Wikipedia editors say 20 percent of new articles were flagged as AI

A user who writes five long Wikipedia articles in five minutes is a reason to suspect AI. That is how one editor described the current test to IEEE Spectrum. The volunteer group that runs the check is called WikiProject AI Cleanup. It was founded in 2023 by Ellia Lebleu and other editors. The group started on Wikipedia's Discord server. IEEE Spectrum published a long account of the project on October 21. A Hacker News discussion of the piece picked up 398 points this week.

What the volunteers used to look for

The old signals were about words. IEEE Spectrum names two example phrases. One is "stands as a testament." The other is "rich cultural heritage." The old signals also included odd section headings. The piece names "Why is this important?" as one of those. A conclusion paragraph that read as a summary was another. Those phrases and that shape pointed to AI-generated articles with no further checking. The report says this method worked in the first phase.

By 2025, the signals shifted. The piece names "the rule of three" and negative parallelisms as later linguistic patterns volunteers watch for. The project also adopted a behaviour-first approach. IEEE Spectrum reports an editor's own example. If a user writes five long articles in five minutes, that is a reason to suspect AI. Volume and speed became the stronger signal than style.

Numbers and a bigger problem

IEEE Spectrum reports that AI-detection tools flagged about 20 percent of new Wikipedia articles and drafts in late 2025. The flagged share fell to 15 to 16 percent by early 2026. The report does not name a cause for the drop. These are figures from the detection tools the project uses, not an independent audit.

The piece also names a deeper problem than fake writing. AI often produces citations that look right but misrepresent the source. IEEE Spectrum writes that an AI system more often finds a real source and gets its content wrong. That happens more often than inventing a reference from nothing. A citation that points at a real article but misrepresents it is harder to catch. A citation that points at nothing is easier to find.

Why this matters outside Wikipedia

A software team that reads user submissions can borrow the same test. A code review, a bug report, a lead form, a job application: in each case the behaviour of the submission is a signal the content is not. Five long submissions in five minutes from one account is a reason to check. The writing can look however the model wants it to. The old word-pattern rules will keep breaking as model output gets more varied. The behaviour rule holds longer. It tests the thing the model cannot cheaply fake. That is the time it takes a person to produce the work.

Source

Wikipedia's Cleanup of AI-Generated Content, IEEE Spectrum.