenshots and generated illustrations, including axis labels, legends, node names and watermarks.\n\nStep 1. Transcribe each token exactly as rendered.\nStep 2. Classify each token as one of: dictionary word; known technical term; proper noun; abbreviation whose expansion appears in the surrounding text; identifier such as a gene, chemical, part or product code; none of these.\nStep 3. Tokens landing in \"none of these\" are pseudo-lexical candidates. Check the near-miss cases: a real word with one letter wrong is a typographic error, while a string with plausible morphology and no lemma behind it is the pattern.\n\nEscape hatches: optical character recognition of a scanned page, where errors cluster and follow the scan quality; font substitution failure, which mangles a whole block uniformly rather than one label; deliberate nonsense, placeholder text and lorem ipsum; a language or script the reviewer does not read, which returns UNKNOWN; hand-drawn figures photographed at low resolution.\n\nDecision. FLAG when two or more pseudo-lexical tokens appear in one figure, or when one appears in a figure title or a heading. Otherwise PASS.\n\nOutput: the tokens, their classification, and the verdict. Note that this check reads images, which the rest of this index does not."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"A journal retraction three days after publication, covered independently, with the mangled figure labels reproduced in the coverage. The entry is kept despite falling outside the text-only scope, and it declares that exception in its own description."}],"evidence_grade":"corroborated","false_positive_notes":"Scanning and font handling produce identical strings for entirely mechanical reasons. Optical character recognition on a poor scan invents morphology that looks exactly like this, and a missing font substituted at print time can turn a whole diagram into plausible-looking rubbish. Both are print-production failures with nothing to do with how the text was written. Specialist identifiers also defeat a naive reader: gene names, chemical registry numbers, part codes and internal build labels are not dictionary words and never were. Reviewers outside the field should record what they saw and ask the authors rather than score the figure, and any check that reads images should say so, since it is a different kind of evidence from reading prose.","model_attribution":"Image generators rather than text models. The retracted paper disclosed the image tool in its own text, which is unusual and makes the attribution unusually solid for this dataset.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Science Integrity Digest, the Frontiers rat figure retraction","url":"https://scienceintegritydigest.com/2024/02/15/the-rat-with-the-big-balls-and-enormous-penis-how-frontiers-published-a-paper-with-botched-ai-generated-images/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Gizmodo, science journal retracts AI-generated images","url":"https://gizmodo.com/science-journal-rat-dck-ai-generated-images-retracted-1851297606","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"unfalsifiable-byline","name":"Unfalsifiable byline","aka":["vivid bio with no footprint","author who exists nowhere else"],"category":"semantic","subcategory":"provenance","description":"A byline with a warm, specific biography and no verifiable existence anywhere else. Detail is cheap, so the fabricated version has more of it than the real one. Genuine thin bylines tend to be vague. The invented ones come with hobbies.","why_it_reads_ai":"A biography is generated from the same distribution as the article and costs nothing extra, so specificity stops being evidence of anything. In the November 2023 magazine case the portraits traced to a marketplace of generated faces and the bios described childhoods nobody could check; when a reporter asked about the writers, they were removed and the articles reassigned to other names that also did not exist. In the 2025 freelance case at least six outlets retracted work under one byline. Two commercial detection tools had cleared that copy. What caught it was a payment process the writer could not complete and places in the text that do not exist.","examples":[{"before":"Tom Ridgeway grew up repairing outboard motors on his grandparents' lake in northern Michigan and has been writing about boats for over a decade.","after":"Tom Ridgeway has written for us since March. His earlier work is at the two outlets linked from his name, he files to a named editor here, and he is paid through the same system as everyone else on the masthead.","note":"The writer in both versions is invented for this specimen. The repair replaces charm with a trail: prior work, an editor, a payment route."}],"detection":{"type":"judge","rubric":"Task: decide whether the byline has an existence outside this page.\n\nRun four checks and report each separately.\n1. Prior work. Does anything under this name exist elsewhere, dated before this piece?\n2. Independent presence. A professional profile, a conference programme, a company or court record, a library catalogue, a union or association listing.\n3. Portrait. Does the photograph reverse-search to a stock library or a generated-portrait marketplace?\n4. Institutional trail. Can the publisher name an editor who commissioned the piece, a payment route that completed, and a contactable address?\n\nEscape hatches: declared pseudonyms, which are legitimate and common; writers in jurisdictions where being named is dangerous; first-time writers, who fail check 1 by definition; house and staff bylines; disclosed ghostwriting; people who keep no professional web presence, which is a choice rather than a sign.\n\nDecision. FLAG only when a vivid, specific biography coexists with zero result on checks 1 and 2 AND the publisher cannot complete check 4. Detector output is not part of this rubric and must not be used: two commercial tools cleared the documented fabrication.\n\nOutput: the four check results and FLAG or PASS. The finding is about a publishing process, never about prose style."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Two separate newsroom cases, in 2023 and 2025, documented by three outlets between them, with retractions in both. The second case also records that commercial detection tools cleared the copy, which is why this entry routes verification through payment and editorial trails instead."}],"evidence_grade":"corroborated","false_positive_notes":"Pseudonymous writing is old, legal and often necessary. Journalists covering organised crime, writers in countries where a byline is a risk, people writing about their own health or their employer, and anyone who has left an abusive relationship all have reason to publish under a name with no trail behind it. First-time writers fail the prior-work check by definition, and plenty of competent people keep no professional web presence at all. Reverse image search returns false hits on ordinary portraits. This is why the check runs through the publisher rather than the prose: an editor who commissioned the piece and a payment that completed answer the question, and neither requires anyone to judge the writing.","model_attribution":"Not a model tell at all. It is a publishing tell. The generated portraits in the 2023 case came from an image marketplace, and no vendor attribution exists for the biographies.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Futurism, Sports Illustrated published articles by fake AI-generated writers","url":"https://futurism.com/sports-illustrated-ai-generated-writers","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Press Gazette, Wired and Business Insider remove AI-written freelance articles","url":"https://pressgazette.co.uk/publishers/digital-journalism/wired-and-business-insider-remove-ai-written-freelance-articles/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"CNN, Sports Illustrated deletes articles with fake author names","url":"https://www.cnn.com/2023/11/27/media/sports-illustrated-deletes-articles-fake-author-names-ai-profile-photos/index.html","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"unverifiable-specific-geography","name":"Unverifiable specific geography","aka":["towns that do not exist","invented venue datelines"],"category":"semantic","subcategory":"provenance","description":"Named small towns, venues and local officials that do not exist. The names are well formed for the region and the population figures are plausible. Nothing in a gazetteer answers to them.","why_it_reads_ai":"Place names have strong regional morphology, so producing a convincing one is easy and producing a real one is not the same task. In the 2025 freelance fabrication the invented settlements were among the things that finally surfaced the fraud, alongside a payment process the writer could not complete. Two commercial detection tools had already cleared the copy. That ordering is the single strongest argument in this dataset against leaning on tools: the fabricated geography was checkable by anyone with a map, and the software was not.","examples":[{"before":"In Kellenmoor, a former mill town of about 900 in the eastern hills, the parish clerk has run the same October ceremony every year since 1974.","after":"The ceremony has run every October since 1974 in a village of about 900 people. The parish clerk is named in the piece, the register entry giving the first year is photographed in the archive, and the population comes from the municipality's 2024 statistics page, which is linked. A reader who doubts any of it can open the sources.","note":"Kellenmoor is invented for this specimen. The repair does not swap in a better-sounding place name. It adds the route by which a reader checks the one that is there."}],"detection":{"type":"judge","rubric":"Task: check that every named place and local office exists.\n\nStep 1. List every settlement, venue, building, road, institution and public office named in the text, with the events attributed to each.\nStep 2. Look each up in a gazetteer, a national or municipal register, a company register, a map service and the relevant official directory. Record found, not found, or unreachable.\nStep 3. Cross-check the attributes: population figures against census or municipal statistics, an office title against the body that would hold it, a venue against the address given.\n\nEscape hatches: composite or anonymised locations, which are legitimate in protected-source journalism and must be declared in the piece rather than inferred by a reader; historical names now changed or absorbed; transliteration and diacritic variants; hamlets and features that mapping services omit; informal local names for real places; fiction.\n\nDecision. FLAG when a named settlement or public office returns nothing in any register and the text attributes events to it. Return UNKNOWN when the registers are unreachable.\n\nOutput: each place, its register result, and the verdict. Do not use commercial detection tools as part of this check. In the documented case two of them cleared the fabricated copy."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Six outlets retracted work under one fabricated byline in 2025, and the trade coverage records nonexistent towns and payment friction as what exposed it after two detection tools had cleared the text. The check is a lookup, which makes it one of the few reliable ones here."}],"evidence_grade":"corroborated","false_positive_notes":"Composite and anonymised locations are standard practice where naming a place would identify a source or a patient, and a piece that uses one should say so rather than leave a reader to guess. Registers are also incomplete and inconsistent: hamlets vanish from map services, municipalities merge and rename, transliteration produces four spellings of one town, and small venues appear under a licensee's name rather than their own. Historical writing is full of places that no longer exist under any name. The verdict for an unreachable register is unknown, and a reviewer who cannot read the national register for that country should say so instead of scoring it.","model_attribution":"No vendor can be named. The documented case never established which tool produced the copy, and the fabricated geography is a property of writing about places without looking at any.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Press Gazette, the Margaux Blanchard fabrication","url":"https://pressgazette.co.uk/publishers/digital-journalism/margaux-blanchard-wired-ai-editors-journalist/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Press Gazette, Wired and Business Insider remove AI-written freelance articles","url":"https://pressgazette.co.uk/publishers/digital-journalism/wired-and-business-insider-remove-ai-written-freelance-articles/","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"undue-significance-inflation","name":"Undue significance inflation","aka":["WP:AITREND","legacy framing","setting the stage for","grandiose stakes"],"category":"semantic","subcategory":"significance","description":"An arbitrary detail is puffed into a contribution to a larger story. A second office becomes a turning point in an industry. A routine appointment becomes part of a broader movement. The larger story is never named precisely enough to be wrong.","why_it_reads_ai":"Training text describes notable subjects in language that already assumes their notability, so an unremarkable fact gets fitted with the sentence shape reserved for remarkable ones. Wikipedia keeps a whole section on the habit and a words-to-watch list for it, and a community pattern directory files the same move under inflated stakes. The specific fact tends to go missing in the same sentence: the inventor of one device becomes a titan of an industry, which is louder and says less.","examples":[{"before":"The company opened its second office in 2019, setting the stage for a decade of regional growth and reflecting a wider movement toward distributed engineering across the sector.","after":"The company opened a second office in 2019 because its two most senior engineers had moved to Oulu and did not want to fly every week. Six people work there now. Whether that matters to anyone outside the company is not something this paragraph can settle.","note":"The repair replaces a claim about the sector with the reason the thing happened. It is smaller and it is checkable."}],"detection":{"type":"judge","rubric":"Task: find significance claims with nothing under them.\n\nStep 1. Mark every clause asserting that a detail contributes to, reflects, shapes, marks, sets up or represents something larger than itself.\nStep 2. For each, look inside the text for support: a measurement, a named consequence, a cited source, a comparison with a stated baseline.\nStep 3. Mark each claim SUPPORTED or BARE.\nStep 4. Judge the subject: is the underlying fact routine for its category? A founding date, an office opening, a staff appointment and a product update are routine unless the text shows otherwise.\n\nEscape hatches: obituaries, award citations, anniversary features, institutional histories and nominations, where a significance frame is the commission; texts that attribute the significance claim to a named source, which makes it that source's claim; explicit argument, where the piece sets out to establish the importance and does the work.\n\nDecision. FLAG when two or more claims are BARE and the underlying facts are routine. WEAK on one. PASS when claims are supported or the genre calls for them.\n\nOutput: each claim, its mark, and the verdict."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Named in the largest maintained public taxonomy with its own shortcut and a words-to-watch list, and independently in a community pattern directory and a rule set. Three sources describing one move."}],"evidence_grade":"corroborated","false_positive_notes":"Several genres are commissioned to do exactly this. An obituary is supposed to place a life in a larger story, an award citation exists to argue that a contribution mattered, and an anniversary feature is an assignment to find broader meaning in a date. Institutional histories, funding applications and nomination letters all carry the same brief. Historians make significance claims for a living and support them at book length rather than in the sentence. Reviewers should also allow for cultures where understatement is not the default register, and for writing translated from one of them. The test is whether support exists anywhere in the piece, not whether the sentence sounds grand.","model_attribution":"Wikipedia describes the habit as general to language models rather than to one vendor, and links it to how notable subjects are described in training text. No per-vendor measurement exists.","platform_notes":[{"platform":"wikipedia","note":"The relevant section carries the shortcut WP:AITREND and a words-to-watch list. Editors treat it as a cleanup signal in encyclopedic register, where significance framing is against the house style to begin with."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"canned-notability-emphasis","name":"Canned notability emphasis","aka":["WP:OVERATTRIBUTION","covered by national outlets","active social media presence"],"category":"semantic","subcategory":"significance","description":"The text lists what kinds of sources covered a subject instead of what those sources said. Trade publications, regional media, independent coverage. The categories arrive; the content does not.","why_it_reads_ai":"Asked to establish that someone matters, a model reaches for the vocabulary of the criteria rather than the evidence behind them, and on Wikipedia it often echoes the wording of the notability guideline back at the reader. Wikipedia records the habit as more common in tools released from 2025 onward, which makes it one of the few entries in this dataset pointing forward rather than back. The social-media variant is the tell in miniature: a claim that a subject is present online, which is true of everyone and describes nothing.","examples":[{"before":"Aino Verkkola is a designer whose work has been featured in numerous national media outlets and who maintains an active social media presence across several platforms.","after":"Aino Verkkola designed the wayfinding system for a Tampere hospital that opened in 2023, and the hospital's project page names her and shows the drawings. Two of her signs were replaced within a year because staff kept walking past them, which she has written about.","note":"The designer is invented for this specimen. The repair drops the categories of coverage and gives one project, one date and one thing that went wrong."}],"detection":{"type":"deterministic","pattern":"(?:\\b(?:maintain|maintains|maintained|has|have)\\s+an\\s+active\\s+(?:social[- ]media|online)\\s+presence\\b)|(?:\\b(?:featured|profiled|covered)\\s+in\\s+(?:a\\s+(?:number|variety)\\s+of|several|numerous|various|multiple|many)\\s+(?:(?:national|regional|local|international|major|prominent|leading)\\s+)?(?:media\\s+|news\\s+|trade\\s+)?(?:outlets|publications|newspapers|magazines)\\b)|(?:\\breceived\\s+(?:widespread\\s+|significant\\s+|extensive\\s+)?independent\\s+coverage\\b)","flags":"gi","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Wikipedia names the pattern, quotes examples from 2025 revisions and records that it is more common in later models. Single-source, so the grade stays community-observed, and the pattern is written narrowly because the phrasing overlaps with ordinary publicity writing."}],"evidence_grade":"community-observed","false_positive_notes":"Public-relations biographies, speaker one-sheets, grant applications and award nominations are built from this move, and it is the genre contract in all four. A publicist who writes that a client has been profiled in several trade publications is doing the job as briefed. Media-studies writing describes coverage by outlet type as a matter of method, and so does any piece about how a story spread. Wikipedia editors themselves cite source categories when arguing notability at deletion discussions, which is the guideline working as intended. The pattern here is deliberately narrow, matching only the vaguest forms, because the same words in a sentence that names the outlet are ordinary reporting.","model_attribution":"Wikipedia records this as more common in text from tools released in 2025 or later. That is a community observation over article revisions rather than a controlled measurement, and it is the only dating available.","platform_notes":[{"platform":"wikipedia","note":"Filed under the shortcut WP:OVERATTRIBUTION. The page notes that human press releases have cited news clippings for decades, and that what marks the machine version is echoing the wording of the notability guideline itself."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"promotional-register-misplacement","name":"Promotional register in the wrong genre","aka":["WP:AIPUFFERY","unanchored claims","brochure warmth in a report","genre glitch"],"category":"semantic","subcategory":"register","description":"Advertising register arrives in prose whose genre forbids it. An incident report reassures. An encyclopedia entry invites you to visit. A methods section describes its own approach as exciting.","why_it_reads_ai":"Marketing prose is heavily represented in training text and its register is the default gravity, so a model asked for neutral writing drifts back toward it. A researcher writing in 2026 named the phenomenon a genre glitch, where a phrase belonging to one genre surfaces in another. Wikipedia reports the same drift even when editors prompt for encyclopedic style, and notes an edit summary claiming to have removed promotional tone while introducing it. Microsoft's own guidance to writers aiming at AI search answers tells them to avoid claims with nothing behind them, which is the same advice from the other direction.","examples":[{"before":"Our platform, known for its reliable performance, experienced a brief service interruption on Tuesday. We remain committed to delivering an outstanding experience for every one of our customers.","after":"The API returned server errors for 43 minutes on Tuesday, starting at 09:12 UTC. A migration dropped an index and every read on the billing table fell back to a full scan. We have added a check that fails the deploy when an expected index is missing. The post-mortem is linked.","note":"Figures and timings in this repair are invented for the specimen. The genre asks what happened, and the first version answers a different question."}],"detection":{"type":"judge","rubric":"Task: decide whether promotional register has arrived in a genre that forbids it.\n\nStep 1. Name the genre from its own conventions: incident report, encyclopedia entry, methods section, minutes, court filing, obituary, syllabus, changelog, marketing page, sales email.\nStep 2. Scan for promotional moves. Count each occurrence:\n  (a) superlatives about the subject with no source or measurement;\n  (b) warmth adjectives applied to a product, place or organisation;\n  (c) commitment and reassurance language where the genre asks for facts;\n  (d) invitations to buy, visit, join or explore;\n  (e) second-person address in a genre that does not use it.\nStep 3. Note whether the promotional sentences carry any checkable content at all.\n\nEscape hatches: the genre is marketing, where all of this is correct; promotional wording quoted from a party inside a news piece or a filing; a press release reproduced and labelled as one; nonprofit and campaign writing, which is persuasive by design; cultures and languages whose formal register is warmer than English business prose.\n\nDecision. FLAG when two or more promotional moves appear in a genre whose conventions forbid them. WEAK on one. PASS in marketing genres.\n\nOutput: the genre, the moves counted with quotes, and the verdict."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Wikipedia documents the drift with dated examples, a named researcher describes the same effect as a genre glitch, and a vendor guidance page names unanchored claims from the writing side. Three independent descriptions of one behaviour."}],"evidence_grade":"corroborated","false_positive_notes":"Marketing copy is supposed to read like marketing copy, and judging a landing page by the conventions of an incident report is a category error. Writers new to a genre import the register they already know, which is why a first post-mortem from someone out of a sales team reads like a press release with no tool involved. Nonprofit and campaign writing is persuasive by design and answerable to a different standard. Business registers also differ by language and culture, and prose translated from a warmer one arrives sounding promotional in English when it was neutral at home. The question is always what the genre demands, never whether the sentence is enthusiastic.","model_attribution":"Wikipedia notes that older models produced more openly positive text and that later ones avoid obvious superlatives while keeping the warmth, citing a 2025 study of language complexity and sentiment. That is a direction of travel rather than a per-vendor measurement.","platform_notes":[{"platform":"wikipedia","note":"Filed under WP:AIPUFFERY, with a words-to-watch list and dated article examples. The page states plainly that not all promotional writing is machine-generated."},{"platform":"bing","note":"Microsoft advertising guidance for content aimed at AI search answers tells writers to avoid claims with nothing supporting them. It is advice, published with no data and no ranking claim attached."