Wikipedia AI Enforcement: Every Published Number on Detection and Deletion (2026)

1 How Many Wikipedia Pages Have Been Tagged As AI-Written?

The shape of that series is the most useful thing on this page, and it is also the easiest thing to misread. A rise in tagging is a rise in tagging. It tells you that more volunteers looked at more pages and applied a template, and it cannot separate that from a rise in how much machine text was actually submitted. The dip to the low two hundreds in the late spring followed by a jump back above six hundred in the early summer is almost certainly a story about backlog drives and editor attention rather than about anything the models did.

Set the totals against the size of the encyclopedia and the scale becomes clearer in the other direction. 7.1 million articles exist in English. (The Guardian, 2026) Nineteen months of tagging has touched a few thousand of them. Whatever is happening, the constraint is not the volume of AI text arriving, it is the number of humans available to look at it, which is the same ceiling that shows up in the broader numbers on how much content detection tools are scanning across the industry.

There is one more way to read the series, and it is the reading the editors themselves seem to work from. Treat the monthly figure as a measure of capacity rather than of contamination. On that reading the encyclopedia discovered in the late summer of last year that it could process somewhere between three and six hundred suspect pages a month, pushed past a thousand once, and has otherwise hovered in that band for a year. Nothing in the published data suggests the band is set by how much machine text arrives. Everything suggests it is set by how many people show up to look.

2 Wikipedia Will Not Delete An Article Because Its Author Admits Using AI

Read the criterion closely and it is a deliberate piece of engineering. The first sign is the model talking to its operator, phrases such as a cheerful offer of your Wikipedia article or a note about a training cutoff. The second is citations that fail on inspection: dead on arrival links, invalid checksums, a digital object identifier that resolves to a paper about beetles in an article about computing. The third is technical residue, the stray markup and token artifacts that leak out of a chat interface. Each one is checkable by a second editor without anyone having to agree about style.

Then comes the sentence that makes this page worth writing. The criterion says the other recognised signs of machine writing are not enough on their own, and it puts an author’s own confession in that same excluded category. A contributor can say outright that a model wrote the article and that still does not satisfy the rule. The reason is consistency: a confession is not an artifact anyone can verify from the page, and a criterion that admits unverifiable evidence invites exactly the disputes that the question of whether detector output counts as proof produces in universities. The encyclopedia chose a narrower rule it could defend over a broader one it could not.

That design choice has a cost the rule accepts openly. A page can be written end to end by a language model, read like it, and survive, as long as whoever pasted it cleaned up the artifacts and the citations hold. The criterion catches carelessness rather than machine authorship, and the editors who wrote it knew that. What they bought in exchange is a deletion that holds up when it is challenged, because the evidence sits on the page where a second person can see it. Most institutions facing this problem have made the opposite trade, accepting a score from a tool and discovering afterwards how hard it is to defend.

3 The Ban Passed 44 To 2, And Two Broader Proposals Failed First

A tally that lopsided looks like an easy decision, and the history says it was anything but. Willemien Froneman’s process trace in AI and Society, the first full account of how this policy came together, describes a community that agreed early and unanimously that machine text was a threat and then failed twice to write a rule about it. (Froneman, AI & SOCIETY, 2026) The second attempt was closed with a finding of no consensus to adopt any wording as either policy or guideline. Her argument is that comprehensive proposals generate too many points of disagreement to survive a consensus process, and that governance succeeds only when it narrows.

What the editors said while failing is more revealing than the final tally. One argued that an early proposal put the cart before the horse, enabling summary removal for a violation without giving editors any way to determine whether it had happened and without recourse for innocent victims of bad judgment. Another dismissed AI detectors as having no source code, no model weights and nothing besides marketing copy. A third concluded there is ultimately no oracle machine that could perfectly distinguish AI text from text written by a person. Those three objections are why the rule that eventually passed looks the way it does, and they are a sharper statement of the problem than most of the published work on how far detectors disagree with each other manages.

4 What Do The Published Detector Measurements On Wikipedia Actually Show?

The 2024 study, by Creston Brooks, Samuel Eggert and Denis Peskoff, is careful in a way the headlines about it were not. The authors ran GPTZero and the open source Binoculars over new pages, calibrated both to a 1% error rate against articles written before the release of the model that started all this, and reported the result as a floor rather than an estimate. It predates the deletion criterion and the ban by more than a year, so it describes a period with no enforcement at all. Treat it as history. The same caution applies to anyone citing it alongside the figures on AI generated research papers, where the detection baseline is also moving under the measurement.

Wiki Education’s numbers are the more interesting half, because the organisation ran a detector across a corpus whose authorship it already knew. Of 3,078 articles produced through its programs, 178 came back flagged. (Wiki Education, 2026) More than two thirds of those flagged articles failed source verification in some form, yet only 7% contained outright fabricated references, which suggests the flag was catching sloppy sourcing at least as often as machine authorship. The organisation also published where its detector misfired: on bibliographies and outlines, pages heavy with non-prose content, which forced a change to how text was prepared before scanning. A vendor rarely volunteers that. It is the same failure mode that turns up in independent testing of that detector and in flagging of neurodivergent writing.

Put the two measurements next to each other and they answer different questions, which is why neither should be quoted as the AI rate on Wikipedia. The 2024 study sampled new pages from an unknown population of authors and asked how many look machine written to a calibrated detector. Wiki Education sampled a known population of supervised students and asked the same question of a different detector. The first is a floor on a population nobody controls. The second is a rate within a group that was being taught and watched, which is roughly the best case, and it still flagged articles at a similar order of magnitude. The honest summary is that both numbers are small, both are uncertain, and the gap between what a detector flags and what a person actually generated is wide enough in both studies that the authors said so themselves.

5 What These Numbers Cannot Tell You

The outcome gap is the serious one. Tagging is public, the noticeboard is public, the deletion queue is public at any given moment, and none of it is aggregated into a figure for how many tagged pages were actually deleted, how many were contested, and how many came back. That means nobody can currently measure the thing that matters most about this regime, which is its error rate. A system built explicitly to avoid false positives does not publish the statistic that would show whether it succeeded.

The language gap compounds it. 3 other editions appeared in the 2024 study with lower flag rates than English, and lower flag rates in a detector-based study can mean less machine text or simply a detector that works less well, a confound explored in the accuracy gap between English and other languages and visible again in the scale of false positives among non-native English writers. Everything else on this page is English Wikipedia alone. Institutions writing their own rules, from the encyclopedia to the detection policies at major universities, are working from evidence that is thinner and more English-centric than the confidence of the policies suggests.