The AI writing tells

What does Wikipedia's catalogue say about itself?

Four statements, all on the page, all easy to check, and all routinely dropped by the pages that cite it.

• It is descriptive, not prescriptive. In its words: it consists of observations, not rules.
• The signs are not the problem. It says in bold that the patterns are only potential signs of a problem rather than the problem itself, and asks readers not to treat the signs as the things to fix.
• It expects false positives by design. It notes that not all text with these markers is machine-written, because the models were trained on human writing, including Wikipedia itself.
• It is scoped to Wikipedia. It describes itself as a field guide for spotting undisclosed AI content on Wikipedia, and says only some of the signs are broadly applicable.
• It is an advice page of WikiProject AI Cleanup and not Wikipedia policy, because it has not been reviewed by the community for consensus.

None of that makes it less useful. It changes what kind of tool it is. It describes what machine writing has tended to look like. It is not a test you can run on a colleague's draft.

Why does the list keep expiring?

Because the tells are model artefacts, and models change. The catalogue dates its own vocabulary by model era and the shape of that is the most useful thing on it.

Eighteen down to four. The catalogue also records the fate of the most famous one: “delve” was heavily overused in 2023 and early 2024, became less frequent later in 2024, then dropped off sharply in 2025.

So a list of AI words has a half-life, and most of the lists in circulation are reporting the vocabulary of a model nobody uses any more. The page itself agrees: as of August 2026 it has a maintenance banner asking for its coverage of the most recent models to be updated.

Remember that before running any of these words as a checker. A rule that flags “delve” in 2026 is testing for a model that has been superseded twice.

Follow the em dash, because people name it first and it has the least evidence behind it: whether the em dash is evidence of anything is a page of its own. The question underneath the whole list, why so much of this output converges on one register, has a published answer that does not expire with the vocabulary.

Which of these have evidence behind them?

Unevenly and the page is honest about it if you look. “Delve” has four separate citations to published work. Other entries have none: “key”, as an adjective, is tagged with a citation-needed marker.

Somebody measured “delve”. Somebody only noticed “key”. Both can be true. Only one is evidence and a list that presents them in the same typeface is inviting you to treat them as equally solid.

If you take anything from this page into your own style rules, take the cited ones and leave the rest as observations.

The same standard is worth turning on the tools that claim to settle the question from the other end. Detector accuracy claims, read against the disclaimers on the same page come out thinner than anything above, which is why the control belongs earlier than the finished draft.

Which of these can a checker catch, and which cannot?

A word list is trivially checkable. A construction usually is not and the difference decides what can go in a style file at all.

Most of the problem is in the bottom two rows, and no linter will help with either. That is not a reason to skip the top rows. It is a reason not to believe a green run means the writing is good.

The rules that survive into a document are the mechanical ones. The download includes a style checker configured with them.

Before you write the rules, know what a checker can and cannot be asked to do: which parts of a voice map onto linter rules covers the mapping, and why no published package is a brand voice on its own.

Why is a tell not proof?

Because the base rate defeats it. These constructions appear in human writing constantly, they appeared there first and the models learned them from us. A marker that is common in both populations cannot separate them, however strongly it feels like a signal.

The catalogue says as much, which is why the honest use of it is to improve your own writing rather than to judge somebody else's. This site will not claim a file makes writing undetectable and it will not sell you the reverse either.

What should you do about it?

Write the prohibitions down and let a tool enforce them, rather than trying to recognise a vibe after the fact. A banned-words list you chose is worth more than a tells list somebody else observed, because it encodes your judgement instead of a model's habits.

Answer the questions and the tool builds that list, with a checker already configured. Your answers stay in your browser unless you ask for a share link.

If you arrived wanting the assistant itself to sound like you rather than wanting to spot when it does not, why asking a model to write like you trains nothing is the page that takes that apart and it ends in the same place this one does.

Where these come from

Every claim above is quoted from one of these, and each was read on the date beside it. If one of them has changed since, the page is wrong and we would like to know.

• Wikipedia: Signs of AI writing and its own caveats checked 2026-08-17
• Wikipedia: WikiProject AI Cleanup checked 2026-08-17