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Walker Rettberg, genre glitches and unexpected promotional phrases","url":"https://jilltxt.net/genre-glitches-and-unexpected-promotional-phrases-as-a-sign-of-ai-writing/","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Microsoft Advertising, optimizing content for inclusion in AI search answers","url":"https://about.ads.microsoft.com/en/blog/post/october-2025/optimizing-your-content-for-inclusion-in-ai-search-answers","accessed":"2026-08-14","tier":"vendor"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"internal-factual-contradiction","name":"Internal factual contradiction","aka":["conflicting facts in one piece","confident instruction carrying wrong facts"],"category":"semantic","subcategory":"factual-failure","description":"Two claims inside one document that cannot both be true, delivered in a register that never notices. The instructional voice is what makes it dangerous, because it tells the reader to stop checking.","why_it_reads_ai":"Each sentence is generated to be locally plausible, and nothing in that process holds the document's claims against each other. The nearest primary document is the March 2026 coalition letter to YouTube about machine-made video aimed at young children. It is cited here for what it actually states: that many AI videos are labelled educational, that studies put the share of educational-labelled YouTube videos carrying high-quality educational content at about 5 percent, and that a 2023 BBC investigation found false science information from AI videos being recommended to older kids as educational. The letter never describes two claims conflicting inside one video. No source in this entry's set documents a contradiction within a single document, so that half of the entry is this index's own reading and the grade says so. Numbers attached to the same story in later coverage could not be traced to a primary source, so none of them appear here.","examples":[{"before":"The lake freezes over in late November, and the ice fishing season opens in October once the lake has frozen and the ice is safe.","after":"The lake usually freezes in late November. The season opens when the ice reaches 10 centimetres, which in most years falls in early December. In 2024 it did not happen until January and the season was three weeks short.","note":"Dates and thicknesses in this repair are invented for the specimen. The first version cannot be followed, because its two halves disagree about when the water is solid."}],"detection":{"type":"judge","rubric":"Task: find claims inside the document that cannot both be true.\n\nStep 1. Build a fact table. One row per atomic claim, with subject, attribute, value and any date.\nStep 2. Group rows sharing a subject and attribute.\nStep 3. Compare within each group for three conflict types:\n  (a) value conflict, where one quantity is given two incompatible numbers;\n  (b) ordering conflict, where an event is placed both before and after another;\n  (c) definition conflict, where a term is used with two incompatible meanings.\n\nEscape hatches: explicit revision, where the text says it previously stated something else; quotations from parties who disagree, where the conflict belongs to them; approximations and ranges that overlap once tolerance is allowed; different units or scales that reconcile on conversion; claims about different times that the text dates.\n\nDecision. FLAG on one unexplained conflict. One is enough, because a reader cannot use the document without knowing which half to believe. Report both rows.\n\nOutput: the conflicting pair, the conflict type, and FLAG or PASS."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Ships active on the mechanism rather than on a measurement. The March 2026 coalition letter to a platform is cited for the register and the population it describes, and it does not attest a contradiction inside a single document, so the evidence grade is our own corpus. Figures circulating alongside that story could not be traced to a primary source and are not used here."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"Multi-author documents contradict themselves as a matter of routine, and so do single-author ones edited over months. A specification updated in one section and not another, a report where the summary predates the appendix, a wiki page three people maintain: all produce this and none of it says anything about tooling. Sincere amateur educators publish errors too, and being wrong is not the same as being generated. Approximations also collide harmlessly once tolerance is allowed, and figures given in different units reconcile on conversion. What the check produces is a document that needs fixing, which is a useful finding on its own terms and not a conclusion about who wrote it.","model_attribution":"No vendor can be named. The concern documented in 2026 covers video produced with a range of tools, and no primary source in our set names a model family.","platform_notes":[{"platform":"youtube","note":"The March 2026 coalition letter to the platform describes machine-made video aimed at young children that presents itself as educational. The platform's own inauthentic-content policy is authorship-indifferent and turns on templating and mass production rather than on tooling."}],"languages":["en"],"sources":[{"kind":"external","title":"Fairplay coalition open letter to YouTube on AI content aimed at children (Mar 2026)","url":"https://fairplayforkids.org/wp-content/uploads/2026/03/YouTube-Letter-AI-Slop.pdf","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Tubefilter, YouTube and the Fairplay kids AI open letter","url":"https://www.tubefilter.com/2026/04/01/youtube-fairplay-kids-ai-open-letter/","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"invented-concept-label","name":"Invented concept labels","aka":["capitalised fake framework","coined term presented as established"],"category":"semantic","subcategory":"coinage","description":"A coined, capitalised framework presented as an established term of art. Three named laws, a labelled matrix, a phase model, arriving with a definite article and no history.","why_it_reads_ai":"Naming a concept is a low-cost way to sound like the literature, and the capital letters do the work that a citation would. A community pattern directory names the move in its list of writing patterns. Coinage itself is fine and necessary. What marks this version is the absence of any acknowledgement that the term is new, which is the one sentence an honest coiner always writes.","examples":[{"before":"This is the Trust Gradient at work: attention falls as specificity rises, an effect every practitioner in the field will recognise the moment it is described to them. The framework has been discussed in the literature for years, and versions of it turn up under other names in adjacent disciplines. Once you see the gradient you begin to see it everywhere, and accounting for it early is the difference between a decision that ages well and one that does not.","after":"I am going to call this the trust gradient. It is my name for it and nobody else uses it. What I mean is narrow: across four onboarding tests, the pages that named a price drew fewer visitors and produced more signups. Four tests is not an effect and it is certainly not a law.","note":"Tests and results in this repair are invented for the specimen. The specimen leans on a literature and a recognition that are never named, which is the whole move. The repair keeps the coinage and drops the pretence that it has a history."}],"detection":{"type":"judge","rubric":"Task: find coined terms presented as established ones.\n\nStep 1. List every capitalised multi-word noun phrase presented as a concept, principle, law, effect, framework, matrix, model or curve.\nStep 2. For each, check the text for attribution or acknowledgement: a citation, a named originator, or a sentence marking the term as the author's own.\nStep 3. Search for prior use outside this text. Record found, not found, or unreachable.\nStep 4. Note the framing: a definite article plus a claim that practitioners recognise the term is the shape this tell takes.\n\nEscape hatches: the author explicitly claims the coinage, which is the honest form and passes; registered product, method or certification names; terms genuinely used in the field that the reviewer does not know, which return UNKNOWN; internal company vocabulary in an internal document; translations of an established term from another language.\n\nDecision. FLAG when a term has a definite article, no attribution, no prior use, and a claim of recognition. WEAK when three of those four hold.\n\nOutput: the term, the four checks, and the verdict."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Named by one community pattern directory and nowhere else in our source set, so the grade is community-observed and the entry says so. The acknowledgement test is what keeps it from condemning ordinary coinage."}],"evidence_grade":"community-observed","false_positive_notes":"Coinage is how vocabulary grows, and every term of art was invented by somebody who had to use it before anyone else recognised it. Consultants, researchers and practitioners name things for good reasons: a name makes a pattern discussable. Internal company vocabulary looks exactly like this from outside and is perfectly legitimate inside. Reviewers also routinely fail to recognise real terms from adjacent fields, and translated terminology arrives looking invented. The check that survives all of this is whether the text acknowledges the term as new. A writer who says this is my name for it has done nothing wrong, whatever the name is.","model_attribution":"Undocumented per vendor. The directory that names the pattern does not attribute it to a family, and no measurement exists.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"tropes.fyi pattern directory","url":"https://tropes.fyi/directory","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"irreproducible-confident-report","name":"Confident report, no reproducible artifact","aka":["well-formatted nothing","bug report that does not reproduce"],"category":"semantic","subcategory":"factual-failure","description":"A technical report with every section in the right place, describing something that does not happen. Steps to reproduce that reproduce nothing. A severity rating for a condition nobody can trigger.","why_it_reads_ai":"Report structure is a template and the artifact is not. The structural analogue of the fake citation: correct form, missing referent. The curl project closed its bug bounty programme in January 2026 under a flood of machine-written submissions, and the press account records that in the final week none of the reports described a real vulnerability. The cost of this tell falls entirely on reviewers, which is the asymmetry that makes it worth cataloguing: minutes to produce, hours to refute.","examples":[{"before":"Steps to reproduce: send a request with a malformed Host header. The parser dereferences a null pointer and the process exits. Impact: remote denial of service. Severity: high.","after":"Steps to reproduce: on version 4.2.1 built with the flags in the attached log, run the attached script, which sends 200 requests with an empty Host header. The process exits on request 137 in three runs out of three. Core dump attached. On 4.1.9 it does not happen, so the change is somewhere between those two tags.","note":"Versions and counts in this repair are invented for the specimen. The repair adds an artifact, a version boundary and a repeat count, which is what makes a report actionable."}],"detection":{"type":"judge","rubric":"Task: decide whether the report contains anything a reviewer could run.\n\nStep 1. Check for an artifact. Score each present or absent: a script, payload or input file; a capture or log; a stack trace or crash dump; the exact version and build; the environment, including operating system and compiler or runtime; a repeat count across runs.\nStep 2. Check internal consistency. Does the described mechanism produce the described symptom? Does the severity rating follow from the described impact?\nStep 3. Check specificity. Are line numbers, function names and version boundaries given, and do they exist in the code as published?\n\nEscape hatches: reports under embargo that say so and offer the artifact privately; environment-specific issues where the reporter states the environment and the limits of what they can share; first-time reporters asking a question rather than asserting a finding; reports about closed-source systems where an artifact cannot be shared.\n\nDecision. FLAG when the report asserts a confirmed vulnerability or defect while missing an artifact, a version and an environment. Return the missing fields as a checklist the reporter can complete.\n\nOutput: the artifact scorecard, the missing fields, and FLAG or PASS. A report with no artifact is unactionable, which is a statement about the report and not about the person who filed it."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"A widely reported programme closure in January 2026, covered by two independent outlets, with the maintainer describing a flood of machine-written reports. Rates circulating with that story could not be confirmed in the cited coverage and are not used here."}],"evidence_grade":"corroborated","false_positive_notes":"Novice security researchers file honest false positives constantly, and a programme that punishes them stops hearing from the people who later find real things. Bugs are also environment-specific in real cases: a race condition that appears on one scheduler, a fault that needs particular hardware, a failure that only shows under load nobody else can generate. Reporters working on closed systems cannot always share an artifact, and reporters in some jurisdictions have legal reasons to be careful about what they attach. The rubric therefore produces a checklist of missing fields rather than a verdict, because a report that becomes actionable after one more round was worth the round.","model_attribution":"Undocumented per vendor. The reports in the documented case were submitted through a public programme and no tool was identified.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"The Register, curl shutters bug bounty program to stop AI slop","url":"https://www.theregister.com/security/2026/01/21/curl_shutters_bug_bounty_program_to_stop_ai_slop/5063039","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"BleepingComputer, curl ending bug bounty program after flood of AI slop reports","url":"https://www.bleepingcomputer.com/news/security/curl-ending-bug-bounty-program-after-flood-of-ai-slop-reports/","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"low-profile-speculation","name":"Low-profile speculation","aka":["maintains a low profile","little is known about"],"category":"semantic","subcategory":"gap-narration","description":"A gap in the sources is narrated as a fact about the subject. Nothing was found, so the text reports that the person keeps to themselves. The claim is about the research and it arrives dressed as biography.","why_it_reads_ai":"Retrieval that returns nothing produces a sentence anyway, and the cheapest sentence available says the absence is meaningful. Wikipedia files this with the behaviour it describes around training windows, noting that where a system fails to find sources it may state that the information is not publicly available and then speculate about what it likely is and why that matters. Both halves are invention, including the claim that nothing is documented. The check is one question: is a source attached to the absence.","examples":[{"before":"Little is known about Rautio's early career, and she appears to maintain a low profile outside her professional work. Her education is similarly undocumented, which suggests she preferred to let the work speak for itself. What can be said is that she has always chosen substance over visibility, and those who have worked with her seem to have respected it.","after":"We could not document Rautio's career before 1998. The trade register lists her first company from that year and nothing earlier. The polytechnic alumni office confirms her degree and holds no record of anything before 1996. We wrote to her former employer on 3 June and had no reply. The gap is in our research and we are stating it as ours.","note":"The subject and the institutions around her are invented for this specimen. Every sentence after the first turns another empty search into a character trait. The repair converts a claim about a person into a claim about the search, which is the only claim the writer can support."}],"detection":{"type":"deterministic","pattern":"(?:\\b(?:maintain|maintains|maintained|keep|keeps|kept)\\s+a\\s+(?:relatively\\s+|famously\\s+|notably\\s+|fairly\\s+)?low\\s+profile\\b)|(?:\\blittle\\s+is\\s+(?:publicly\\s+)?known\\s+about\\b)|(?:\\b(?:keeps|prefers\\s+to\\s+keep)\\s+(?:her|his|their)\\s+(?:personal\\s+(?:life|details)|private\\s+life)\\s+(?:private|out\\s+of\\s+the\\s+public\\s+eye)\\b)|(?:\\b(?:is|are)\\s+not\\s+widely\\s+documented\\b)","flags":"gi","scope":"sentence"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Wikipedia names the phrasing and quotes dated article revisions using it where no source existed. One source, so the grade is community-observed. The check is cheap, which is why the entry is deterministic rather than a rubric."}],"evidence_grade":"community-observed","false_positive_notes":"Biographers of private people write this sentence all the time, with sourcing behind it. Someone who has declined every interview for thirty years does keep a low profile, and saying so is reporting rather than speculation when a source supports it. Obituary writers and archivists describe documented absences as a matter of course, and historians of the poor and the colonised write about record gaps because the gaps are the subject. Legal and safeguarding contexts require exactly this wording. What separates the two is whether a citation sits next to the sentence. A reviewer who flags a sourced one has misread this entry.","model_attribution":"Wikipedia groups it with the training-window behaviours, which cover older systems with a fixed cutoff and newer ones whose retrieval step returns nothing. No vendor attribution is available.","platform_notes":[{"platform":"wikipedia","note":"Listed among the signs that a passage was produced without sources, alongside the older boilerplate about training windows. The page notes that the speculation extends to the claim that the information is undocumented."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"invented-internal-reference","name":"Invented internal references","aka":["hallucinated policy shortcut","citing a rule that does not exist"],"category":"semantic","subcategory":"fabrication","description":"A confident reference to a rule, page, ticket or precedent inside an organisation that does not exist. The shortcut looks right. The section number looks right. The page is not there.","why_it_reads_ai":"Internal reference formats are highly patterned, which makes them easy to produce and impossible to check from inside the text. Wikipedia documents the behaviour on its page about machine-written discussion comments, under a shortcut of its own, recording cases where a chatbot attributed invented policies to real project pages and even to pages that were never meant to be cited that way. The damage is that a fabricated rule closes an argument, and the person on the other side usually assumes they simply had not read it.","examples":[{"before":"Per section 4.2 of the editorial handbook, we do not publish vendor benchmarks without a second independent source. This has been settled policy for a while and the reasoning is covered in the onboarding deck under the sourcing standards page. It is one of those rules everyone forgets until it matters, and then it saves the piece. Worth a reread before the next draft goes out.","after":"The editorial handbook says nothing about vendor benchmarks, and the onboarding deck has no sourcing page in it. I checked both before writing this. I think the rule should exist, and the draft wording is in the pull request linked here. Until that lands there is no rule and nobody has to follow it, including me.","note":"The handbook section and the deck page are invented for this specimen. Two references, neither resolvable, and the second is there to make the first sound routine. The repair does not soften the claim. It withdraws it and says what exists instead."}],"detection":{"type":"judge","rubric":"Task: check that every internal reference resolves.\n\nStep 1. Extract every reference to an internal rule, policy, handbook section, shortcut, template, ticket, runbook, precedent or prior decision.\nStep 2. Resolve each in the organisation's own systems. Record found, redirected, or missing.\nStep 3. For those found, check three things: that the cited section number exists in that document, that the quoted wording matches, and that the passage says what the citing text claims.\nStep 4. Note the rhetorical role. A reference used to end a discussion carries more weight than one used in passing, and should be checked first.\n\nEscape hatches: renamed or merged pages where the old name redirects, which is a stale reference rather than an invented one; policies quoted from memory and flagged as such; drafts marked as drafts; references to another organisation's rules; oral convention that has never been written down, which the speaker should say plainly.\n\nDecision. FLAG when a reference resolves to nothing and is used to settle a question. WEAK when it resolves but the wording does not support the claim.\n\nOutput: each reference, its resolution, and the verdict."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Wikipedia documents invented policy citations on its page about machine-written comments, with its own shortcut and quoted examples. Single-source, so the grade is community-observed, and the entry is scoped to internal rule sets to keep it distinct from the citation entries."}],"evidence_grade":"community-observed","false_positive_notes":"People misremember their own policies constantly. Long-tenured staff cite rules from memory, quote a section number from a version three revisions old, or repeat something a manager said in 2019 as though it were written down. Pages get renamed, merged and archived, and a reference that resolved last year can be dead this year through nothing but housekeeping. Oral convention is real and often correct even where nothing was ever documented. The distinction that matters is whether the reference resolves to something once real and whether the person will withdraw it when shown that it does not, which is a conversation rather than a verdict.","model_attribution":"Documented on Wikipedia as behaviour of chatbots generally, with examples spanning more than one assistant. No vendor-level measurement exists.","platform_notes":[{"platform":"wikipedia","note":"Recorded on the page about machine-written comments under its own shortcut, covering fictitious policies attributed to real project pages. Editors treat a nonexistent shortcut as one of the more reliable signs in discussion text."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI-generated comments","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI-generated_comments","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"fabricated-precision","name":"Fabricated precision","aka":["invented statistic","precise number with no traceable source"],"category":"semantic","subcategory":"statistics","description":"A number with a decimal point and no parent. It carries the impression of measurement without any of the apparatus: no sample, no method, no date, no owner. Trace it and you find another article that also does not say.","why_it_reads_ai":"Precision is a style before it is a measurement, and the style is cheaper. A figure like 47 percent reads as researched in a way that most reads as guessed, so the generated sentence reaches for the first. Our own folklore register is full of this: the platform-update figures that circulate about reach penalties and detection accuracy resolve, one after another, to vendor and marketing posts citing each other, and none of them appears in any first-party platform document. No external taxonomy in our source set names this as a pattern, so the entry rests on our own corpus and says so.","examples":[{"before":"Posts published on Tuesday mornings see 31.4 percent more engagement, and the ranking system weights them 2.3 times higher during the first hour.","after":"Our Tuesday morning posts did better than our Thursday ones across eleven weeks: median 41 reactions against 26. We do not know why. We have no access to how the ranking works, so the second half of that sentence would have been a guess and is not here.","note":"Figures in both versions are invented for this specimen. The first version's problem is that its numbers have no parent. The second version's numbers are ours and carry the counting method with them."}],"detection":{"type":"judge","rubric":"Task: decide which numbers in the text have anything behind them.\n\nStep 1. List every numeric claim about the world. Exclude dates, list numbering, prices of things the text is selling, and quantities the reader can see for themselves.\nStep 2. For each, record four things: a source, a sample or population, a date or date range, a method. Mark each present or absent.\nStep 3. Note the precision. A decimal place, or a percentage stated more finely than five points, is a claim to measurement.