Wikipedia AI policy: how editors remove AI text

Wikipedia banned AI-written article text in March 2026

Since March 2026, English Wikipedia's AI policy forbids using ChatGPT, Claude, or any other LLM to write or rewrite article content, and editors who keep doing it lose the ability to edit. The rule lives in a guideline called Writing articles with large language models, usually shortened to WP:LLM. An LLM, or large language model, is the kind of AI behind chatbots like ChatGPT.

The guideline names ChatGPT, Claude, Copilot, DeepSeek, Gemini, Grammarly, and Grok as examples. Grammarly is on the list even though most people think of it as a grammar checker rather than a chatbot.

Two exceptions: simple copyedits and translation

An LLM may only suggest fixes for simple mistakes in your own writing or help translate an article from another language's Wikipedia, and a person has to review the result either way. The copyediting exception covers spelling, punctuation, capitalization, grammar, and similar small errors. After any such fix, the text still has to match what the cited sources say.

Translation is the second exception, and it comes with its own guidance page on LLM-assisted translation. The guideline warns that a poorly checked machine translation tends to break the same rules as text generated from scratch.

Editors who keep using AI get blocked

On Wikipedia, a block depends on the full pattern of an editor's recent edits and whether they meet core content policies, not on whether the writing merely sounds like a chatbot. Repeated LLM use outside the two exceptions counts as disruptive editing and ends in a block, a measure that stops an account from editing. That consequence was added through a community discussion that ran from March to May 2026.

Some people simply write in a style that resembles chatbot output. That's why a hunch about tone or word choice isn't enough to sanction anyone.

WikiProject AI Cleanup hunts for AI text

Volunteers have worked through AI-written pages since 2023, and 324 cases were still open on November 2, 2026, with text added from 2022 to 2026 still on the site. WikiProject AI Cleanup is the group of editors doing that work, and since the ban came years after the first chatbot text appeared, much of the job is cleaning up older edits. The project lists five goals:

• Find and remove AI-written text that breaks WP:LLM
• Find AI-generated images and make sure they are used appropriately
• Keep track of editors who use AI without seeing its weaknesses as a writing tool
• Teach editors about the problems with AI text and what English Wikipedia expects
• Protect the site from disruption by AI agents (related: Rogue AI Agents Tied to OpenAI Hit Wikimedia Sites)

It started with the same odd phrases in December 2023

WikiProject AI Cleanup began when several editors realized they had all run into strange articles full of the same phrases, such as "stands as a testament." Founding member Ellia Lebleu told IEEE Spectrum that the moment came around December 2023, in a conversation on Wikipedia's Discord server. The articles also leaned on lines like "rich cultural heritage" and ended with sections titled "Conclusion" or "Why is this important?"

That style was typical of ChatGPT at the time and looked out of place on Wikipedia, which made the pages easy to catch. Lebleu set up an official project so more editors could help, and the earliest signature on the participant list is dated December 4, 2023. Lebleu is now working toward a Ph.D. in AI at the Inria Center of the University of Rennes in France.

Suspect articles get a notice, and real sources are kept

WikiProject AI Cleanup volunteers tag suspect articles with a notice, remove the text that breaks WP:LLM, warn the user who added it, and save any genuine sources for a human rewrite. The notice is the AI-generated template at the top of the article, and it tells readers the page hasn't been fully cleaned up yet.

AI text sometimes cites sources that really exist. When that happens, editors move those sources to the article's talk page, the discussion page attached to every article, so a person can use them later.

The signs of AI writing Wikipedia editors look for

Obvious stock phrases are fading, so editors now weigh sentence structure, what the sources actually say, and how fast someone writes, and no single sign proves anything. The editors keep their observations on a page called Signs of AI writing. It collects real examples from Wikipedia articles, drafts, and comments, and it describes patterns rather than setting rules.

From stock phrases to sentence structure

Since about 2025, Wikipedia editors look for patterns that mean little alone but add up, such as listing things in threes or "not just X, but Y" constructions. In 2023 and early 2024, AI articles gave themselves away with lines like "rich cultural heritage" and tacked-on conclusion sections, but by 2025, Lebleu says, the giveaways had moved from vocabulary into sentence structure. As of October 2026, the guide lists these as the most common signs in new AI text:

• Markdown formatting, such as asterisks for bold, pasted into the article
• Edit summaries that were themselves written by AI
• Hedging about whether a source is available or how it was used
• Vague claims that two things are connected
• Boilerplate praise of sources and media coverage

Citations that don't back the claim, and impossibly fast writing

The most common problem on Wikipedia now is a real source that doesn't say what the article claims, and writing five long articles in five minutes is also a red flag. Chatbots can now look up sources, so completely invented citations have become rare. The typical failure today is a genuine book or paper that, once opened, says something different from the sentence it supports.

