# Excess-vocabulary cluster (delve, underscore, intricate) Part of [The AI Tells Index](https://feedsquad.com/ai-tells). A tell signals low effort. It does not identify an author. Skilled writers produce every shape listed here, some of them daily, and automated detectors misread those writers for it at rates measured above 60 percent on non-native English prose. Nothing in this index proves that a machine wrote anything. Read an entry as one piece of evidence to weigh against the false-positive notes printed beside it. ## Facts - Id: delve-excess-vocabulary - Category: Lexical (lexical) - Subcategory: excess-vocabulary - Also known as: 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 - Status: Fading. Signal is weakening, usually because model vendors trained the habit out. - Severity: medium - Evidence grade: peer-reviewed (a published study measures the pattern) - Languages: en - Added: 2026-08-14 - Updated: 2026-08-17 - Page: https://feedsquad.com/ai-tells/delve-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 as machine-written 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. ## Detection Type: statistical (a measured metric against a threshold with a stated basis) Metric: distinct-excess-vocabulary-markers-per-1000-words Threshold: 3 (fires when 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. ## Examples Constructed specimens. Written for this index. Never quoted from anyone's posts. ### Specimen 1 Before, exhibiting the tell: > 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, repaired: > 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. ## False positives Who legitimately writes this way. 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 - wikipedia: 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. - linkedin: 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. ## Sources 1. Kobak, Gonzalez-Marquez, Horvat, Lause: Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Science Advances 11(27) https://www.science.org/doi/10.1126/sciadv.adt3813 (tier: peer-reviewed; accessed 2026-08-14) 2. Same study, arXiv HTML v1 (excess frequency ratios and gaps) https://arxiv.org/html/2406.07016v1 (tier: primary-doc; accessed 2026-08-14) 3. Kousha and Thelwall: How much are LLMs changing the language of academic papers after ChatGPT? https://arxiv.org/abs/2509.09596 (tier: primary-doc; accessed 2026-08-14) 4. Yakura et al.: Empirical evidence of Large Language Model's influence on human spoken communication https://arxiv.org/abs/2409.01754 (tier: primary-doc; accessed 2026-08-14) 5. Merrill, Chen, Kumer: What are the clues that ChatGPT wrote something? We analyzed its style, Washington Post https://www.washingtonpost.com/technology/interactive/2025/how-detect-chatgpt-em-dash/ (tier: press; accessed 2026-08-14) 6. Liang, Yuksekgonul, Mao, Wu, Zou: GPT detectors are biased against non-native English writers, Patterns 4:100779 (PubMed Central copy) https://pmc.ncbi.nlm.nih.gov/articles/PMC10382961/ (tier: peer-reviewed; accessed 2026-08-14) 7. Vanguard Nigeria: Nigerians tackle American author for claiming delve is only used by ChatGPT https://www.vanguardngr.com/2024/04/nigerians-tackle-american-author-for-claiming-delve-is-only-used-by-chatgpt/ (tier: press; accessed 2026-08-14) 8. Wikipedia: Signs of AI writing https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing (tier: community; accessed 2026-08-14) 9. Juzek and Ward: Why Does ChatGPT Delve So Much? (COLING 2025) https://arxiv.org/abs/2412.11385 (tier: peer-reviewed; accessed 2026-08-14) 10. Geng and Trotta: coevolution of human and LLM writing (Findings of ACL 2025) https://aclanthology.org/2025.findings-acl.657/ (tier: peer-reviewed; accessed 2026-08-14) 11. LexA-Index, CC0 per-language overuse dataset https://github.com/fsu-nlp/lexa-index (tier: primary-doc; accessed 2026-08-14) ## Status history Ids are permanent. A retired tell keeps its id and its page. - 2026-08-14, Fading: 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. ## License CC BY 4.0. https://creativecommons.org/licenses/by/4.0/ Attribution: The AI Tells Index, feedsquad.com/ai-tells Reuse the data, including commercially. Keep the attribution line. --- Dataset version 1.0.0. Schema version 1. Part of [FeedSquad](https://feedsquad.com). Built by [Herman Foundry](https://hermanfoundry.com) from Levi, Finnish Lapland.