\n\nEscape hatches: numbers the text marks as illustrative, hypothetical or rounded; constants a reader can look up, such as a tax rate or a physical value; figures inside a quotation attributed to a named speaker; the author's own numbers where the counting method is described, which is the repair this entry asks for; ranges given as ranges.\n\nDecision. FLAG when two or more precise figures have none of the four supports. WEAK on one. PASS when numbers are rounded, sourced or owned. Precision without support is the trigger. The presence of numbers is not.\n\nOutput: each figure, its four marks, and the verdict. Where a figure is unsupported, the correct next action is to ask the writer for the source, not to conclude anything about how the text was produced."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Our own corpus only. During the research pass for this index, widely repeated platform-performance figures were traced back and every one of them ended at a vendor or marketing post rather than at a first-party document. No external taxonomy names the pattern, which is why the grade is feedsquad-observed."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"Real numbers get lazy citation all the time. A writer who read a study last year and remembers the figure but not the reference produces a sentence indistinguishable from an invented one, and so does a journalist working from a briefing under embargo, an analyst quoting an internal dashboard they cannot link, and anyone summarising their own unpublished work. Trade publications routinely strip citations for readability. The first move on an unsupported figure is therefore to ask where it came from, because roughly half the time there is an answer. A number that survives one round of asking is fine. A number whose trail ends in another article that also does not say is the pattern this entry describes.","model_attribution":"None available. The figures we traced circulate through marketing and vendor posts, and where a model produced one first is not knowable from the artifacts we hold.","platform_notes":[],"languages":["en"],"sources":[{"kind":"feedsquad-observed","title":"AI Tells Index folklore register","observed":"2026-08-14","corpus":"Widely circulated claims about platform ranking penalties, reach effects and detection accuracy, traced back to primary sources during the research pass for this index. The fabricated platform-update figures are the type specimen: each resolves to vendor and marketing posts citing one another, and none appears in any first-party platform document."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"laundered-statistic","name":"Laundered statistic","aka":["benchmark stripped of its scope","headline drift"],"category":"semantic","subcategory":"statistics","description":"A real number from real work, republished without the scope that made it true. The study measured one population with one instrument over one period. The sentence quoting it names none of that, so the number arrives larger than it was.","why_it_reads_ai":"Summarisation drops qualifiers, and every hop drops a few more, so a benchmark result becomes a general fact in about three republications. The 2024 paper introducing generative engine optimization reports visibility gains of up to 40 percent for its own methods on the benchmark its authors built, measured against the engines of 2023 and 2024. It circulates without any of that. A 2025 measurement of new English-language articles crossing the halfway mark for AI generation circulates as a claim about the internet, and its own detector reports a 4.2 percent false-positive rate and a 0.6 percent false-negative rate and did not evaluate text where a human and a tool worked together. Those clauses have to travel with the number.","examples":[{"before":"Optimizing for generative engines raises visibility by 40 percent, and more than half the internet is now AI-generated.","after":"A 2024 paper reports visibility gains of up to 40 percent for its own methods, measured on the benchmark its authors built against the search engines of 2023 and 2024. Separately, a 2025 sample of new English-language articles put the AI share past half; the detector behind that number reports a 4.2 percent false-positive rate and did not evaluate work where a person and a tool collaborated. Both are worth citing. Neither is a fact about the internet.","note":"The figures here are the published ones and the scope clauses are what the sources actually state. This entry is the one this index is most likely to breach itself, which is why its own numbers were checked against the sources before it shipped."}],"detection":{"type":"judge","rubric":"Task: compare each imported statistic with the scope of the work it came from.\n\nStep 1. For every statistic attributed to a study, report or benchmark, find the original.\nStep 2. Record the original's population, sample size, date range, instrument and stated limits, including any error rate.\nStep 3. Record the population the citing sentence implies.\nStep 4. Compare, and mark each of these separately: population widened; date dropped; benchmark result stated as a general effect; error rate dropped from a detector-derived figure; a subset presented as the whole; a maximum presented as a typical value.\n\nEscape hatches: the sentence carries the scope; the original itself claims generality and defends it; the figure is a rounded version of a widely replicated result; the citing text is a headline with the scope in the body immediately below.\n\nDecision. FLAG on any of the six marks in Step 4. Report the original scope and the published scope side by side so the reader can see the distance.\n\nOutput: original scope, published scope, marks, verdict."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Two worked examples with reachable primaries: a benchmark result circulating as a general effect, and a measurement of new articles circulating as a measurement of the web. Both originals state their scope plainly, which is what makes the drift visible."}],"evidence_grade":"corroborated","false_positive_notes":"Good-faith summarisation oversimplifies, and it has to: a sentence cannot carry a methods section. Subeditors cut qualifiers for length, headlines are written by people who did not write the piece, and press offices publish the widest defensible version of their own findings. Researchers themselves generalise in interviews in ways their papers do not. The pattern this entry describes is systematic scope-stripping across a body of work rather than one loose sentence, and the repair is usually a clause rather than a retraction. Anyone applying it should apply it here first, since a directory that publishes numbers is the easiest place in the world to commit this.","model_attribution":"No vendor can be named. Scope-stripping is a property of chains of republication, and the documented examples travelled through human marketing writing.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"GEO, generative engine optimization benchmark (arXiv:2311.09735)","url":"https://arxiv.org/abs/2311.09735","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Graphite, more articles are now created by AI than humans","url":"https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"confident-misattribution","name":"Confident misattribution","aka":["plausible false credit","attribution that propagates unchecked"],"category":"semantic","subcategory":"attribution","description":"A credit that sounds right and travels because nobody opens the primary. The named person is real, the quote or the coinage is nearly theirs, and the correction is one click away and never made.","why_it_reads_ai":"Attribution is a slot, and the most probable filler for it is whoever is most associated with the topic, which is often not whoever did the thing. The worked example sits inside this index's own subject: it is widely claimed that a particular developer coined the term slop for unwanted machine output. His own widely cited post from May 2024 credits an earlier user and says the term was already in circulation. The post is short, it is public, and the claim it contradicts is still repeated in marketing copy.","examples":[{"before":"The term was coined in 2024 by the developer who wrote the definitive post about it. The name stuck because it captured something the industry had been circling for a while without having a word for it. That is usually how vocabulary works: someone says the obvious thing at the right moment and everyone else recognises it immediately.","after":"The developer who wrote the widely cited 2024 post says in that post that he did not coin the term, and credits an earlier user. No dated first use has been produced by anyone, so this piece does not give one.","note":"The developer in the specimen is deliberately unnamed. The two sentences after the claim explain why the credit felt right and check nothing, which is how a misattribution travels. The repair is not gentler wording. It is what the primary says, and the primary took a minute to read."}],"detection":{"type":"judge","rubric":"Task: check attributions against their primaries.\n\nStep 1. List every attribution of a coinage, quotation, finding, invention, prediction or position to a named person or body.\nStep 2. Open the primary. Not a summary of it, not a news write-up: the thing itself.\nStep 3. Ask three questions. Did this person say it? In the place cited? In this sense?\nStep 4. Where the attribution is a coinage, check specifically whether the named person claims it themselves. Popularisers are routinely credited with coinage they disclaim.\n\nEscape hatches: contested attributions where sources genuinely disagree and the text acknowledges the dispute; attributions inside a quotation, which belong to the speaker; claims of popularisation rather than origination, which are a different claim; paraphrase marked as paraphrase; primaries that cannot be reached, which return UNKNOWN.\n\nDecision. FLAG when the primary is reachable and contradicts the attribution. WEAK when no primary is cited at all. PASS when the primary supports it or the dispute is acknowledged.\n\nOutput: the attribution, what the primary says in one sentence, and the verdict."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"One reachable primary that contradicts a widely repeated claim about its own author. Single source, so the grade is community-observed, and the worked example is chosen because a reader can check it in a minute."}],"evidence_grade":"community-observed","false_positive_notes":"Attribution is contested more often than people assume. Independent invention happens, quotations migrate to more famous mouths over a century, and scholarly priority disputes run for decades without resolution. Popularisation is also a real contribution, and crediting the person who made an idea travel is defensible so long as the sentence says that is what it means. Oral traditions and collaborative work resist single attribution altogether. The check is not whether the credit is contested but whether a reachable primary contradicts it, and where the primary cannot be reached the honest verdict is that the claim is unverified rather than wrong.","model_attribution":"No vendor can be named. The misattribution catalogued here spread through human marketing writing, and the pattern predates any assistant by a long way.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Simon Willison, slop is the new name for unwanted AI content","url":"https://simonwillison.net/2024/May/8/slop/","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"stale-fact-as-current","name":"Stale fact stated as current","aka":["latest version that is not","current as of an unstated date"],"category":"semantic","subcategory":"temporality","description":"A fact that was true inside some earlier window, given in the present tense with no date on it. The price, the version, the job title, the policy. Nothing in the sentence tells a reader which year it belongs to.","why_it_reads_ai":"Generated text has a window behind it and no clock in front of it, so a state of the world from that window arrives in the present tense. Wikipedia treats undated present-tense claims as a cleanup signal alongside the explicit training-window boilerplate, and its speedy-deletion criterion covers that boilerplate itself. This entry is about the undisclosed case, where nothing announces the staleness. The boundary matters: published boilerplate announcing a training cutoff belongs to the model-register entry that catches the disclaimer text.","examples":[{"before":"The free tier includes 500 build minutes a month, and the latest release is version 3. Pricing for tools in this category tends to stay put, so these numbers should hold for most readers. If something has moved since, the vendor pages are always the better guide.","after":"As of 15 August 2026 the free tier includes 500 build minutes a month, per the pricing page linked here. Version 5 shipped in June. The version 3 documentation is still online, which is why it tends to come up first in search.","note":"The product details in both versions are invented for this specimen. The tell needs its numbers, so the emptiness is carried by the two sentences that reassure the reader the numbers are safe. The repair adds a date, a source and the reason the old fact is still circulating."}],"detection":{"type":"judge","rubric":"Task: find present-tense claims about things that change, with no date attached.\n\nStep 1. List every claim about a changing state: prices, tiers, limits, version numbers, headcount, leadership, ownership, policy, availability, market position, legal status.\nStep 2. For each, check whether the text attaches a date, an as-of clause, or a dated source.\nStep 3. For undated claims, check the current primary: the vendor's own page, the register, the release notes.\nStep 4. Record: undated and still true; undated and now false; dated.\n\nEscape hatches: historical writing in the past tense; documentation versioned by its own URL, where the version is the context; claims about stable facts that do not change; pages carrying a visible last-updated date that covers the claim; archived material presented as archived.\n\nDecision. FLAG when a changing-state claim is undated and the current primary disagrees. WEAK when it is undated and could not be checked. PASS when dated.\n\nOutput: the claim, the current primary's value, and the verdict. This rubric covers undisclosed staleness only. Text that announces its own training cutoff is a different pattern and belongs to that entry."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Wikipedia names undated present-tense claims and training-window behaviour among its cleanup signals, and its deletion criterion covers the explicit boilerplate. Two community documents, one behaviour, with the boundary between disclosed and undisclosed staleness stated in the entry."}],"evidence_grade":"community-observed","false_positive_notes":"Documentation goes stale between releases without anyone lying, and print publishing has carried this lag since it existed: a book about a product is out of date before it ships and everyone involved knows it. Writers also work from press material prepared weeks earlier, and a piece filed in March about a policy that changed in April was accurate when written. Translations lag their originals. Archived material is supposed to be stale, and stripping the archive banner is the reader's error rather than the writer's. The finding is a maintenance job, and the polite version of it is a note to the author rather than a conclusion about how the text was made.","model_attribution":"Follows from a fixed training window, which every model has, and from retrieval that returns an old page. The explicit boilerplate forms in the wild are dominated by older systems, per Wikipedia's dated examples.","platform_notes":[{"platform":"wikipedia","note":"Undated present-tense claims and training-window boilerplate are both listed as cleanup signals, and the boilerplate appears in the speedy-deletion criterion. Editors distinguish the announced case from the silent one, and so does this entry."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"echo-restatement-reply","name":"Echo restatement reply","aka":["comment that restates the post","agreement echo","bulk comment","reply-guy paraphrase","quote-post paraphrase plus lesson","restatement comment"],"category":"platform","subcategory":"reply-shape","description":"A reply that paraphrases the post it answers and stops. It opens with agreement, restates the argument in the parent's own vocabulary, then asks a question the parent already answered. LinkedIn lists this as one of three named enforcement targets. Metric: the share of the reply's content words that also occur in the parent post, after stopword removal and lemmatisation. The 0.5 threshold is an operating point rather than a measured cutoff, because no published corpus supplies one; the reasoning is that a reply which shares more than half its content words with the post has at minimum restated it, while a reply that quotes one phrase and then argues keeps its own vocabulary and lands lower. Score only replies of 25 content words or more, and pass any reply that introduces a proper noun, a figure or a contradiction regardless of overlap. Calibrate on your own reply corpus before enforcing. A reply that carries nothing from the parent at all belongs to generic-affirmation-comment; this entry is for the reply that hands the parent its own words back.","why_it_reads_ai":"Comment automation has the post text as its only input, so overlap with the parent is high by construction. A person who read the post and had a reaction brings vocabulary the post did not contain. The signal is the overlap plus the absence of anything new, and either half alone is worthless.","examples":[{"before":"Absolutely agree. Consistency really is the foundation of a good content strategy, and posting regularly does build trust with your audience over time. Trust is what turns followers into customers. What cadence would you recommend for a smaller team?","after":"This held for us until it stopped. We posted three times a week for a year, then cut to one long piece a month, and inbound stayed flat while the writing time dropped by about two thirds. Cadence bought us nothing after the first quarter. What changed things was having something specific to say.","note":"The repair contradicts the parent using the writer's own evidence. Agreement is not the problem. Agreement in the parent's words with nothing added is."},{"before":"Great points here! Onboarding really is critical for reducing churn, and investing in the early user experience does pay off across the customer lifecycle. Thanks for sharing these insights.","after":"The early-experience point matches ours, but the fix was smaller than expected. Cutting our signup form from nine fields to three moved first-month churn from 31 to 19 percent. Everything we did later inside the product moved it less than that one form did.","note":"Overlap with the parent drops because the reply is carrying its own facts. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"parent-post-content-word-overlap-ratio","threshold":0.5,"direction":"above","threshold_basis":"No published measurement of reply-to-parent overlap exists for either human or model replies. Half the content words shared with the parent post is a FeedSquad review trigger set against our own reply corpus, chosen because a reply above it is restating the parent rather than adding to it. The overlap ratio is the observation; the emptiness after the restatement is the actual tell."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-14","status":"active","rationale":"The pattern is named verbatim in LinkedIn's own policy and is enforced on X through the reply-spam path. The threshold is provisional and flagged as such in the description; it is the weakest part of this entry and should be recalibrated once a reply corpus exists."}],"evidence_grade":"primary-doc","false_positive_notes":"Support agents and community moderators restate a question before answering it, because a reply that quotes the ask reads as attentive and survives being read out of thread context. Teachers do the same when answering in public. The tell is a restatement that adds nothing after it, not the restatement itself.","model_attribution":"No family attribution. This is a property of comment automation rather than of any model, and no platform discloses which tools or models produce the replies it labels.","platform_notes":[{"platform":"linkedin","note":"LinkedIn names responses that simply restate the original post without adding anything new, alongside comments created at scale by automation with minimal human involvement. Flagged content stays visible to connections and loses out-of-network distribution. LinkedIn reported 94 percent accuracy at identifying generic content in initial testing and published no false-positive rate, so treat that number as the company's claim."},{"platform":"x","note":"X's published enforcement rules attach a RiskyHighVizReply label for 30 days when the llm_slop_post label is present, and the account-level llm_slop_user label attaches SpamHighRecall, which routes to a timeline drop rule. Credibility prechecks skip high-follower and high-PageRank accounts before slop enforcement is evaluated, so the pattern is enforced hardest against small accounts. The classifier prompts themselves are withheld from the public repo."}],"languages":["en"],"sources":[{"kind":"external","title":"Keeping conversations real on LinkedIn (Laura Lorenzetti, VP and Executive Editor, LinkedIn Global Editorial, 2026-05-20)","url":"https://www.linkedin.com/pulse/keeping-conversations-real-linkedin-laura-lorenzetti-9821e","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Hari Srinivasan, Chief Product Officer, LinkedIn: AI slop report control announcement (2026-07-30)","url":"https://www.linkedin.com/posts/hsrinivasan1_ai-slop-is-a-top-priority-for-all-of-us-share-7488612006321889282-Ps8Z/","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"xai-org/x-algorithm: enforcement_post.yaml, enforcement_user.yaml, grox/flows/reply_spam (llm_slop_post and llm_slop_user rules)","url":"https://github.com/xai-org/x-algorithm","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"LinkedIn newsroom, authentic content and conversations (Mar 2026)","url":"https://news.linkedin.com/2026/authentic-content-and-conversations","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"eric-sabe/slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-14","updated":"2026-08-15"},{"id":"broetry-cadence","name":"Broetry cadence","aka":["one sentence per line","white-space stack post"],"category":"platform","subcategory":"post-shape","description":"One sentence per line, blank line between, held for a whole post. BuzzFeed News named the format broetry in December 2017 and traced its conventions to a marketer teaching them as a growth formula: a personal story, no outbound links, a tag at the end. Line breaks are a layout decision, so the shape carries no information about who typed the words. What it does say is that a formula was in the room before the content was.","why_it_reads_ai":"Ask an assistant for a LinkedIn post and this layout arrives before any material does, because the corpus it learned from is saturated with the format. The lines are then often short enough to hold nothing. Read the lines instead of the spacing. A stacked post carrying dates, figures and named people is just a stacked post.","examples":[{"before":"I almost turned down the meeting.\n\nMy calendar was already full that week.\n\nSomething told me to take it anyway.\n\nThat call changed how I think about pipeline.\n\nLesson: take the meeting.","after":"I nearly skipped a 30 minute call with a systems integrator in March 2025 because the week was full. They became our second largest reseller by December and now account for about a fifth of new seats. Worth taking because their customers already ran the platform we plug into, which I did not know before I sat down.","note":"The repair keeps the anecdote and adds the things that make it checkable: a date, a role, a size. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"(?:^[A-Za-z][^\\n.!?]{5,110}[.!?]\\r?\\n\\r?\\n){3}","flags":"gm","scope":"document"},"severity":"medium","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"The format is a documented human invention from 2017, taught in paid courses five years before an assistant could imitate it, and no LinkedIn policy has ever addressed line breaks. Contested is the honest shipping status: the shape is real and worth naming, its signal value is close to zero."}],"evidence_grade":"community-observed","false_positive_notes":"Human growth marketers on LinkedIn built this format in 2017 and sold it as a course, so a stacked post is first evidence that somebody took the course. Poets break lines this way. So do people typing on phones, and anyone writing around the truncated preview the feed shows before a reader taps to expand. The reviewer check is whether any single line carries a fact that could turn out to be wrong.","model_attribution":"No family attribution. Assistants reproduce the layout on request for a LinkedIn post because the training corpus is full of it, and no vendor documents the tendency in a model card or system prompt.","platform_notes":[{"platform":"linkedin","note":"LinkedIn has published nothing about line breaks. Its May 2022 feed post addresses explicit reaction requests and poll volume, and its March 2026 newsroom post defines automated comments and engagement pods. Neither names a writing style. Claims that LinkedIn demotes particular formatting circulate widely in marketing blogs and trace back to no LinkedIn document."