# Internal factual contradiction 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: internal-factual-contradiction - Category: Semantic (semantic) - Subcategory: factual-failure - Also known as: conflicting facts in one piece, confident instruction carrying wrong facts - Status: Active. Currently signals low-effort writing. - Severity: high - Evidence grade: feedsquad-observed (our own corpus only, no external attestation) - Languages: en - Added: 2026-08-15 - Updated: 2026-08-15 - Page: https://feedsquad.com/ai-tells/internal-factual-contradiction ## 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 as machine-written 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. ## Detection Type: judge (a rubric for a lightweight model judge) Rubric for a lightweight model judge: ```text Task: find claims inside the document that cannot both be true. Step 1. Build a fact table. One row per atomic claim, with subject, attribute, value and any date. Step 2. Group rows sharing a subject and attribute. Step 3. Compare within each group for three conflict types: (a) value conflict, where one quantity is given two incompatible numbers; (b) ordering conflict, where an event is placed both before and after another; (c) definition conflict, where a term is used with two incompatible meanings. Escape 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. Decision. FLAG on one unexplained conflict. One is enough, because a reader cannot use the document without knowing which half to believe. Report both rows. Output: the conflicting pair, the conflict type, and FLAG or PASS. ``` ## Examples Constructed specimens. Written for this index. Never quoted from anyone's posts. ### Specimen 1 Before, exhibiting the tell: > 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, repaired: > 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. ## False positives Who legitimately writes this way. 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 - youtube: 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. ## Sources 1. Fairplay coalition open letter to YouTube on AI content aimed at children (Mar 2026) https://fairplayforkids.org/wp-content/uploads/2026/03/YouTube-Letter-AI-Slop.pdf (tier: primary-doc; accessed 2026-08-14) 2. Tubefilter, YouTube and the Fairplay kids AI open letter https://www.tubefilter.com/2026/04/01/youtube-fairplay-kids-ai-open-letter/ (tier: press; accessed 2026-08-14) ## Status history Ids are permanent. A retired tell keeps its id and its page. - 2026-08-15, Active: 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. ## 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.