Editors also look past the article itself. Someone who adds five articles of several thousand words each within five minutes is almost certainly not typing them by hand.

A sign is a clue, not proof

People write this way too because chatbots learned from human text, but anything that was already in an article before ChatGPT launched on November 30, 2022 is very unlikely to be AI-generated. The Signs of AI writing guide says so up front. Many of the same habits turn up in editorials, blogs, and fan fiction.

That is why the ChatGPT launch date works as a practical cutoff. A sentence that predates it can be ruled out, while newer text has to be judged on its sources.

Obvious AI pages are deleted without discussion

Only pages with traces that a person would almost never leave can skip the usual deletion discussion. Speedy deletion lets an administrator remove a page right away instead of waiting for other editors to weigh in. The criterion for AI pages is G15, "Unambiguously LLM-generated pages."

Three signs that qualify for speedy deletion

A Wikipedia administrator can delete a page on sight if it contains a chatbot talking to its user, a reference that doesn't exist, or a technical trace such as "oaicite." One of the three is enough. Each is something a human writer would almost never produce.

The three G15 signs for speedy deletion

Sign	Examples
Talking to the user	"Here is your Wikipedia article on…", knowledge-cutoff disclaimers, "as a large language model", unfilled placeholders like "[Birth Date]"
References that don't exist	Dead links, ISBNs with bad checksums, unrelated DOIs, news reports dated before the event
Technical traces	Model citation bugs such as oaicite or turn0search0, wikitext code blocks

Source: English Wikipedia, Criteria for speedy deletion, G15.

Broken citations can also come from typos or link rot. So the examples stick to errors that a typo or a dead link can't easily explain.

Everything else goes to a deletion discussion

Style quirks, a confession, or a high detector score are not enough for speedy deletion, so those pages go through proposed deletion or a deletion discussion. The common signs alone don't trigger G15, and neither does an author admitting to using an LLM.

Those pages can be proposed for deletion through LLMPROD, nominated at Articles for deletion, or raised on the AI noticeboard. The Signs of AI writing guide adds that a high percentage from a detector does not meet G15 either.

Detection tools help, but editors make the call

Tools sort through large volumes of edits, but a person who has checked the sources decides whether text is AI-generated. WikiProject AI Cleanup uses a detector to estimate how much AI text is coming in. Other tools narrow down which edits an editor needs to open (related: AI Image Detection: 3 Free Tools to Spot AI Photos).

Pangram estimates 15–16% of new articles are AI-generated

The share of new articles flagged as AI fell from about 20% in late 2025, and nobody knows yet whether the policy worked or the detector is falling behind newer models. Lebleu works with the AI detection company Pangram to measure how many new articles and drafts are AI-generated. The early 2026 figure was 15 or 16 percent.

Before relying on it, Lebleu ran Pangram on Wikipedia text written before ChatGPT existed and got zero false positives. A tool that rarely flags human writing tends to miss some AI text instead, so Lebleu treats the figure as a very conservative lower bound.

A source checker and Paste Check lighten the load

An AI that checks sources against claims and a prompt that asks whether pasted text is your own catch problems before an editor has to look. Another Wikipedia editor built the source checker, which opens each citation on a page and compares it with the sentence it supports. It is an AI model too and makes mistakes, but Lebleu calls it a good first filter when there are thousands of pages to get through.

Paste Check comes from the Wikimedia Foundation. When someone pastes text into an article, it asks whether they wrote it, and it can flag text copied straight from some AI chatbots.

The foundation tested Paste Check on 22 Wikipedias in 2025. From late June 2026, the mobile version also triggers when at least 100 characters and 10 words are entered at once, in addition to pastes of 50 characters or more.

Both detectors and people misjudge AI text

Detectors miss paraphrased text and unfamiliar models, and a 2025 study found people no better than chance, so Wikipedia never relies on either one alone. The Signs of AI writing guide tells editors not to depend on tools like GPTZero or Pangram. They beat guessing, but their error rates are still significant.

Paraphrasing, formatting changes, or a model the detector never saw in training can all throw them off. Human readers do not do much better.

A 2025 study found that people could not tell LLM text from human writing better than chance. Another 2025 study, based on German theses, found people correctly identified 57% of AI texts and 64% of human ones.

How to apply Wikipedia's test to what you read

Wikipedia's editors end up judging AI text by whether the sources support it, and anyone reading an article can run that same check. Style fools both people and software, but a citation either says what the sentence claims or it doesn't. Checking that takes a click and no expertise.

On English Wikipedia, start by looking for an AI-generated notice at the top of the page. For any claim that matters to you, open the cited source and see whether it matches. The View history tab shows when a passage was added, and anything older than November 30, 2022 is very unlikely to be AI text.

Other platforms are leaning on people too. LinkedIn uses member reports to cut the reach of AI slop (related: LinkedIn AI Slop Button Passes a Million Users). On both sites, people, not a detector score, make the final call.