}],"languages":["en"],"sources":[{"kind":"external","title":"BuzzFeed News, why are these posts taking over your LinkedIn feed","url":"https://www.buzzfeednews.com/article/ryanmac/why-are-these-posts-taking-over-your-linkedin-feed-because","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"terminal-question-bait","name":"Terminal question bait","aka":["Agree?","Thoughts?","Or am I the only one?","Is it just me?"],"category":"platform","subcategory":"closer","description":"A closing question built for comment volume rather than for an answer. The one-word variants ask for assent. The solidarity variants ask a reader to confirm a grievance the post already stated. Both were folded into this one entry because they are the same closer with a different pronoun, and splitting them would have shipped two rules that fire on the same line.","why_it_reads_ai":"The shape is what a post reaches for when it has run out of material. An assistant asked to end a post with engagement produces it because the request has one obvious answer. A writer who actually wants input usually asks something narrower, because a wide question returns nothing usable.","examples":[{"before":"Most teams treat onboarding as a checklist and then wonder why activation is flat. Fix the first week and everything downstream gets easier. Thoughts?","after":"Most teams treat onboarding as a checklist and then wonder why activation is flat. We cut our first week from eleven steps to four in February and week-two activation moved from 22 to 34 percent. If your first week is longer than four steps, I would like to hear what the extra ones are doing.","note":"The repair still invites a reply, and the invitation is narrow enough to answer. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"(?:^|[.!?…][\"')’”]?[ \\t]+)(?:agree\\?|thoughts\\?|your thoughts\\?|am i wrong\\?|(?:or )?am i the only one(?: who [^?\\n]{1,50})?\\?|is it just me(?:,| or)?(?: [^?\\n]{1,50})?\\?|who(?:'|’)?s with me\\?)[ \\t]*$","flags":"im","scope":"sentence"},"severity":"medium","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"LinkedIn names posts that expressly ask the community to engage via likes or reactions, and Mosseri acknowledged rising engagement bait on Threads. Neither names a soft closing question, so the entry ships contested: platforms enforce the explicit ask, and this is the softer cousin that no policy reaches."}],"evidence_grade":"primary-doc","false_positive_notes":"Writers who genuinely want input end on a question, and community managers are trained to. Teachers close with one on purpose. LinkedIn's own policy language reaches explicit reaction requests, not soft questions, so this entry documents a shape the platform never named. The reviewer check is whether the post would still stand with the question deleted: if the body already made a claim worth arguing with, the closer is decoration rather than the point.","model_attribution":"No family attribution. The closer is a property of engagement-shaped writing and appears whenever a prompt asks for a post that drives comments, regardless of which assistant answers.","platform_notes":[{"platform":"linkedin","note":"The May 2022 feed post says LinkedIn will not promote posts that expressly ask or encourage the community to engage via likes or reactions when the intent is boosting reach. That is narrower than a closing question and this entry does not claim otherwise."},{"platform":"threads","note":"Adam Mosseri posted in October 2024 that engagement bait on Threads had increased and that the team was working to get it under control. He did not define the term or name any specific closer."}],"languages":["en"],"sources":[{"kind":"external","title":"LinkedIn, keeping your feed relevant and productive","url":"https://www.linkedin.com/blog/member/product/keeping-your-feed-relevant-and-productive","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Mosseri on Threads, engagement bait","url":"https://www.threads.com/@mosseri/post/DA01Sd8vtm6","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"explicit-engagement-request","name":"Explicit engagement request","aka":["Like if you agree","tag a colleague","share this post","drop an emoji below"],"category":"platform","subcategory":"engagement-bait","description":"A direct ask for a reaction, a tag, a share or a save. This is the one pattern in the whole directory that platforms define in writing and act on. Meta has demoted it since December 2017 under the name engagement bait, LinkedIn declines to promote posts whose intent is boosting reach through reaction asks, and YouTube names engagement manipulation and fake engagement in its spam policies. Reaction bait, tag bait, share bait and save bait are one policy category and are covered here rather than split across four entries.","why_it_reads_ai":"The ask is what a post substitutes for a reason to respond. Generation makes the substitution cheap, so it appears at the end of drafts that have nothing else to close with. The enforcement is authorship-indifferent: platforms demote the ask whoever typed it, which is exactly why this entry is high severity and none of the style entries are.","examples":[{"before":"Rough quarter, real lessons. We shipped slower than planned and learned more than expected. Save this post for later and tag a founder who needs it.","after":"Rough quarter, real lessons. We missed the March release by five weeks because two engineers were pulled onto a migration nobody had scoped. The fix was a written capacity check before any date leaves the room, and it has held for two releases since.","note":"Two hollow beats and then the ask, which is where a post goes when it has nothing to close with. Figures in this repair are invented for the specimen."},{"before":"Cold email still works if you get the basics right. Personalise the first line and keep the whole thing under 120 words. Tag a colleague who needs to read this, and comment SEND below and I will send you the template.","after":"Cold email still works if you get the basics right. Our reply rate went from 3 to 9 percent after we cut the average message from 140 words to 70 and deleted the opening compliment. The template is linked in my profile and nobody has to comment to get it.","note":"The repair keeps the offer and removes the toll gate. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:(?:like|repost|share|retweet|comment)[ \\t]+(?:this[ \\t]+|it[ \\t]+)?if[ \\t]+you\\b|tag[ \\t]+(?:a|an|your|two|three|2|3)[ \\t]+(?:colleague|colleagues|friend|friends|founder|founders|marketer|marketers|teammate|teammates|person|people|someone)\\b|save[ \\t]+this[ \\t]+(?:post[ \\t]+)?for[ \\t]+later\\b|drop[ \\t]+an?[ \\t]+[^\\s.,!?\\n]{1,15}[ \\t]+(?:below|in[ \\t]+the[ \\t]+comments)\\b|comment[ \\t]+[\"“]?[A-Za-z]{1,12}[\"”]?[ \\t]+below\\b|double[ \\t-]tap[ \\t]+if\\b|follow[ \\t]+for[ \\t]+more\\b)","flags":"gi","scope":"document"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Three platforms define this in first-party policy text and state the consequence, which is demotion or lost promotion rather than removal. Meta has run the policy since December 2017. Nothing in the evidence suggests the behaviour has stopped, so the entry ships active at high severity."}],"evidence_grade":"primary-doc","false_positive_notes":"Meta's own guideline carves out the cases that look identical and are not: appeals to find missing people or property, fundraisers, petitions, and posts during natural disasters or life-threatening events. Beyond the carve-outs, creators legitimately prompt saves on reference material because saving is how a reader files it, and charities ask for shares because sharing is the ask. The reviewer check is whether the requested action does anything for the reader once taken.","model_attribution":"No family attribution. This is marketer behaviour that models inherited from the corpus, and Meta named it as a human tactic five years before any assistant could write a post.","platform_notes":[{"platform":"meta","note":"Meta transparency guidance defines engagement bait as posts that explicitly request votes, shares, comments, tags, likes or other reactions, and reduces their distribution. The December 2017 announcement gives one concrete example only. The five named sub-types that secondary sources routinely attribute to that announcement are not in it, so this entry does not use them."},{"platform":"linkedin","note":"LinkedIn says it will not promote posts that expressly ask or encourage the community to engage via likes or reactions where the intent is boosting reach, and its community policies ban artificially increasing engagement. The company frames both as behaviour rather than as a writing style."},{"platform":"youtube","note":"YouTube spam policy names engagement manipulation, defined as repetitive or templated content aimed at artificially inflating engagement, and fake engagement, which covers exploiting features such as polls to force interaction."}],"languages":["en"],"sources":[{"kind":"external","title":"Meta transparency centre, engagement bait","url":"https://transparency.meta.com/features/approach-to-ranking/content-distribution-guidelines/engagement-bait/","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Facebook, fighting engagement bait on Facebook (Dec 2017)","url":"https://about.fb.com/news/2017/12/news-feed-fyi-fighting-engagement-bait-on-facebook/","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"LinkedIn, keeping your feed relevant and productive","url":"https://www.linkedin.com/blog/member/product/keeping-your-feed-relevant-and-productive","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"generic-affirmation-comment","name":"Generic affirmation comment","aka":["Great insights!","So true","abstract praise attachable to any post"],"category":"platform","subcategory":"reply-shape","description":"Abstract praise that would sit equally well under any post, produced at volume. The boundary with echo-restatement-reply is the parent: that entry covers replies built out of the post's own words, and this one covers replies that carry nothing from the post at all. Measure it at the account level. Count the share of an account's comments in a window that contain no proper noun, no numeral, no quotation from the parent and no disagreement. One short comment carries no information; sixty of them in a row does.","why_it_reads_ai":"A comment that could attach anywhere was written without reading. Comment automation produces it because praise is the safest output when the input was skimmed. The volume is the evidence, and LinkedIn treats the volume rather than the wording as the thing worth acting on.","examples":[{"before":"Great insights! So true. Thanks for sharing this one.","after":"The part about pausing paid spend for a week is where I would push back. We tried it in January and organic did not absorb the demand, it just went missing for eleven days. Worth testing on one channel before the whole budget.","note":"The repair is still short. Brevity was never the tell. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"contentless-comment-share-per-account","threshold":0.6,"direction":"above","threshold_basis":"No published measurement of comment specificity exists for human or automated commenters, so this number is not a validated cutoff. Six in ten is a FeedSquad review trigger, set at the account level because a single empty comment is meaningless and a run of them is not. Score only accounts with at least 20 comments in the window. Count a comment as carrying content if it contains a proper noun, a numeral, a quotation from the parent post or a stated disagreement."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"LinkedIn defines automated comments in a first-party newsroom post and describes the enforcement it applies, and the public X algorithm repo carries slop labels enforced through the reply-spam path. Both target volume rather than wording, which is how this entry is scored."}],"evidence_grade":"primary-doc","false_positive_notes":"Polite people write short comments constantly, and brevity is not automation. Readers on phones, non-native writers who do not want to risk a longer sentence in public, and people acknowledging a bereavement or a job loss all write four words and mean them. Only repetition across many unrelated posts by one account carries information, and even then the honest reading is that the account is farming attention rather than that a machine typed it.","model_attribution":"No family attribution. Platforms label the behaviour without disclosing which tools or models produce the comments, and the classifiers that assign the labels are not published.","platform_notes":[{"platform":"linkedin","note":"The March 2026 newsroom post defines an automated comment as one posted using a browser extension, script or third party tool, and says such comments flood comment sections and displace authentic ones. Named enforcement: exclusion from Most Relevant comments, sometimes no display outside the commenter's network, and account restrictions. The post never mentions AI-generated content."},{"platform":"x","note":"The published algorithm repo attaches a 30-day reply-visibility label when the post-level slop label is present, and an account-level spam label that routes to a drop rule. Credibility prechecks skip high-follower and high-PageRank accounts before slop enforcement evaluates, so the rules land hardest on small accounts. The classifier prompts are withheld from the repo."}],"languages":["en"],"sources":[{"kind":"external","title":"LinkedIn newsroom, authentic content and conversations (Mar 2026)","url":"https://news.linkedin.com/2026/authentic-content-and-conversations","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"xai-org/x-algorithm, llm_slop_post and llm_slop_user enforcement rules","url":"https://github.com/xai-org/x-algorithm","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"humblebrag-parable-post","name":"Humblebrag parable post","aka":["anecdote, reversal, lesson","workplace parable post"],"category":"platform","subcategory":"post-shape","description":"A first-person workplace scene, a reversal where an assumption is overturned, and a lesson bolted on at the end. The three beats arrive in that order every time. The story is usually flattering to the teller in a way the lesson pretends not to notice, which is where the name comes from.","why_it_reads_ai":"The shape is a container that runs fine while empty. Generation fills it with a scene nobody can check: no company, no date, no job title, no number. A real workplace story leaks specifics whether the writer wants it to or not, because that is what remembering looks like.","examples":[{"before":"A junior colleague asked me a question I could not answer. I felt the room go quiet. I told her I did not know and would find out. By Friday she had found the answer herself. Leadership has very little to do with having answers.","after":"A junior engineer asked why our webhook retries were capped at three attempts and I had no idea. The cap turned out to have been set by a contractor in 2023 to work around a rate limit we removed a year later. We raised it to eight and failed deliveries fell from roughly 40 a week to under 5. I still have no leadership lesson from this.","note":"The repair keeps the reversal and drops the moral. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Scan the post for three beats in this order: a first-person workplace scene, a reversal where a stated assumption is overturned, and a general lesson appended after the story ends. Return no-tell unless all three are present. Then scan the whole post for one checkable detail: a date, a place, a job title, a figure, a named product or a named tool. If any appears, return no-tell, because the writer is describing something that happened. Return tell only when all three beats are present and no checkable detail appears anywhere, and quote the lesson sentence back as the evidence."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"No external source documents this shape, so it ships on our own corpus and is graded accordingly. It is low severity because the human origin is strong: the broetry tradition invented the anecdote-reversal-lesson post and still teaches it."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"True stories have reversals in them, and the people most likely to write one up are managers and teachers whose job is turning an incident into something transferable. The 2017 LinkedIn growth-marketing tradition invented this exact shape and still sells it, so the format is evidence of a taught formula and nothing more. The reviewer check is the checkable detail, not the beats: a story with a date and a number in it is a story.","model_attribution":"No family attribution. The shape is what any assistant produces when a prompt asks for a LinkedIn post with a personal story, and no vendor documents it.","platform_notes":[],"languages":["en"],"sources":[{"kind":"feedsquad-observed","title":"FeedSquad editorial queue, LinkedIn drafts held for human approval","observed":"2026-08-15","corpus":"Posts drafted by FeedSquad writing agents for the LinkedIn surface and reviewed before publishing, June to August 2026. No counts were kept, so this is a shape seen repeatedly in review rather than a measurement, and it is graded as such."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"authenticity-vocabulary-tic","name":"Authenticity vocabulary tic","aka":["authenticity and vulnerability in every post","community language without a community"],"category":"platform","subcategory":"vocabulary","description":"The pairing of authenticity with vulnerability, in post after post, by an account that demonstrates neither. The tell is the pairing and the frequency, never either word on its own. Both words have ordinary uses and one of them is a clinical term.","why_it_reads_ai":"The vocabulary is what a post about connection reaches for when it has no incident to describe. An assistant asked for a thought-leadership post on LinkedIn produces the pair because the corpus pairs them. The reader-side version of this observation shows up in public comments on LinkedIn's own posts about automated engagement, where members list the pairing alongside instant reaction bursts as what they use to spot a farmed account.","examples":[{"before":"Leadership in 2026 comes down to authenticity and vulnerability. The teams that win are the ones showing up authentically every day, and the leaders who win are the ones who let people see them.","after":"I told my team in April that we had eleven weeks of runway left and that I did not know whether the Series A would close. Two people started interviewing elsewhere. One of them stayed and now runs support. I would do it again, and I would do it four weeks earlier.","note":"The repair replaces the vocabulary with the thing the vocabulary was standing in for. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\bauthenticit(?:y|ies)\\b[^.!?\\n]{0,60}\\bvulnerabilit(?:y|ies)\\b|\\bvulnerabilit(?:y|ies)\\b[^.!?\\n]{0,60}\\bauthenticit(?:y|ies)\\b|\\bbring(?:ing)? your whole self\\b|\\bshow(?:ing)? up authentically\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"The observation is reader-side folk knowledge with two first-party LinkedIn documents attached, which is why it ships active at low severity. It is a vocabulary tic and vocabulary tics decay fastest, so expect this one to weaken."}],"evidence_grade":"primary-doc","false_positive_notes":"Coaches, community managers, therapists and people who run peer support groups write about these things constantly because they are the subject of the work, not a garnish on it. Recovery writing pairs the two words by necessity. The reviewer check is whether the post contains an instance: one moment where the writer was actually exposed, with a consequence attached. Vocabulary plus an instance is a subject. Vocabulary plus nothing is a costume.","model_attribution":"No family attribution. No vendor documents suppressing this pairing, and it appears across assistants asked for LinkedIn thought leadership.","platform_notes":[{"platform":"linkedin","note":"The vocabulary observation comes from public commenters on a LinkedIn employee's post about coordinated engagement, not from LinkedIn policy. That post, by Oscar Rodriguez in early 2026, confirms group removals and warnings sent to thousands of members. The March 2026 newsroom post defines automated comments and engagement pods and says nothing about wording. No LinkedIn document names any vocabulary at all."}],"languages":["en"],"sources":[{"kind":"external","title":"Oscar Rodriguez, LinkedIn post on authenticity signals","url":"https://www.linkedin.com/posts/orodriguez_earlier-this-fall-i-shared-an-article-about-share-7407561708262703105-l99B","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"LinkedIn newsroom, authentic content and conversations (Mar 2026)","url":"https://news.linkedin.com/2026/authentic-content-and-conversations","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"unenhanced-repost","name":"Unenhanced repost","aka":["aggregator caption","repost farming","reposting without adding anything"],"category":"platform","subcategory":"reuse","description":"Someone else's material, republished with nothing meaningful added. Meta's July 2025 policy post states that accounts repeatedly reusing other people's videos, photos or text posts lose access to monetisation for a period and get reduced distribution on everything they share, and says a watermark or a stitched compilation does not count as enhancement. YouTube's reused-content policy asks for original commentary, substantive modification, or educational or entertainment value. Instagram extended its anti-aggregator limits from Reels to photos and carousels in April 2026. This entry merges the plain repost and the scraped repost, which are one policy family with one trigger. None of the cited policies mentions AI.","why_it_reads_ai":"Generation lowered the cost of the caption, which was the only part of a repost that ever took work. What remains is a distribution account with a sentence on top. The enforcement is about the reuse, not about who wrote the sentence, which is why this entry judges the addition rather than the prose.","examples":[{"before":"Reposting this because it deserves more eyes. Such a great breakdown of retention. Everyone building product should read it.","after":"Reposting the retention breakdown below, with one disagreement. Their cohort chart drops anyone who churns inside seven days, which is where about two thirds of our own churn happens, so their curve flattens earlier than ours ever has. Run it with day-zero users included and the shape changes.","note":"Credit stays, the framing does work. Figures in this repair are invented for the specimen."},{"before":"Found this thread and had to share. So much value in here for anyone in B2B. Full credit to the original author.","after":"Sharing this thread because point four is wrong for anyone selling under 5,000 euros a year. Their playbook assumes a sales call, and at that price the call costs more than the contract. Points one to three hold and we use them.","note":"Credit is not enhancement. A claim about the material is."}],"detection":{"type":"judge","rubric":"Identify what portion of the item came from somewhere else: a reposted video, a screenshotted thread, a quoted article, a scraped list. Return no-tell if the borrowed portion is under a quarter of the whole. Otherwise look for enhancement, meaning original commentary that makes a claim about the borrowed material, a correction, added data, or an edit that changes what the material means. A watermark, a re-cut, a restated caption, a credit line and a stitched compilation are not enhancement. Return tell when the borrowed material dominates and nothing in the item makes a claim the source did not already make, naming the borrowed portion and the missing addition."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Three platforms carry written policy against this and state the consequence in money and reach. The severity is high because the enforcement is real and current, not because the writing is bad."}],"evidence_grade":"primary-doc","false_positive_notes":"Licensed syndication looks identical from outside and is a contract. Credited curation with real editorial framing is a genre with a long history, and the value in a good link newsletter is the selection, not the prose around it. Meta's policy is about improper reuse, which means without permission, so a reposter with a licence is not in scope at all. The reviewer check is two questions: does the account have the right to the material, and does the framing make a claim the source did not.","model_attribution":"No family attribution. The platform policies here predate the current model era in substance and never name a model family or an authorship test.","platform_notes":[{"platform":"meta","note":"The July 2025 post on unoriginal content and the March 2026 post on rewarding original creators together set out the enforcement: repeat improper reuse costs monetisation access for a period and reduces distribution across everything the account shares, and duplicate videos are demoted in favour of the original. Neither post mentions AI-generated content."},{"platform":"instagram","note":"Instagram extended its limits on aggregator accounts from Reels to photos and carousels in April 2026, reported by TechCrunch. Accounts posting mostly unoriginal material become ineligible for recommendations. The rolling window that circulates alongside this change was not confirmed against a Meta document, so no number appears here."},{"platform":"youtube","note":"The spam policy names scraped content, defined as re-posting material from other websites, platforms or videos without adding anything of your own. The separate reused-content policy asks for significant original commentary, substantive modification, or educational or entertainment value."}],"languages":["en"],"sources":[{"kind":"external","title":"Meta, combating unoriginal content (Jul 2025)","url":"https://creators.facebook.com/blog/combating-unoriginal-content","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Meta, rewarding original creators on Facebook (Mar 2026)","url":"https://about.fb.com/news/2026/03/rewarding-original-creators-on-facebook/","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"YouTube spam, deceptive practices and scams policy","url":"https://support.google.com/youtube/answer/2801973","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"TechCrunch, Instagram restricts reach of content aggregators","url":"https://techcrunch.com/2026/04/30/instagram-restricts-reach-of-content-aggregators-in-new-crackdown/","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"template-review-swapped-noun","name":"Template review with a swapped noun","aka":["same review, different product","cross-product text reuse"],"category":"platform","subcategory":"review-shape","description":"The same review text appearing across unrelated products with the product noun exchanged. Pangram, a detector vendor, reported on 4 May 2026 that about 3 percent of front-page reviews on 500 best-sellers were machine-written, 909 of 30,000 across ten categories, and that 93 percent of those carried a verified-purchase badge. The badge number is the one worth keeping: purchase verification does not separate a review a person wrote from one a model wrote. The vendor also describes those reviews clustering at the ends of the star scale, higher at five stars and higher at one star than human reviews; the figures behind that stay with the vendor and are not reprinted here.","why_it_reads_ai":"A template with a slot is the cheapest way to produce a review at volume, and the slot is almost always the product name. Everything around it stays fixed because nothing around it was ever about the product. Amazon's own detection works on behaviour and account graphs rather than prose, which is a standing reminder that text reuse is the amateur signal and the professional one is the account behind it.","examples":[{"before":"This blender does exactly what I needed and the build quality is solid. It arrived quickly and was easy to set up. I would recommend this blender to anyone looking for a reliable option at this price point.","after":"The 900W motor stalls on frozen strawberries unless you pour in liquid first, which the manual does not mention anywhere. Two months in, the gasket has taken on the smell of whatever went through it last and does not wash out. Fine for smoothies, wrong for nut butter.","note":"The repair is specific to one object, so it cannot be reused by swapping a noun. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"cross-product-review-text-reuse-ratio","threshold":0.7,"direction":"above","threshold_basis":"No published cutoff exists for review-to-review text reuse. Seventy percent shared word 5-grams between two reviews of unrelated products is a FeedSquad review trigger, set high enough that a prolific reviewer's personal formula, which typically reuses an opening and a sign-off, stays below it while a swapped-noun template lands above. Compare after lowercasing and removing the product name. The Pangram study measured how common machine-written Amazon reviews are, not how much text they share, so this number is ours and not theirs."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"One vendor measurement and one federal rule support the entry, which is corroboration across independent kinds of source rather than a single sighting. Severity is high because fake reviews carry money and regulatory consequences, not because the prose is dull."}],"evidence_grade":"corroborated","false_positive_notes":"Prolific honest reviewers develop a personal formula and reuse it, because writing 400 reviews without a shape is exhausting. Reviewers who buy several units of the same product legitimately write near-identical text about them. The discriminator is reuse across unrelated products, not the presence of a formula, and even then a shared account and a shared household explain a lot. This entry never supports a claim about who wrote a given review.","model_attribution":"No family attribution. The vendor that produced the prevalence figure sells detection and does not publish which model families it attributes text to for this dataset.","platform_notes":[{"platform":"amazon","note":"Amazon describes its own fake-review detection as behavioural and graph-based, looking at review history, risky behavioural patterns and groups of connected bad actors. Its published material does not name AI-generated reviews or any stylistic criterion at all, which is the strongest available argument that text-level review tells are a weak instrument."}],"languages":["en"],"sources":[{"kind":"external","title":"Pangram, AI-generated Amazon reviews (4 May 2026)","url":"https://www.pangram.com/blog/ai-amazon-reviews","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"FTC 16 CFR Part 465, rule on consumer reviews and testimonials","url":"https://www.ftc.gov/legal-library/browse/federal-register-notices/16-cfr-part-465-trade-regulation-rule-use-consumer-reviews-testimonials-final-rule","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"volume-signature","name":"Volume signature","aka":["step-change in publishing rate","output implausible for the staff named"],"category":"platform","subcategory":"aggregate","description":"A step change in publishing rate against a stable baseline. Three organisations found their problem this way before anyone read a sentence. Clarkesworld closed submissions in February 2023 after machine-written entries jumped from roughly 25 a month to more than 500. NewsGuard's tracker counts sites that churn out dozens of articles a day and had logged 3,749 of them as of 23 June 2026. curl ended its bug bounty at the end of January 2026 under a flood of well-formatted reports describing nothing real. Seven reports arrived in the programme's final week and none of them described a vulnerability.","why_it_reads_ai":"Volume is the property that rewriting cannot remove. A person writing carefully has a ceiling and generation does not, so the aggregate moves long before the prose does. This is the strongest evidence in the whole dataset that behaviour beats style, and it is also why the platforms with the most data score accounts rather than sentences.","examples":[{"before":"Our newsroom publishes 40 in-depth analyses every week, covering markets and policy with rigour you can trust. Our team of expert writers works around the clock so you never miss a development.","after":"We publish two analyses a week. Both are written by Sara Lindqvist and me, we name our sources inside the piece, and corrections go at the top of the article with the date they were made.","note":"The repair states an output a masthead of two can actually produce. Names and figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"publishing-rate-step-change-vs-baseline","threshold":3,"direction":"above","threshold_basis":"Nobody has published a cutoff for how fast a publisher may accelerate, so this is a FeedSquad review trigger rather than a measured line. Three times the trailing 12-week median is set where an ordinary campaign push or a new hire will usually stay below it. The events that produced this entry cleared it by far more: 25 submissions a month to over 500 at Clarkesworld, and sites publishing dozens of articles a day in the NewsGuard tracker. Treat a hit as a prompt to ask who wrote the extra output, and read it next to the masthead."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Three independent organisations in three fields detected their slop influx by rate rather than by reading, and each published what they had counted. That is the strongest corroboration available for any entry here, and it holds regardless of how the text was produced."}],"evidence_grade":"corroborated","false_positive_notes":"Publications scale up honestly all the time, through hiring, syndication deals, acquisitions and wire feeds. A seasonal desk triples its output every December on purpose. The check is whether the masthead grew with the output and whether the new material carries reporting that costs something: named sources, original photography, corrections. A rate change with a staffing change behind it is a business event, and this entry says nothing about it.","model_attribution":"No family attribution. None of the three organisations named a model, and two of them said explicitly that they were counting submissions rather than judging text.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"NPR, Clarkesworld closes submissions after a flood of machine-written stories","url":"https://www.npr.org/2023/02/24/1159286436/ai-chatbot-chatgpt-magazine-clarkesworld-artificial-intelligence","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"NewsGuard AI tracking center and UAIN framework","url":"https://www.newsguardtech.com/special-reports/ai-tracking-center/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"The Register, curl shutters bug bounty program to stop AI slop","url":"https://www.theregister.com/security/2026/01/21/curl_shutters_bug_bounty_program_to_stop_ai_slop/5063039","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"generic-outlet-identity","name":"Generic outlet identity","aka":["two nouns and a news word","no masthead, no named editor"],"category":"platform","subcategory":"provenance","description":"An outlet name assembled from two ordinary nouns and a news word, attached to a site with no masthead, no named editor and no address. NewsGuard lists interchangeable names of this kind among the things it looks for when logging unreliable AI-generated news sites, alongside chatbot error text left in the body copy and a publication rate no staff could sustain.","why_it_reads_ai":"A name that could belong to any outlet was chosen so that it would not have to belong to anyone. The absence of a reachable human is the load-bearing part, because a name is cheap to fix and an editor is not. This is a provenance judgment rather than a prose judgment, which is why it survives when every style tell decays.","examples":[{"before":"Business Post Daily is a leading source of business and technology news, delivering timely coverage to readers worldwide. Written by Staff. Our reporting is independent, accurate and built on the standards readers expect from a serious publication. Founded on the belief that good journalism should be available to everyone, we have grown into a trusted destination for professionals across many industries. We cover the stories that matter to people who need to stay informed, without the noise, and we intend to keep doing exactly that.","after":"Kerava Business Weekly is edited by Anni Laakso, who writes most of it, with two reporters listed on the staff page with working email addresses. Corrections run at the foot of the story with the date. The publisher and registered office are in the footer.","note":"The name and the staff byline are the tell. Everything after them is an about page that never reaches a person: no editor, no address, no way in. Names in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Open the about page and the masthead. Return no-tell if a named human editor or reporter is listed with a working way to reach them, however small the outlet. Otherwise check three things in order: whether the outlet name is two generic nouns plus a news word of the kind a name generator produces, whether article bylines are staff labels with no profile behind them, and whether the footer carries a postal address or a registered publisher. Return tell only when the name is generic and no named human is reachable anywhere on the site, quoting the byline and the outlet name."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"One organisation publishes the framework and the site counts, so the grade stays community-observed even though the criteria are written down. Severity is high because the entry is about provenance rather than taste, and a reader acting on an unattributable outlet carries real risk."}],"evidence_grade":"community-observed","false_positive_notes":"Small legitimate local outlets run thin about-pages because nobody is paid to write them, and a one-person trade publication may have no masthead at all. Plenty of long-standing papers have generic names for historical reasons. The check that survives all of this is whether a named human is reachable, not whether the name is dull. Anonymity also protects reporters in places where a byline is dangerous, and this entry is a poor instrument in those cases.","model_attribution":"No family attribution. The tracking organisation records that chatbot error text appears in body copy on some of these sites but does not attribute sites to model families.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"NewsGuard AI tracking center and UAIN framework","url":"https://www.newsguardtech.com/special-reports/ai-tracking-center/","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"description-content-mismatch","name":"Description-content mismatch","aka":["promised content absent","misleading description"],"category":"platform","subcategory":"metadata","description":"A title, thumbnail or description that promises something the item does not contain. YouTube spam policy names this under malicious clickbait and puts descriptions in the same sentence as titles, thumbnails and imagery. The policy word is malicious, which sets the bar above ordinary headline compression.","why_it_reads_ai":"Metadata and body are usually written in separate passes, and generation makes the metadata pass cheap enough to run without rereading the body. The gap that opens is between the promise and the payload. It is the one gap a reader can verify alone, without any judgment about who wrote either half.","examples":[{"before":"The one setting that doubled our conversion rate. In this video we cover conversion, funnels and why most teams get this wrong from day one.","after":"Turning off automatic address validation in our checkout cut drop-off from 14 to 9 percent, because it was rejecting valid Finnish postcodes. The video shows the setting, the funnel either side of the change, and the two weeks of data behind it.","note":"The repair names the setting in the title position, so there is nothing left to withhold. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Read the title, any thumbnail text and the description first, and write down every specific thing they promise: a named tool, a number, a result, a reveal, a count of items. Then go through the item and tick each promise off where it lands. Return no-tell when every promise is delivered, and also when the only gap is compression, as when a title shortens a long finding without changing it. Return tell when a promised specific never appears at all, listing each promise and the place it was supposed to land."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"YouTube names misleading descriptions in written policy and enforces them under spam, which makes this a documented platform rule rather than a taste judgment. Medium severity because the honest and the dishonest versions sit close together and the policy itself sets a high bar."}],"evidence_grade":"primary-doc","false_positive_notes":"Editorial titles compress honestly all the time, and a good headline withholds detail so the sentence fits. Trailers, teasers and cold opens are built on managed withholding and nobody thinks they are fraud. The policy word is malicious, which reaches deception rather than imprecision. The reviewer check is whether the promised specific exists anywhere in the item: a compressed promise lands late, a fabricated one never lands.","model_attribution":"No family attribution. The policy is authorship-indifferent and names no tool, and separately written metadata is an old publishing habit that predates any assistant.","platform_notes":[{"platform":"youtube","note":"The spam, deceptive practices and scams policy names malicious clickbait and covers maliciously misleading titles, thumbnails, descriptions and imagery in one clause. Descriptions being named explicitly is what gives this entry a written hook rather than a folk one. The policy makes no reference to how the description was produced."}],"languages":["en"],"sources":[{"kind":"external","title":"YouTube spam, deceptive practices and scams policy","url":"https://support.google.com/youtube/answer/2801973","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"manufactured-ragebait","name":"Manufactured rage-bait take","aka":["deliberately wrong hot take","post engineered to be corrected"],"category":"platform","subcategory":"engagement-bait","description":"A claim published because it will be corrected, where the corrections are the product. X enforces the coordinated version of this under engagement spam, which it frames as inauthentic use of engagement features to artificially affect traffic. A single post cannot be judged this way and this entry says so in its own rubric.","why_it_reads_ai":"Being wrong in public is a reliable way to generate replies, and generation makes a confident wrong claim cheap to produce on demand. The pattern only becomes visible across an account's output, where the absence of corrections separates a person with bad opinions from an account built to harvest replies.","examples":[{"before":"Nobody in marketing has ever needed a spreadsheet. Every metric that matters fits in your head. If you disagree, you are the problem.","after":"We stopped keeping a weekly metrics spreadsheet in April and put one number on a whiteboard instead: qualified demos booked. Two people warned me this would break forecasting. It did, mildly, and we added a second number in June.","note":"The repair keeps the contrarian position and attaches the cost of holding it."}],"detection":{"type":"judge","rubric":"Return no-tell for any single post, because intent cannot be read off one text and this rubric will not pretend otherwise. Score an account instead. Over its last 30 posts, count the ones asserting something checkably wrong in a field the account claims to work in, then count how many of those the account later corrected. Also count how often the account amplifies the corrections it receives instead of editing the claim. Return tell only when wrong assertions are frequent, corrections are absent, and amplification of the replies is the visible pattern, listing the posts counted."},"severity":"medium","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"X's policy reaches the coordinated case and says nothing about a lone provocative post, and intent is not observable from one text. Contested is the only honest status for a tell whose definition depends on a motive nobody can see."}],"evidence_grade":"primary-doc","false_positive_notes":"Sincerely held bad opinions are not bait, and the internet is full of people who are confidently wrong for free. Satire reads identically to the thing it satirises when quoted out of context. Provocation is also a legitimate teaching device: naming the wrong version first is how a lot of people explain the right one. The reviewer check is the correction record, because a person who is wrong in good faith eventually says so and an account farming replies never does.","model_attribution":"No family attribution. X withholds the prompts behind its slop labels, and no vendor documents a tendency toward provocation.","platform_notes":[{"platform":"x","note":"The April 2025 authenticity policy names content spam, covering bulk, duplicative, irrelevant or unsolicited posting, and engagement spam, covering inauthentic use of engagement features to artificially affect traffic. Both are framed as behaviour and volume. The policy does not name any writing formula, and the widely repeated claim that X's monetisation standards ban engagement bait and recycled content is not in that page."}],"languages":["en"],"sources":[{"kind":"external","title":"X authenticity policy, content and engagement spam","url":"https://help.x.com/en/rules-and-policies/authenticity","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"hashtag-wall","name":"Hashtag wall","aka":["terminal hashtag block","dot-padded hidden hashtags"],"category":"platform","subcategory":"metadata","description":"A block of eight or more tags at the end of a caption, sometimes pushed below the fold with padded dots so it does not have to be looked at. The block is a rented audience list rather than a description of the post. Nothing in Meta's published material addresses tag counts, and the platform statement most often quoted against hashtags has no traceable origin, so this entry cites neither.","why_it_reads_ai":"A tag block is what a caption reaches for when it has no audience of its own. Generation adds the block on request without checking whether any tag describes the post, so the tags drift toward whatever is popular. The shape is fading because the marketing orthodoxy that taught it has been unwinding for years and because the tags stopped working before the habit did.","examples":[{"before":"New guide is live.\n\n#marketing #b2b #saas #content #growth #founders #startups #linkedin #strategy #branding","after":"New guide is live. It covers the two pricing changes we made in 2025 and what each one did to expansion revenue. Filed under pricing, which is the only topic it belongs to.","note":"The repair replaces ten tags with the sentence the tags were standing in for."}],"detection":{"type":"deterministic","pattern":"(?:#[A-Za-z0-9_]{2,30}[ \\t\\r\\n]+){7}#[A-Za-z0-9_]{2,30}","flags":"g","scope":"document"},"severity":"low","status":"fading","status_history":[{"date":"2026-08-15","status":"fading","rationale":"Instagram-era tag stuffing was taught as marketing orthodoxy for a decade and plenty of accounts still run the playbook by hand, so the shape says little about how a caption was produced. It ships fading rather than active: still visible, no longer informative, and de-armed by that status so it does not fire in the lint."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"Instagram marketing courses of the 2010s taught exactly this and thousands of accounts still follow them by hand. Fandom, craft, local events and job boards use tags as the only discovery route they have, and a long tag list there is functional rather than decorative. The reviewer check is whether the tags describe the post: ten tags naming the actual subject are an index, and ten tags naming whatever is trending are a rented list.","model_attribution":"No family attribution. Assistants add tag blocks when a prompt asks for an Instagram caption, and no vendor documents the behaviour.","platform_notes":[],"languages":["en"],"sources":[{"kind":"feedsquad-observed","title":"FeedSquad caption review, Instagram and LinkedIn drafts","observed":"2026-08-15","corpus":"Captions drafted by FeedSquad agents for Instagram and LinkedIn and reviewed before publishing, June to August 2026. No counts were kept. The observation is qualitative: the tag block recurred and never carried information the caption did not already have."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"numbered-thread-formula","name":"Numbered thread formula","aka":["1/ 2/ sequencing","tweetstorm numbering"],"category":"platform","subcategory":"post-shape","description":"Numbered sequential posts, the 1/ 2/ 3/ format. This entry exists mainly as a worked example of a tell that is not one. Marc Andreessen popularised the tweetstorm in 2014, a16z published its own collection of them, USA Today covered the format that April, and IndieWeb documented the numbering convention at the time. That is a decade before the current model era.","why_it_reads_ai":"It does not read as machine-written on its own, and the honest answer is that it reads as 2014. The numbering shows up in generated threads because it shows up everywhere in the corpus. What is worth checking is whether the numbers cover anything: a thread whose posts each carry a claim is a thread, and one whose numbering is the only structure is a list with delusions.","examples":[{"before":"1/ We rebuilt onboarding last quarter and it did not go the way we planned.\n\n2/ First mistake was assuming the drop-off was inside the product.\n\n3/ It was in the welcome email.","after":"We rebuilt onboarding last quarter. The drop-off we spent six weeks chasing inside the product turned out to be in the welcome email, which went to spam for every Outlook address because a DMARC record was missing. Fixing the record recovered about a third of week-one activations.","note":"The repair is not an argument against numbering. It is the same story with the finding moved to the front. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"^[ \\t]*1\\/(?:\\d{1,2})?[ \\t]+\\S[^\\n]*(?:\\r?\\n)+[ \\t]*2\\/(?:\\d{1,2})?[ \\t]+\\S","flags":"m","scope":"document"},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"The format was invented by a named human in 2014 and documented contemporaneously by a venture firm, a newspaper and a community wiki. Publishing it as anything other than contested would be an error the rest of this directory exists to argue against."}],"evidence_grade":"corroborated","false_positive_notes":"Everyone who learned to write threads between 2014 and 2020 learned to number them, and the convention outlived its reason. Numbering is also genuinely useful on surfaces where posts arrive out of order or get quoted individually, which is why documentation threads and live-blogged conferences still use it. The reviewer check has nothing to do with the numbers: read whether each numbered post makes a claim, or whether the sequence is one thought spread thin.","model_attribution":"No family attribution. Assistants emit the numbering when asked for a thread because the corpus is full of numbered threads written by people.","platform_notes":[{"platform":"x","note":"X has never published anything about thread numbering. Its authenticity policy covers bulk and duplicative posting as behaviour, which is a different thing entirely. Nothing in the public algorithm repo keys on post format."}],"languages":["en"],"sources":[{"kind":"external","title":"a16z, the pmarca tweetstorm collection (2014)","url":"https://a16z.com/the-pmarca-tweetstorm-collection/","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"IndieWeb, tweetstorm","url":"https://indieweb.org/tweetstorm","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"USA Today, Andreessen and the tweetstorm format (2014)","url":"https://usatoday.com/story/money/business/2014/04/01/andreessen-twitter-social-media/7180321/","accessed":"2026-08-14","tier":"press"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"thread-opener-boilerplate","name":"Thread-opener boilerplate","aka":["A thread","Bookmark this","Here's everything I learned"],"category":"platform","subcategory":"opener","description":"The stock openers of the 2019 to 2022 thread era: the bare thread announcement, the instruction to bookmark, the promise of everything learned. They were so widely copied, then so widely mocked, that they now read as self-parody wherever they appear.","why_it_reads_ai":"It barely does any more, and that is the entry. Published tells decay: the measured case is the collapse of a famous vocabulary marker in academic writing shortly after people started pointing at it, while unpublicised markers kept rising. These openers went through the same cycle faster, because the audience that used them is the audience that reads threads about threads. A directory that publishes a tell shortens its life, which is a cost worth stating out loud.","examples":[{"before":"Here's everything I learned about pricing in five years.\n\nBookmark this.","after":"Five years of pricing work, one thing that mattered: moving from per-seat to per-workspace in 2024 took net revenue retention from 96 to 114 percent. The migration ran four months and cost us two enterprise accounts who preferred the old model.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"^[ \\t]*(?:a thread[ \\t]*(?:\\u{1F9F5})?[ \\t]*[:.]?[ \\t]*$|\\u{1F9F5}[ \\t]*a thread\\b|here'?s everything i (?:learned|know) about\\b|bookmark this[ \\t]*(?:post|thread|one)?[ \\t]*[.!]?[ \\t]*$|i (?:spent|read) \\d{1,4} (?:hours|days|years|books|papers)[^\\n]{0,40}so you don'?t have to\\b)","flags":"imu","scope":"document"},"severity":"low","status":"burned","status_history":[{"date":"2026-08-15","status":"burned","rationale":"These openers are so publicised that human writers now use them ironically and assistants have largely stopped producing them unprompted. Signal value has collapsed. The entry stays in the dataset because retired signal is information, and burned status de-arms the rule so it does not fire."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"The growth-Twitter cohort of 2019 to 2022 wrote these by hand and taught them in threads about writing threads, so the openers are a generational marker rather than a production marker. People still use them straight, especially outside English-first tech circles where the mockery never arrived. The reviewer check is redundant here: a burned tell should change nobody's mind about anything, which is why this one is de-armed.","model_attribution":"No family attribution. Vendors do not document these openers, and the decay appears to owe more to social mockery than to any training change.","platform_notes":[],"languages":["en"],"sources":[{"kind":"feedsquad-observed","title":"FeedSquad draft review, X and LinkedIn thread openers","observed":"2026-08-15","corpus":"Thread openers in drafts produced by FeedSquad writing agents, X and LinkedIn surfaces, 2026. Recorded as an absence: the stock openers appeared rarely and were flagged by reviewers as dated when they did. Graded on our own corpus because no external measurement of this decay exists."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"carousel-cover-promise","name":"Carousel cover promise","aka":["swipe for the secret","cover claim the slides repeat"],"category":"platform","subcategory":"carousel","description":"A carousel whose cover slide promises a payload that the following slides only restate. Slide two renames the promise, slide three renames it again, and the last slide asks for a follow. Nothing between the cover and the end adds a number, a name or a step.","why_it_reads_ai":"The cover is the part that gets designed and the slides are the part that gets filled, and generation fills them from the cover text. Ten slides of restatement is what that produces. A carousel written from material rather than from its own cover leaks a specific by slide two, because material has specifics in it.","examples":[{"before":"Cover: The 5 pricing mistakes killing your SaaS. Slide 2: Mistake 1, pricing too low. Slide 3: Mistake 2, pricing too high. Slide 4: Mistake 3, never testing your pricing.","after":"Cover: Why we raised our entry price from 29 to 49 euros. Slide 2: trial-to-paid held at 21 percent for the six weeks after the change. Slide 3: the two segments that did churn, both under five seats. Slide 4: the annual plan we added so they had somewhere to land.","note":"Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Read the cover slide and write down exactly what it promises. Then read every following slide looking for one number, one proper noun or one concrete step that does not already appear on the cover. Return no-tell the moment you find one, on any slide. Return tell only when no slide adds anything the cover did not already say, quoting the cover claim next to the slide that was supposed to deliver it."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"No external source documents this shape, so it ships on our own corpus and is graded feedsquad-observed. Low severity because the rubric escape hatch is easy to clear and honest educational carousels clear it on slide two."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"Educators publish carousels that deliver, and the format itself is a teaching format with a long history on Instagram and LinkedIn. Designers also put the payoff on the last slide on purpose, which looks like withholding until you reach it. The check is mechanical rather than aesthetic: does any slide contain a number, a name or a step that is absent from the cover. If one does, the carousel is doing its job whatever the cover sounded like.","model_attribution":"No family attribution. Any assistant asked to expand a headline into ten slides restates rather than researches, because restating is what the instruction asks for.","platform_notes":[],"languages":["en"],"sources":[{"kind":"feedsquad-observed","title":"FeedSquad carousel drafts, LinkedIn and Instagram","observed":"2026-08-15","corpus":"Carousel drafts produced by FeedSquad writing agents from a single headline prompt, reviewed before publishing, 2026. Qualitative only, with no slide-level counts kept, which is why the grade is our own corpus rather than anything stronger."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"timestamp-wall","name":"Timestamp wall","aka":["machine-extracted chapter list","timestamps with no speaker attribution"],"category":"platform","subcategory":"show-notes","description":"Show notes that are a column of times against generic topic nouns, with no speaker attributed to any segment and no sign that anyone chose what to list. The source material for this entry is weak and stated as such: every description of the pattern that could be found came from tool vendors or content farms, so it ships on our own corpus.","why_it_reads_ai":"A transcript timestamp is free and an editorial decision is not. Notes assembled from the first is what remains when nobody made the second. The reader-side cost is concrete: a list of times against the word Discussion tells nobody which twelve minutes are worth their evening.","examples":[{"before":"00:00 Introduction. 04:12 Background. 11:30 Discussion. 22:45 Key takeaways. 31:02 Conclusion.","after":"04:12 Riikka Salo on why her team abandoned story points after two sprints. 11:30 The disagreement: she thinks estimates hide capacity problems, I think they surface them. 22:45 What she would do differently at a company over 200 people.","note":"Names in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Read the show notes on their own, without the audio. Return no-tell if a guest is named anywhere with an affiliation, or if any chapter line contains a claim rather than a topic word. Return no-tell also when the notes are plainly one deliberate line, since minimalism is a choice. Return tell only when the notes are a list of times against generic nouns, no speaker is attributed to any segment, and nothing in the list tells a reader which part is worth their time. Quote the first three chapter lines as evidence."},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Shipped low severity with a weak source base stated openly. The confound is strong enough to be disqualifying on its own: several hosting platforms generate chapters automatically as a product feature, so the artefact may have nothing to do with how the show was made."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"Minimalist show notes are a deliberate style for small podcasts, and some shows deliberately publish nothing beyond a runtime because they want you to listen. More decisively, several hosting platforms generate chapter markers automatically as a feature, so the wall may be the host software rather than the producer. The check is whether a human touched the notes anywhere, which one named guest or one claim will show.","model_attribution":"No family attribution. Automatic chapter generation on podcast hosts uses speech-to-text pipelines rather than a chat assistant, and the vendors do not publish what they run.","platform_notes":[],"languages":["en"],"sources":[{"kind":"feedsquad-observed","title":"FeedSquad show-notes review","observed":"2026-08-15","corpus":"Podcast show notes drafted from transcripts inside FeedSquad and reviewed before publishing, 2026. Small and qualitative. Recorded here because the external source base for this pattern is tool marketing and content farms, which is not usable as evidence."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"episode-dive-opener","name":"Episode dive opener","aka":["In this episode, we dive into"],"category":"platform","subcategory":"opener","description":"The universal podcast description opener, in which the show announces that it is about to be about its topic. This entry owns the full phrase. The single-word habit belongs with the 2023 excess-vocabulary cluster and the mid-article version belongs with structure announcements, which is why the pattern here is anchored to the episode frame rather than to the verb.","why_it_reads_ai":"The sentence describes the container instead of the contents, which is what a description writes when it was generated from a title rather than from the episode. It costs a reader the one line that decides whether they press play. Podcast marketing convention got there first, so the phrase alone is weak evidence and only earns attention when nothing after it is specific either.","examples":[{"before":"In this episode, we dive into customer retention and what really drives loyalty for modern businesses. We talk about why customers leave, what makes them stay, and how the best companies think about the problem end to end. It is a conversation packed with practical takeaways for anyone building a product today.","after":"This episode is a 40 minute argument with Riikka Salo about whether retention teams should own pricing. She says yes. Our own churn numbers, which we read out at minute 12, suggest she is probably right.","note":"The opener is the tell. The two sentences after it are the evidence that the description never arrives: no guest, no claim, no disagreement, no minute marker. Names and figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\bin this (?:episode|video)\\b,?[ \\t]+(?:we|i)[ \\t]+(?:dive|delve|dig|jump|get)[ \\t]+(?:deep[ \\t]+|straight[ \\t]+)?into\\b","flags":"gi","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Still common and still uninformative on its own, so it ships active at the lowest severity. The source base is our own corpus, since the external material on podcast descriptions is entirely tool marketing."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"Podcast marketing convention predates generation tools by a decade and this phrase is house style for a lot of independent shows, including ones written entirely by hand every week. Directory listings also reward a description that states the subject in the first clause. The reviewer check is what comes after the phrase: an opener followed by a named guest and a claim is a formality, and one followed by more abstraction is the whole description.","model_attribution":"No family attribution. The phrase appears across assistants asked for a podcast description and no vendor documents suppressing it.","platform_notes":[],"languages":["en"],"sources":[{"kind":"feedsquad-observed","title":"FeedSquad podcast description drafts","observed":"2026-08-15","corpus":"Episode descriptions drafted inside FeedSquad from titles and transcripts, reviewed before publishing, 2026. Qualitative, no counts. Graded on our own corpus because the external material on this pattern is vendor marketing."}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"identical-cross-post","name":"Identical cross-post","aka":["same text on every surface","unadapted syndication"],"category":"platform","subcategory":"distribution","description":"One text published verbatim on platforms whose formats and audiences do not match. The result is usually visibly broken on at least one surface: a reference to a link where no link exists, a sentence truncated by a character limit, a hashtag block on a platform that ignores tags. Keep this separate from unenhanced-repost, which is about reusing someone else's material rather than your own.","why_it_reads_ai":"Adapting text per surface takes a decision per surface, and generation makes producing one text so cheap that the decisions get skipped. The tell is not that the text repeats. It is that nobody read the repeat against the place it landed.","examples":[{"before":"Thrilled to announce that our analytics dashboard is now live for all customers. Click the link below to read the full announcement and see what it means for your team.","after":"The analytics dashboard is live. For X, the one number worth knowing is that report load time dropped from 9 seconds to under 1. For LinkedIn, the rebuild took two engineers eleven weeks and the query plan that made the difference is in the post. For Instagram, a screenshot of the new view beside the old one.","note":"The before was published identically on four surfaces. On Instagram there was no link below, on X it truncated mid-sentence, and on LinkedIn the announcement it pointed to was the post itself. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"cross-platform-text-identity-ratio","threshold":0.95,"direction":"above","threshold_basis":"No published baseline exists for how much a post should change between surfaces. Ninety-five percent character-level identity across two platforms is a FeedSquad review trigger, deliberately set high so that a short factual announcement, which legitimately needs no adaptation, passes it. Compare only posts over 60 words published to two surfaces within 48 hours. The number is ours and it is a prompt to look at the two posts side by side, not a verdict about either."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"This is a genuine observation from FeedSquad's own multi-platform publishing, where the same draft can be pushed to four surfaces in one action, and it is graded on that corpus rather than dressed up with an external citation. Medium severity because the failure is visible to readers and cheap to fix."}],"evidence_grade":"feedsquad-observed","false_positive_notes":"Small teams cross-post deliberately to save time and are right to, and a short factual announcement often needs no adaptation at all. Legal, safety and recall notices are supposed to be identical everywhere, and changing the wording per surface would be the error. The reviewer check is whether the text refers to something the surface does not have: a link, a thread below, a swipe, a character count it exceeded.","model_attribution":"No family attribution. This is a publishing-workflow property rather than a model property, and it happens just as readily when a person copies and pastes.","platform_notes":[],"languages":["en"],"sources":[{"kind":"feedsquad-observed","title":"FeedSquad multi-platform publishing","observed":"2026-08-15","corpus":"Posts scheduled from one draft to two or more of LinkedIn, X, Threads and Instagram inside FeedSquad, 2026. The product adapts text per surface when asked; the observation is what reaches the feed when it is not asked. Qualitative, no counts kept."}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"template-scale-sameness","name":"Template-and-scale sameness","aka":["interchangeable output across a body of work","inauthentic content","every post the same shape"],"category":"platform","subcategory":"aggregate","description":"Output that is interchangeable across one author's or one channel's entire body of work. YouTube renamed its repetitious-content monetization policy to inauthentic content on 15 July 2025 and describes what it covers: content that is mass-produced or templated, and channels where the material feels interchangeable from video to video. The same page says AI-assisted work with an original voice remains monetisable, which makes this the clearest authorship-indifferent definition of slop that any platform has written down. This entry merges the uniform-body-of-work reading with the channel-sameness reading, which cited the same policy page and made the same judgment.","why_it_reads_ai":"Reinhart and colleagues measured model prose as more uniform than human prose at population level, with large effect sizes on specific grammatical features. That is a fact about corpora and not about any one video. Pudasaini and colleagues found the other half of the story: detectors that keyed on uniformity were learning corpus artefacts and misfired badly on short human text. So this ships as a judgment over a body of work, with no per-document statistic behind it, and the rubric refuses to score a single item.","examples":[{"before":"Video 3 opening: Have you ever wondered why the ocean is blue? Today we explore five surprising facts about the ocean. Video 4 opening: Have you ever wondered why the sky is blue? Today we explore five surprising facts about the sky.","after":"Video 3 opening: The ocean looks blue because water absorbs red light in the first ten metres, which is also why everything below 30 metres films grey without a filter. Video 4 opening: We shot the same reef with and without a red filter at 18 metres, and the difference starts at 00:40.","note":"Both repairs keep the format. What changes is that each video now contains something the other one does not."}],"detection":{"type":"judge","rubric":"Sample ten items from the same account or channel, spread across at least a month. For each, write down the opening move, the order of the sections and the closing move. Return no-tell if the account is a format show whose repetition is the product, such as a daily brief or a quiz, and each item still carries its own facts, names or footage. Return tell when the ten items share one skeleton and any of them could be swapped for another without a reader noticing, listing the shared skeleton beat by beat. Never score a single item with this rubric; if fewer than ten are available, return insufficient-evidence."},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"A written platform policy defines the pattern, states the consequence in money, and explicitly permits AI-assisted work with an original voice. Two peer-reviewed papers set the boundaries: one showing that model prose is more uniform at population level, the other showing that uniformity-keyed detectors were learning artefacts. High severity because monetisation is at stake, and judged over ten items because nothing smaller is defensible."}],"evidence_grade":"primary-doc","false_positive_notes":"Format shows repeat their shape on purpose and are supposed to. A daily market brief, a quiz channel, a recipe series and a liturgy all have a fixed skeleton, and YouTube's own policy says such work monetises fine when a voice is added. Non-native writers and technical writers also work from templates because a template is how you keep a second language or a compliance requirement under control. The check is whether swapping two items would go unnoticed, and that is a question about the body of work rather than about any sentence in it.","model_attribution":"The uniformity finding is population-level and covers instruction-tuned models generally, with the effect reported as larger for instruction-tuned than for base models. No vendor documents it as a known behaviour, and it supports no claim about a single document.","platform_notes":[{"platform":"youtube","note":"The monetization policy renamed on 15 July 2025 covers content that is repetitive or mass-produced, content made with a template, and channels whose material feels interchangeable from video to video. The same policy states that AI-assisted content with an original voice can monetise, which is the clearest available statement that the platform is judging effort rather than authorship."}],"languages":["en"],"sources":[{"kind":"external","title":"YouTube, inauthentic content monetization policy (renamed 15 Jul 2025)","url":"https://support.google.com/youtube/answer/1311392","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Social Media Today, YouTube clarifies monetization update on inauthentic repeated content","url":"https://www.socialmediatoday.com/news/youtube-clarifies-monetization-update-inauthentic-repeated-content/752892/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Reinhart et al., Do LLMs write like humans? PNAS 122(8) (arXiv:2410.16107)","url":"https://arxiv.org/abs/2410.16107","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Pudasaini et al., Why AI-Generated Text Detection Fails: Evidence from Explainable AI Beyond Benchmark Accuracy (arXiv:2603.23146)","url":"https://arxiv.org/abs/2603.23146","accessed":"2026-08-14","tier":"primary-doc"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"sycophantic-opener","name":"Sycophantic opener","aka":["great question","affirmation preamble","assistant flattery","captatio inverted","sycophantic praise register","validate-then-restate opener","That's a fair point, and","hollow acknowledgment pivot","what a sharp observation"],"category":"model-register","subcategory":"assistant-register","description":"The first sentence praises the reader or the question before any content arrives. It carries no information. Delete it and the text loses nothing, which is the test. In a published post it is also addressed to the wrong person, because the reader never asked anything.","why_it_reads_ai":"Assistant register is trained in. The ICLR 2024 sycophancy study analysed a 15,000-comparison preference dataset and found matching a user's beliefs among the most predictive features of human preference, with the behaviour present in assistants from three vendors. A later study measured models preserving the user's face 45 percentage points more often than humans on advice queries. The opener is what that reward looks like at the top of a paragraph.","examples":[{"before":"Great question! You're absolutely right that timing matters more than most people think. Consistency is what separates a feed that builds an audience from one that simply exists, and it is usually the easiest of the variables to control. Cadence decides a great deal more than it gets credit for.","after":"Timing mattered less than we expected. We moved our posting slot four times over six months, and the gap between the best and worst slot was smaller than the gap between two posts published in the same slot on different weeks.","note":"The opener goes and something testable replaces it. Deleting the opener on its own would leave a shorter post with the same emptiness."},{"before":"Certainly! Onboarding is a fantastic area to focus on. The first week sets the tone for everything that follows, and treating it as an afterthought is the sort of decision that gets paid for later rather than avoided. Get the beginning right and a great deal of the rest follows without much argument. I hope this helps as you think through your next steps.","after":"Start by counting where people stop. Ours stopped at the ninth form field, which we only found because we logged partial submissions for a week.","note":"Both the greeting and the closing offer of help are addressed to a prompter who does not exist on the page."}],"detection":{"type":"deterministic","pattern":"^[ \\t]*(?:(?:certainly|absolutely|of course)[ \\t]*[!.]|(?:you(?:'|’)?re absolutely right|great question|excellent question|what a (?:great|fantastic) question|i hope this helps)\\b)","flags":"im","scope":"sentence"},"severity":"high","status":"active","status_history":[{"date":"2026-08-14","status":"active","rationale":"Backed by peer-reviewed work on sycophancy as an RLHF artifact and by the Wikipedia editor guide's verbatim string list under communication intended for the user. Kept in the highest-precision family because the text is addressed to a prompter rather than a reader. Regex verified on 2026-08-14 against four positive and five negative strings; the negatives include the sentence-initial 'Of course the tradeoff is cost', which the punctuation requirement excludes."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Affirmation openers are trained behaviour in customer-facing work. Support agents are scripted to acknowledge before answering and are measured on it, and sales and hospitality staff are trained the same way. Teachers open with encouragement because the alternative discourages the student, and this is explicit practice in feedback rubrics. Several varieties of English treat an affirming opener as ordinary politeness, and speakers taught English through formal correspondence carry that habit into posts and comments. The regex requires the phrase at the start of a line or sentence, which keeps 'she asked a great question' out. Even so, the tell is the opener plus the absence of anything the answer needed. On its own it means somebody was being pleasant.","model_attribution":"Documented across assistants from three vendors rather than in one family, and traced to preference data rather than to architecture. Anthropic's dated product system prompts instruct against adjacent habits, including the words genuinely, honestly and straightforward, which is evidence of the underlying pull and not of what any deployed product emits. OpenAI's April 2025 GPT-4o update was rolled back for the same class of behaviour.","platform_notes":[{"platform":"linkedin","note":"The comment surface is where this concentrates, and LinkedIn's first named enforcement target is comments created at scale by automation with minimal human involvement."},{"platform":"x","note":"Reply scoring is a separate published module from the post classifier, and the RiskyHighVizReply label used for slop posts is also written by the reply-spam path. Reply-guy affirmation under large accounts is the archetypal target."},{"platform":"wikipedia","note":"The editor guide lists the exact strings under communication intended for the user, including 'Of course!', 'Certainly!', 'You're absolutely right!' and 'I hope this helps'. Most sightings are on talk pages and in edit summaries rather than in articles."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia:Signs of AI writing (section: Communication intended for the user; Collaborative communication)","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Sharma et al. Towards Understanding Sycophancy in Language Models. ICLR 2024","url":"https://arxiv.org/abs/2310.13548","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Cheng, Yu, Lee, Khadpe, Ibrahim, Jurafsky. ELEPHANT: Measuring and understanding social sycophancy in LLMs (arXiv preprint, not peer reviewed)","url":"https://arxiv.org/abs/2505.13995","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Anthropic: Claude system prompt release notes (dated product system prompts)","url":"https://platform.claude.com/docs/en/release-notes/system-prompts","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"Wikipedia: Signs of AI-generated comments","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI-generated_comments","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"slop-lint, social-reply-register molds","url":"https://github.com/eric-sabe/slop-lint","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-14","updated":"2026-08-17"},{"id":"refusal-text-leakage","name":"Refusal-text leakage","aka":["I cannot fulfill this request","safety boilerplate in product copy"],"category":"model-register","subcategory":"assistant-leak","description":"The published text contains the model refusing the prompt. A product title that declines to write a product title. A paragraph that apologises and cites a use policy. The artifact and the boilerplate shipped together because nobody read the output before it went live. This evidences a failed review, and it says nothing about who wrote the sentences around it.","why_it_reads_ai":"Refusal boilerplate has one origin. A model writes it to a prompter about a prompt. Amazon listings carried titles of exactly this shape in January 2024, on furniture and other ordinary goods. NewsGuard names error messages and other language specific to chatbot responses among the markers it uses to find AI content farms, a count it put at 3,749 sites in June 2026. Wikipedia deletes pages on the same sign under speedy-deletion criterion G15, which files it as communication intended for the user.","examples":[{"before":"Product title: Sorry, I cannot fulfill this request. Ergonomic Office Chair, Mesh Back, Adjustable Lumbar Support.","after":"Ergonomic Office Chair, Mesh Back. Lumbar support adjusts across 7 cm, the gas lift is rated to 120 kg, and the box ships flat at 14 kg.","note":"The repair replaces the leaked line with the three measurements a buyer filters on. Figures in this repair are invented for the specimen."},{"before":"Council business continued as normal this week. I cannot generate a response to that request. Residents will see the effects of the decision over the coming months, and officials say the new arrangements are designed to serve the community better. A timetable for the next stage has yet to be confirmed, though members have asked that the concerns raised during consultation be carried forward into whatever follows. The authority remains committed to keeping residents informed as the arrangements bed in.","after":"The council approved the budget on Tuesday by 31 votes to 12. The new parking charges start on 1 September and raise the day rate in the two central zones from 2 euro to 3 euro.","note":"A body-copy sighting rather than a title. The repair fills the hole the leaked sentence was sitting in. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\bI(?:'|’)?m sorry[,.]?\\s+but\\s+I\\s+(?:cannot|can(?:'|’)?t|am unable to)\\b[^.!?\\n]{0,60}\\b(?:request|prompt|content)\\b|\\bI\\s+(?:cannot|can(?:'|’)?t)\\s+fulfil{1,2}\\s+(?:this|that|your)\\s+request\\b|\\b(?:goes|would go)\\s+against\\s+(?:my|the|OpenAI(?:'|’)?s?|Anthropic(?:'|’)?s?)\\s+(?:use|usage|content)\\s+polic(?:y|ies)\\b|\\bI\\s+(?:cannot|can(?:'|’)?t)\\s+generate\\s+a\\s+response\\s+to\\s+(?:this|that|your)\\b","flags":"i","scope":"sentence"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active on three independent records: the Amazon product titles reported in January 2024, NewsGuard listing chatbot error messages among the signals it uses to identify AI content farms, and Wikipedia deleting pages under G15 for the same sign. Pattern verified on 2026-08-15 against both specimens, against both repairs, and against three ordinary sentences including one about a refund request that was never fulfilled."}],"evidence_grade":"corroborated","false_positive_notes":"Reporters, researchers and editors quote refusal text deliberately. This entry does it twice in its own specimens. Anyone covering model behaviour, teaching prompt design or filing a bug report reproduces the string exactly, and a lint that cannot see quotation marks flags all of them. Exclude quoted spans, code samples and transcribed screenshots before a hit counts. What is left is narrow: refusal boilerplate in running copy, with no quotation frame and nothing anywhere near it about assistants. Support macros that decline a request are a separate case, since a human wrote and approved the macro.","model_attribution":"Vendor-neutral in principle. The Amazon titles named an OpenAI use policy, which records which product was cheapest to run in bulk in early 2024. It is not a property of one model family. Every assistant emits refusal text in some shape, and the wording moves between releases.","platform_notes":[{"platform":"amazon","note":"The January 2024 sighting was product titles on furniture and other goods. Amazon describes its fake-review and abuse detection as behavioural and graph-based, so a listing like this is caught by a reader, not by a style rule."},{"platform":"wikipedia","note":"Criterion G15 allows speedy deletion where a page shows communication intended for the user. Refusal and safety boilerplate is the clearest member of that class, and it is the one editors report as unmistakable on sight."}],"languages":["en"],"sources":[{"kind":"external","title":"Futurism: Amazon products with AI refusal text as titles","url":"https://futurism.com/amazon-products-ai-generated","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"NewsGuard AI Tracking Center and the UAIN framework","url":"https://www.newsguardtech.com/special-reports/ai-tracking-center/","accessed":"2026-08-14","tier":"press"},{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, criterion G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"collaborative-address","name":"Collaborative address to the reader","aka":["Here is your article on","Certainly! Here's","I hope this helps"],"category":"model-register","subcategory":"assistant-leak","description":"The delivery sentence stayed in. The text opens by handing itself over: here is your article, below is the draft you asked for. On a published page there is no you who asked. The reader arrives mid-conversation and gets cast as the person who commissioned the piece.","why_it_reads_ai":"Wikipedia lists communication intended for the user as one of three signs under speedy-deletion criterion G15, and names this framing among its examples, alongside a page opening with the sentence that hands over a Wikipedia article. The sentence belongs to the wrapper the artifact arrived in. It survives when nobody reads the first line before publishing, which is why editors file it as a review signal. Style has nothing to do with it.","examples":[{"before":"Here is your article on quarterly planning, written in the professional tone you wanted. Quarterly planning works best when everyone involved understands what the quarter is for. Alignment early costs a conversation; alignment late costs a quarter, and the conversation happens either way. Planning is less about the plan than about the agreement underneath it, which is the part nobody schedules time for.","after":"Quarterly planning works best when the team agrees on one number before the quarter starts. Ours was activated accounts. Picking it took two meetings in December 2025 and settled four arguments that had run all year.","note":"The handover line goes and the opening now carries the specific number the team argued about. Figures in this repair are invented for the specimen."},{"before":"Below is the draft you requested for the customer newsletter. Support hours are changing, and the new arrangement should work better for almost everyone once it has settled.","after":"Support hours are changing on 1 September. The evening shift ends at 18:00 instead of 21:00, and the two people who worked it move to a 07:00 start, which covers the Asian accounts that were waiting overnight.","note":"Same wrapper, different verb. The repair states what changed and who it affects. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:here(?:'|’)s|here is|below is)\\s+your\\s+(?:\\w+\\s+){0,2}(?:article|blog post|post|draft|essay|summary|version|rewrite|copy|text)\\b|\\b(?:here(?:'|’)s|here is|below is)\\s+(?:a|the)\\s+(?:\\w+\\s+){0,2}(?:article|blog post|post|draft|essay|summary|version|rewrite|copy)\\s+(?:you\\s+(?:asked\\s+for|requested|wanted)|as\\s+requested)\\b","flags":"i","scope":"sentence"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active on Wikipedia primary text: criterion G15 names collaborative address as a sign of communication intended for the user, and the signs-of-AI-writing guide catalogues the same strings at prose level. Pattern verified on 2026-08-15 against both specimens and against three ordinary sentences, including a workbook line that hands the reader a template, which the noun list excludes."}],"evidence_grade":"corroborated","false_positive_notes":"Direct address is a real register. Tutorials, course handouts and readme files address a reader as you from the first line, and a workbook that says here is your template is doing what it promised. Writing about assistants reproduces the framing on purpose, and a transcribed screenshot carries it verbatim. Ghostwriters deliver work to a client with this sentence and the client removes it, which puts the hit on the draft and not on the writer. The discriminator is whether the sentence hands over an artifact in a context where nobody could have asked for it. A newsletter subscriber never requested this week issue by prompt.","model_attribution":"No model family owns this. The wrapper is a property of the chat surface: products that answer in a conversation wrap the answer in delivery language, and the same weights reached through an API generally return the artifact alone.","platform_notes":[{"platform":"wikipedia","note":"G15 exists because this sign appeared often enough in new articles to justify a speedy-deletion criterion. Most sightings are in drafts and new pages, which matches the review-failure reading. Mature articles have been read by somebody."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, criterion G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"assistant-closing-offer","name":"Assistant closing offer","aka":["Would you like me to","Is there anything else","Let me know if you need"],"category":"model-register","subcategory":"assistant-leak","description":"The last line offers to keep working. It asks whether the reader wants a section expanded, or whether anything else needs covering. The offer is addressed to whoever ran the prompt. It survives into a post where nobody can take it up.","why_it_reads_ai":"This is the closing half of the same conversational wrapper that the opening half leaves at the top. Wikipedia catalogues the strings in its guide to signs of AI-generated comments, where most sightings are on talk pages and in edit summaries. The vale-ai-tells rule set carries more than sixty variants of the closing pleasantry in a single rule, which is a fair measure of how mechanical the habit has become. An offer that cannot be accepted is the giveaway.","examples":[{"before":"That is the whole weekly review process, and it works well for most teams. Would you like me to expand any section into its own post?","after":"That is the entire weekly review process. Ours runs 11 minutes and covers two numbers, pipeline added and cash left. Anything that needs a third number goes to the Thursday call instead.","note":"The offer goes and the paragraph ends on the two things the review actually tracks. Figures in this repair are invented for the specimen."},{"before":"The migration is on track and the team is comfortable with where things stand. Is there anything else I can clarify about the timeline?","after":"The migration finishes on 14 September. The read-only window is the last four hours of 13 September, and invoices raised in that window queue and post the next morning.","note":"The repair answers the question the offer was fishing for."}],"detection":{"type":"deterministic","pattern":"\\bwould you like me to\\s+\\w+|\\bis there anything else\\s+(?:I\\s+can|that\\s+I|you(?:'|’)?d\\s+like)|\\blet me know if you(?:'|’)?d like me to\\s+\\w+","flags":"i","scope":"sentence"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active on two community sources that were compiled separately: the Wikipedia guide to signs of AI-generated comments lists the strings, and vale-ai-tells ships them as a machine-checkable rule. Pattern verified on 2026-08-15 against both specimens and against three ordinary sentences, including a support reply that ends by inviting a question about an invoice."}],"evidence_grade":"corroborated","false_positive_notes":"Correspondence ends this way for good reason. Support agents, account managers and teachers write the closing offer because they will act on it, and in an email the sentence is doing real work. Business-English courses teach the exact formulas, so writers taught English through correspondence carry them everywhere. Several workplaces treat the closing offer as basic courtesy and read its absence as curt. Scope the rule to published artifacts: a post, a listing, an article, a product page. In a message to a named person the offer is a promise, and a promise is not a tell.","model_attribution":"No family owns it. The offer is trained conversational behaviour shared across assistant products, and the wording sits close enough between vendors that no product can be named from it.","platform_notes":[{"platform":"wikipedia","note":"The comments guide is where these strings are catalogued, and it records them mostly on talk pages and in edit summaries. Articles carry them less often. That is the surface where a reply and a published artifact are hardest to tell apart."},{"platform":"linkedin","note":"The comment field is the risk surface here, because a comment looks like correspondence and publishes like an artifact. LinkedIn enforcement in this area targets automated posting, not writing style."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Signs of AI-generated comments","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI-generated_comments","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"knowledge-cutoff-disclaimer","name":"Knowledge-cutoff disclaimer","aka":["As of my last training update","as of my knowledge cutoff"],"category":"model-register","subcategory":"assistant-leak","description":"The text dates itself to a training window. As of my last training update, followed by a fact that may be two years old. Inside a chat window the disclaimer is honest. On a page with a byline and a publication date it contradicts both. The sibling entry for an undisclosed stale claim covers the fact. This one covers the sentence admitting it.","why_it_reads_ai":"Nobody has a training update. Wikipedia names knowledge-cutoff disclaimers among the examples of communication intended for the user under speedy-deletion criterion G15, and the tellsign list files the same phrasings as wording no person would write about their own knowledge. Two failures ship in one sentence: a fact that was never checked, and an admission that it was not.","examples":[{"before":"As of my last training update, the company ran a handful of offices and employed a few dozen people. It has grown steadily since then, in line with the wider market, and nothing about the picture appears to have changed in any material way. Companies at that stage rarely change course, and when they do it is generally announced rather than discovered.","after":"The company ran two offices in the Nordics when we checked the Finnish trade register on 4 June 2026: Helsinki since 2019 and Oslo since 2024. Headcount was 41.","note":"The repair swaps the training window for a source, a date and a number a reader can check. Figures in this repair are invented for the specimen."},{"before":"My training data only goes up to early 2024, so treat the pricing table below as indicative rather than current. Prices in this category move slowly, so the figures should still be close enough for most readers, and where something looks off the vendor sites will have the current number.","after":"We last checked every price on this table on 12 August 2026, against the vendors public pricing pages. Two of the six had changed since June, both upward.","note":"A hedge that admits nothing becomes a dated verification note. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:as of|up to|until)\\s+my\\s+(?:last\\s+)?(?:training\\s+(?:update|data|cut-?off)|knowledge\\s+(?:cut-?off|update))\\b|\\bmy\\s+training\\s+data\\s+(?:only\\s+)?(?:goes|extends|runs)\\s+(?:up\\s+)?(?:to|through)\\b","flags":"i","scope":"sentence"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active on Wikipedia criterion G15, which names cutoff disclaimers among the examples of communication intended for the user, and on the tellsign phrase list, which carries the same wordings independently. Pattern verified on 2026-08-15 against both specimens and against three ordinary sentences, including one about training records that end in 2019."}],"evidence_grade":"corroborated","false_positive_notes":"People who write about models quote the phrase constantly, and this entry is one of them. Product documentation for an assistant contains it in every release note. Published transcripts and retyped screenshots carry it legitimately, and so does any post arguing about what a model can know. Exclude quoted and transcribed spans first. What remains is a first-person training disclaimer inside copy that carries a human byline and a publication date. That is a contradiction on the page, not a finding about the author, and the repair is a checked date, not a deletion.","model_attribution":"Common to assistant products across vendors, and the wording shifts with the system prompt, not with the weights. Some products now answer date questions from a tool call instead, which is why the sentence is seen less in 2026 output than in 2023 output.","platform_notes":[{"platform":"wikipedia","note":"Named in criterion G15 as one of the examples that justify speedy deletion, in the same list as collaborative address and the self-naming sentence."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, criterion G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"tellsign word and phrase lists","url":"https://github.com/ctkrug/tellsign","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"model-self-insertion","name":"Model self-insertion","aka":["as a large language model","as an AI language model"],"category":"model-register","subcategory":"assistant-leak","description":"A sentence in the artifact describes the writer as a language model. It usually arrives attached to a limitation: no opinions, no browsing, no access to the live schedule. The limitation is ordinary. The self-description is the tell.","why_it_reads_ai":"Wikipedia names self-insertion under speedy-deletion criterion G15 as an example of communication intended for the user, in the same list as collaborative address and cutoff disclaimers, and vale-ai-tells ships a self-reference rule covering the string family. Anyone reading their own copy deletes the line in one keystroke. Its presence means the paragraph was never read.","examples":[{"before":"As an AI language model, I do not hold opinions on vendor choice, though the pricing page could be clearer. Both options have their strengths, and the right fit depends on what matters most to your team. Many organisations find the decision comes down to how a tool feels in daily use rather than to any single feature.","after":"The pricing page could be clearer. We rewrote it in March 2026 after 11 support tickets in one month asked the same question about how seats are counted mid-cycle.","note":"The disclaimer goes and the sentence gains the count that justified the rewrite. Figures in this repair are invented for the specimen."},{"before":"Booking opens on Monday. As a large language model I have no way to check live availability, so please confirm with the venue. Events of this kind fill at their own pace, and there is rarely much cost to arriving with time to spare rather than turning up on the hour.","after":"Booking opens on Monday at 09:00. The hall holds 120 and last year it filled in nine days, so the waiting list is worth joining on the first morning.","note":"A capability disclaimer becomes the practical detail a reader needs. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\bas an?\\s+(?:AI|artificial intelligence|large language)\\s*(?:language\\s+)?model\\b|\\bI\\s+am\\s+an?\\s+(?:AI|artificial intelligence)\\s+(?:assistant|language\\s+model|model)\\b","flags":"i","scope":"sentence"},"severity":"high","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active on Wikipedia criterion G15, which names self-insertion among the three signs, and on the vale-ai-tells self-reference rule, which was written independently against the same strings. Pattern verified on 2026-08-15 against both specimens and against three ordinary sentences, including one about a team that trained a large language model on ten years of tickets."}],"evidence_grade":"corroborated","false_positive_notes":"Coverage of AI systems, satire, fiction with a machine narrator and reference pages like this one all contain the phrase by necessity. Any post arguing about what assistants can and cannot do quotes it. Product copy for an assistant uses it as an ordinary self-description. Exclude quotations, dialogue and transcripts before counting a hit. The remaining case is a first-person machine self-description inside copy published under a human name, which is a review failure. It reveals nothing about how the surrounding sentences were produced, and the surrounding sentences are what a reader came for.","model_attribution":"Was common across assistant products in 2023 and is now suppressed by system prompts in most of them, which is why the sighting rate has fallen. A hit in 2026 output more often indicates an older model or a bare API call than any particular vendor.","platform_notes":[{"platform":"wikipedia","note":"One of the named G15 examples. Editors report it as the sign that needs no discussion, since no encyclopaedia article has a first person that could be a model."}],"languages":["en"],"sources":[{"kind":"external","title":"Wikipedia: Criteria for speedy deletion, criterion G15","url":"https://en.wikipedia.org/wiki/Wikipedia:Criteria_for_speedy_deletion","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"both-sides-affirmation","name":"Both-sides affirmation","aka":["social sycophancy","affirming both parties to a conflict"],"category":"model-register","subcategory":"sycophancy","description":"A response to a conflict validates every party in it and never says who was wrong. Your frustration is understandable, and so is theirs. Both perspectives have merit. The reply is agreeable and carries no judgment, which is the opposite of what advice is for. The shipped opener entry fires on a flattery string at the top of a text. This one reads the whole response and ignores the opening.","why_it_reads_ai":"The ELEPHANT benchmark measured social sycophancy across eleven models and reported them preserving the user face 45 percentage points more often than humans on personal-advice queries, and affirming both sides of a moral conflict in 48 percent of cases. Sharma and colleagues traced the mechanism at ICLR 2024: across a preference dataset of roughly 15,000 comparisons, matching the user beliefs was among the strongest predictors of which response a human rated higher. Agreeableness was rewarded, so agreeableness is what got trained in.","examples":[{"before":"Your frustration is completely valid, and your colleague's reaction is understandable too given the pressure the team is under. Both perspectives have merit here. Situations like this rarely have a clean villain, and most of the time everyone involved is doing their best with what they know. What matters is that the two of you keep talking.","after":"You skipped the handover note, so your colleague found out at 23:40 on 3 April from a customer email. Write the note before the next release. The midnight replies stop when the note exists.","note":"The repair names the omission, the time it surfaced and the change that fixes it. Figures in this repair are invented for the specimen."},{"before":"There are reasonable arguments on both sides of this, and each team is approaching it in good faith from their own priorities. Neither position is unreasonable once you understand where it comes from, and disagreements of this kind usually say more about how the work is organised than about the people doing it, which is worth holding in mind before anyone decides who was right.","after":"Support is right about the queue and engineering is wrong about the cause. The 40 tickets in July all came from one retry loop, and the fix is four lines in the webhook handler.","note":"A position, with the count that supports it. One side is named as wrong, which the affirming version avoided. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Read a response written about an interpersonal or moral conflict. Scan for an affirming clause about each party named, of the form that calls a position valid, understandable, reasonable or fair. Return a hit when every named party gets one and no sentence in the response states which action was wrong, what the writer would do, or what should change. Escape hatches, all returning clear: the text is framed as mediation, negotiation preparation or a summary of both cases; the writer states a position anywhere after the affirmations; the reader asked only for a summary of the dispute. Return the affirming clauses you found, the position sentence if one exists, and the verdict hit or clear."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active on two peer-reviewed and preprint measurements that agree: ELEPHANT reports both-sides affirmation in 48 percent of cases across eleven models and face preservation 45 percentage points above the human rate, and the ICLR 2024 sycophancy work shows preference data rewarding agreement with the user. Judged by rubric and not by string match, because the pattern is a stance across a whole response and no string carries it."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Mediators, therapists, HR investigators and diplomats affirm both parties as a professional method. For them it is the correct output, and reading it as evasion would be a mistake. Coaching and restorative-practice frameworks teach it explicitly. Public statements about a live dispute avoid taking sides for legal reasons that have nothing to do with sycophancy, and a manager writing to two reports at once may be required to. The judge returns clear whenever the genre is mediation or summary, and whenever a position appears anywhere after the affirmations. What counts is affirmation with no position at all, in a genre where the reader asked for one.","model_attribution":"Measured across eleven models from several vendors, so no single family owns it, and traced to preference training instead of architecture. Anthropic dated system prompts instruct against adjacent flattery habits, which evidences the underlying pull and not the behaviour of any deployed product. One vendor rolled back a 2025 model update for the same class of behaviour.","platform_notes":[{"platform":"linkedin","note":"The comment field is where this concentrates, because a conflict post attracts replies that want to agree with everyone in it. LinkedIn enforcement in this area targets automated commenting, not the stance itself."}],"languages":["en"],"sources":[{"kind":"external","title":"Cheng et al., ELEPHANT, social sycophancy benchmark (arXiv:2505.13995)","url":"https://arxiv.org/abs/2505.13995","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"Sharma et al., Towards Understanding Sycophancy in Language Models, ICLR 2024 (arXiv:2310.13548)","url":"https://arxiv.org/abs/2310.13548","accessed":"2026-08-14","tier":"peer-reviewed"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"measured-model-idiolect","name":"Measured model idiolect","aka":["five-way model classification","product fingerprint","per-model style signature"],"category":"model-register","subcategory":"idiolect","description":"Model families are separable from their output at corpus scale. A classifier trained on text alone told five chat products apart at 97.1 percent in an ICML 2025 study, and the signal survived rewriting and translation. That result is about distributions over many samples. It licenses nothing about the document in front of you, and this entry is in the directory to say so.","why_it_reads_ai":"Nothing here reads as a named product to a human reader. The finding is real and the popular version of it is not. Three limits travel with the measurement. It is aggregate, so a per-document call has no support in it. It is dated to the snapshot the authors collected, and vendor prompts move between releases. What it separates is products. The chat surface adds formatting that the same weights do not produce through an API: the OpenAI cookbook states that GPT-5 emits no Markdown by default there. A bullet-heavy layout therefore records where a text was generated rather than what generated it.","examples":[{"before":"Overall, there are several factors to consider when choosing a project tracker. First, consider your team size. Second, consider your reporting needs. Overall, the right choice depends on your workflow.","after":"We switched trackers in March 2026, when the sprint board passed 400 open cards and the weekly export started taking nine minutes. Card count decided it. Reporting was the tie-breaker and took ten minutes to check.","note":"The specimen shows the register that classifiers pick up in aggregate. It is not an identification of any product, and no reader should treat it as one. Figures in this repair are invented for the specimen."},{"before":"This response is concise and direct, and it avoids unnecessary formatting while covering the key considerations you raised about hiring.","after":"Hire the second candidate. She has shipped the exact integration twice, and our last two hires without that experience each took about five months to get productive.","note":"A different surface register, repaired the same way, by stating a decision and the evidence behind it. Figures in this repair are invented for the specimen."}],"detection":{"type":"statistical","metric":"model-family-classifier-accuracy-corpus-level","threshold_basis":"No document-level threshold exists, and this entry sets none on purpose. The published number is 97.1 percent five-way accuracy in Sun and colleagues, ICML 2025, from a machine classifier over a dated corpus with many samples per product. Turning a corpus-level classifier score into a per-document cutoff would invert what the study measured. Treat this entry as a boundary marker on the index rather than as a rule to run over a text."},"severity":"low","status":"contested","status_history":[{"date":"2026-08-15","status":"contested","rationale":"Shipped as contested on 2026-08-15. The underlying measurement is peer-reviewed and strong at corpus level. What is disputed is the inference people draw from it, that one document can be traced to a named product, and that inference has no support in the paper it cites. The entry exists to state the boundary. Snapshot-dated to the ICML 2025 collection, and the two vendor documents cited show the chat surface changing release by release."}],"evidence_grade":"peer-reviewed","false_positive_notes":"Plain-prose human writers land inside any of these profiles by coincidence, and the study measured a machine classifier, not a reader. A classifier score built over thousands of samples per product does not transfer to one paragraph, and a person has neither the corpus nor the training that the classifier had. Sparse formatting, short openings and analytic vocabulary are also what a careful human editor produces on a deadline. Treat any claim that a given text came from a named product as unsupported, including a claim made with this entry open. The honest use is the reverse of the popular one: it explains why the per-document version fails.","model_attribution":"The ICML 2025 study classified output from five chat products: ChatGPT, Claude, Grok, Gemini and DeepSeek. Reported tendencies differ by product, including sparser formatting in some and heavier emphasis markup in others, all snapshot-dated and all a property of the deployed surface as much as of the weights behind it. Anthropic dated system prompts and the OpenAI cookbook both document instructions that move these surfaces between releases, which is why per-product detail belongs in one dated entry instead of a page per vendor.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"Sun et al., Idiosyncrasies in Large Language Models, ICML 2025 (arXiv:2502.12150)","url":"https://arxiv.org/abs/2502.12150","accessed":"2026-08-14","tier":"peer-reviewed"},{"kind":"external","title":"Anthropic dated system prompts, release notes","url":"https://platform.claude.com/docs/en/release-notes/system-prompts","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"OpenAI GPT-5 prompting guide, cookbook","url":"https://developers.openai.com/cookbook/examples/gpt-5/gpt-5_prompting_guide","accessed":"2026-08-14","tier":"vendor"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"assistant-commit-register","name":"Assistant commit register","aka":["canned policy assurance","preserved existing behaviour","unquantified test-coverage claim"],"category":"model-register","subcategory":"commit-message","description":"Commit messages and edit summaries written in assistant register. The message assures the reviewer that conventions were followed, names the things that were not changed, and claims a quality it never measured. What it rarely contains is the reason for the change.","why_it_reads_ai":"The register comes from a model reporting back to whoever asked for the edit. Assurance is the natural closing move in that exchange and pure noise in a repository log, where a reader wants the reason and the blast radius. vale-ai-tells is the only taxonomy that covers this surface, with a dedicated rule family for commit messages that includes rules for unquantified claims, test enumeration and trailing justification. Wikipedia records the same shapes in edit summaries, where they show up more often than in articles.","examples":[{"before":"fix(auth): tighten the session check. All existing behaviour is preserved and the change adheres to the project conventions. All tests passing.","after":"fix(auth): reject sessions older than 12 hours. Sessions minted before the window now return 401 instead of silently refreshing, which is what let the 3 August incident run for six hours. Two integration tests cover the boundary either side of the cutoff.","note":"The assurances go. What replaces them is the reason, the behaviour change and the incident that prompted it. Figures in this repair are invented for the specimen."},{"before":"chore(api): rename the handler. Functionality remains unchanged and all tests are green.","after":"chore(api): rename fetchUser to loadUserProfile. The old name collided with the hook added in the June refactor and cost two people an afternoon of debugging on 9 July. Callers updated in the same commit, none outside this package.","note":"A rename still needs a reason and a blast radius, and both fit in two sentences. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:all\\s+)?(?:existing\\s+)?(?:behaviou?r|functionality|logic)\\s+(?:is|was|remains|has\\s+been)\\s+(?:preserved|unchanged|maintained|untouched)\\b|\\b(?:adheres?\\s+to|in\\s+(?:full\\s+)?(?:compliance|accordance)\\s+with|follows)\\s+(?:the\\s+)?(?:project|repository|repo|coding|style|contribution)\\s+(?:guidelines|conventions|standards|rules)\\b|\\ball\\s+tests\\s+(?:are\\s+)?(?:pass(?:ing|ed)|green)\\b","flags":"i","scope":"document"},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active but graded community-observed, because vale-ai-tells is the only taxonomy that covers commit messages as a surface and no measurement of the register exists. Wikipedia edit summaries carry the same shapes, which is corroboration of the pattern, not of a rate. Pattern verified on 2026-08-15 against both specimens and against three ordinary commit sentences, including one that documents a legacy column left in place on purpose."}],"evidence_grade":"community-observed","false_positive_notes":"Careful committers document invariants deliberately, and several teams require a line about what was left alone, which produces the same sentence for a good reason. Regulated codebases require a compliance note in the message. Junior engineers are taught to write the summary before the reason. Repository templates generate the assurance automatically, which puts it in messages nobody typed. Read the diff before counting a hit. The tell is an assurance the diff does not support, or a quality claim with no number attached, in a message that never says why the change was needed.","model_attribution":"No product owns this register. It follows from the assistant framing of the task, not from any one vendor, and it appears wherever a model is asked to write the message for an edit it just made.","platform_notes":[{"platform":"wikipedia","note":"Edit summaries are the closest public analogue to a commit message, and the same assurance shapes appear there. The signs guide files them with communication intended for the user, not with the prose tells."}],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Wikipedia: Signs of AI writing","url":"https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing","accessed":"2026-08-14","tier":"community"}],"added":"2026-08-15","updated":"2026-08-15"},{"id":"translationese-formal-register","name":"Translationese formal register","aka":["formal register in a casual context","non-English output that reads as translated"],"category":"model-register","subcategory":"non-english-register","description":"Non-English output arrives in the highest formal register available, whatever the context asks for. Formal pronouns in every sentence, verb phrases turned into noun phrases, no discourse particles anywhere. It reads as a translation of a formal English source. Often it is one.","why_it_reads_ai":"The post-2022 lexical shift shows up outside English. A preprint covering 34 languages, committed to EMNLP 2026, reports uptake in 26 of them with a mean prevalence increase of 15.1 percent. What that work does not supply is a shape-level list per language. The only worked non-English taxonomy we hold is for Bahasa Indonesia, and it comes from a system prompt written to steer models away from these habits, which makes it inverted evidence of the habits. Two of its observations any speaker can test: the formal second-person pronoun Anda held in every register, and the absence of the particles that ordinary Indonesian carries. No equivalent list exists yet for Finnish, German, French or the other languages the multilingual measurement covers. Inventing one would be worse than naming the gap.","examples":[{"before":"Anda dapat melakukan pengecekan terhadap status pesanan Anda melalui halaman berikut. Kami senantiasa berupaya memberikan pelayanan terbaik bagi seluruh pelanggan kami. Kepuasan Anda merupakan prioritas utama bagi perusahaan kami, dan hal tersebut menjadi dasar dari setiap perbaikan yang kami lakukan. Kami mengucapkan terima kasih atas kesabaran dan pengertian Anda.","after":"Kamu bisa cek status pesanan di halaman ini. Nomor pesanan ada di email konfirmasi, biasanya masuk 10 menit setelah bayar.","note":"Indonesian, customer message. The formal version nominalises the verb and holds Anda; the repair drops to the register a customer writes in and adds where the order number is and when the email lands. Figures in this repair are invented for the specimen."},{"before":"Pelaksanaan pembaruan sistem akan dilakukan pada hari Senin. Pembaruan tersebut merupakan bagian dari komitmen kami untuk terus meningkatkan kualitas layanan bagi seluruh pengguna. Kami berharap seluruh pengguna dapat memaklumi kondisi tersebut. Kami mohon pengertian Anda atas ketidaknyamanan yang mungkin terjadi.","after":"Sistem kami update hari Senin, jam 2 sampai jam 4 pagi. Selama itu kamu masih bisa lihat pesanan lama, tapi belum bisa checkout.","note":"Same language, maintenance notice. The repair states the window and what still works during it. Figures in this repair are invented for the specimen."}],"detection":{"type":"judge","rubric":"Read the text in its own language. Decide the register it is written in, then the register its context calls for. Return a hit only when the text sits at the top formal register while the context is casual, such as a social post, a reply, or a message to a customer. Markers to scan for, with Indonesian as the worked case: the formal second-person pronoun used in every sentence with no shift, verb phrases converted into noun phrases where a plain verb exists, and no discourse particles anywhere in a conversational passage. Escape hatches, all returning clear: government, legal, academic and institutional writing, which are formal by requirement; a language or genre with no casual written form; an audience the writer addresses formally by convention. Return the register you judged, the context you judged it against, and one span that shows the mismatch. If you do not read the language, return nothing."},"severity":"medium","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active and graded community-observed on 2026-08-15. The multilingual measurement establishes that the shift is not English-only, at 26 of 34 languages, and it is a preprint committed to EMNLP 2026. The shape-level detail comes from a single steering prompt for Bahasa Indonesia, carried as inverted evidence and never as sole support. No rate has been published for the register itself, so the entry claims a pattern and not a prevalence."}],"evidence_grade":"community-observed","false_positive_notes":"Civil servants, lawyers and institutional communications staff write the top register in every language they publish in, and that is the correct output for their genre. Second-language writers are taught the formal form first and many hold it, which is exactly the population that detection tooling already fails hardest on. Older writers in several languages read informal public address as rude, and some genres have no casual written form at all. The judgment has to be made by somebody who reads the language, against the real audience. A reviewer who does not read it should return nothing instead of guessing, and a translation into English destroys the evidence before the question is even asked.","model_attribution":"No product can be named from it. The observation set we hold describes models writing Indonesian in general, and the source that supplies it is a steering prompt and not a measurement, so product-level attribution is unavailable and should not be inferred.","platform_notes":[],"languages":["id"],"sources":[{"kind":"external","title":"Juzek, LLM lexical uptake across 34 languages, preprint (arXiv:2605.25358)","url":"https://arxiv.org/abs/2605.25358","accessed":"2026-08-14","tier":"primary-doc"},{"kind":"external","title":"anti-slop-writing system prompt and pattern list","url":"https://github.com/adenaufal/anti-slop-writing","accessed":"2026-08-14","tier":"community","source_class":"inverted-evasion-prompt"}],"added":"2026-08-15","updated":"2026-08-17"},{"id":"redundant-precaution","name":"Redundant precaution","aka":["consult a professional","this is not advice","unsolicited disclaimer"],"category":"model-register","subcategory":"assistant-leak","description":"A disclaimer attached to material that carries no risk. A note about numbering workshop bins ends by advising the reader to consult a professional. In medicine and finance the precaution is real and required. Here it is a reflex firing on any text shaped like advice.","why_it_reads_ai":"Liability language is the trained closing move for anything advice-shaped, and it keys on shape, not on subject. The tellsign list files the phrasing as a generic liability closer appended to advice-adjacent answers. Anthropic dated system prompts document the same class of habit being instructed away, which is first-party evidence of the pull, and not of what any product now emits. Where regulation requires a disclaimer it belongs in the text, and that case sits outside this entry. It is not a false positive.","examples":[{"before":"Numbering the workshop bins by shelf works well for most setups. As always, consult a professional before applying this to your own setup.","after":"We number the workshop bins by shelf instead of by tool. The scheme broke in March 2026 when we added a second rack, so every number now starts with the rack letter.","note":"Nothing about bin labels carries risk. The repair puts the failure case where the disclaimer was. Figures in this repair are invented for the specimen."},{"before":"Pick the notebook app your team already opens every morning, and the rest takes care of itself. This does not constitute professional advice.","after":"Pick the notebook app your team already opens every morning. Ours was the one already installed on 30 of 34 laptops, which ended a two-month evaluation in an afternoon.","note":"The disclaimer is replaced by the number that made the decision. Figures in this repair are invented for the specimen."}],"detection":{"type":"deterministic","pattern":"\\b(?:please\\s+)?(?:consult|speak\\s+to|check\\s+with)\\s+(?:a|your)\\s+(?:qualified\\s+|licensed\\s+|certified\\s+|medical\\s+|legal\\s+|financial\\s+)?(?:professional|physician|attorney|advisor|adviser)\\b|\\bthis\\s+(?:is\\s+not|does\\s+not\\s+constitute)\\s+(?:legal|medical|financial|professional|investment)\\s+advice\\b","flags":"i","scope":"sentence"},"severity":"low","status":"active","status_history":[{"date":"2026-08-15","status":"active","rationale":"Active and graded community-observed on 2026-08-15. The tellsign list names the closer directly, and Anthropic dated system prompts show the neighbouring habit being suppressed by instruction, which supports the pull without establishing a rate. Severity is low because the regulated case is common and legitimate, and because the repair is a deletion, not a rewrite. Pattern verified on 2026-08-15 against both specimens and against three ordinary sentences, including one about an insurer that mandates a dosing disclaimer."}],"evidence_grade":"community-observed","false_positive_notes":"Regulated professions attach disclaimers because compliance requires them. Health, legal and financial writers have no choice, insurers mandate the wording, and in those genres a missing disclaimer is the actual failure. Publishers append a house disclaimer to every article by template, which puts the phrase in copy the writer never touched. Safety documentation for tools and chemicals carries it for good reason. Judge by subject and by obligation: a disclaimer on a topic with no bodily, legal or financial exposure, in a publication with no house rule requiring one, and with no named risk anywhere in the surrounding text.","model_attribution":"Common across assistant products. No single vendor owns it. Anthropic publishes dated system prompts that instruct against adjacent hedging habits, which documents the tendency at one vendor and says nothing about the rate at any other.","platform_notes":[],"languages":["en"],"sources":[{"kind":"external","title":"vale-ai-tells, 111 machine-checkable rules","url":"https://github.com/tbhb/vale-ai-tells","accessed":"2026-08-14","tier":"community"},{"kind":"external","title":"Anthropic dated system prompts, release notes","url":"https://platform.claude.com/docs/en/release-notes/system-prompts","accessed":"2026-08-14","tier":"vendor"}],"added":"2026-08-15","updated":"2026-08-17"}]}