Wikipedia AI-generated content: how editors find it, delete it and now ban it, and what the numbers show

Wikipedia AI-generated content is handled by volunteers, not the company behind the site, and their approach has hardened in three steps since 2023. A WikiProject finds and fixes chatbot-written text; a speedy-deletion rule added in August 2025 lets administrators delete obviously unreviewed AI articles without a week of debate; and since March 20, 2026, an English Wikipedia guideline has prohibited using large language models "to generate or rewrite article content." The topic drew fresh attention on October 27, 2026, when IEEE Spectrum's interview with a cofounder of the cleanup effort reached 473 points on Hacker News. For AI search, it matters twice over: Wikipedia is one of the sources chatbots lean on most, and the cleanup is one of the largest attempts to keep machine-written text out of a human-edited reference. Open-source communities face a version of the same problem, as our look at AI-generated code policies shows.

This analysis is based on English Wikipedia's own project pages (the speedy-deletion policy, the guideline on writing articles with LLMs and the request for comment that adopted it, WikiProject AI Cleanup and "Signs of AI writing"), the Wikimedia Foundation's Diff posts on crawler traffic and reader trends, IEEE Spectrum's October 21, 2026 interview with Ellia Lebleu, and reporting by 404 Media, PCWorld, TechCrunch, Decrypt, SiliconANGLE, MediaNama and Quartz, checked October 28 and 29, 2026.

The short version

• WikiProject AI Cleanup: a volunteer group formed in December 2023 to find and fix AI-generated text that breaks Wikipedia's rules.
• Speedy deletion (G15): since August 2025, administrators can quickly delete pages that show unambiguous signs of LLM generation, such as text addressed to the user, references that don't exist, or chatbot markup.
• The March 2026 guideline: English Wikipedia prohibits LLM-written or LLM-rewritten article text, with two exceptions: copyediting your own writing and first-pass translation, both checked by a human.
• The signs: stock phrases, puffed-up importance claims, chatbot leftovers and citations that don't support the text. Style alone isn't enough for sanctions.
• The Foundation's role: it says volunteers set editorial policy. Its own concerns are AI scrapers, which it says drove a 50% rise in multimedia bandwidth, and a roughly 8% fall in human pageviews.

How is Wikipedia dealing with AI-generated content?

Through community rules that tightened over three years:

Last verified: October 29, 2026

What is WikiProject AI Cleanup?

It's a volunteer project, described in 404 Media's October 2024 report as "a collaboration to combat the increasing problem of unsourced, poorly-written AI-generated content on Wikipedia." Its project page was created on December 4, 2023, according to Wikipedia's page history. The page says its first goal is to identify AI-written text and "scrutinize such text to make sure it follows Wikipedia's policies," removing unsourced or likely inaccurate claims first. IEEE Spectrum identifies the cofounder it interviewed as Ellia Lebleu; 404 Media's 2024 report and Quartz's 2026 report name a founding member Ilyas Lebleu. Lebleu told IEEE Spectrum that around December 2023 Lebleu and other editors noticed similar AI-like articles and decided to create an official project. Early on, members generated text with ChatGPT themselves to learn its catchphrases, then used those to spot the most obvious cases, 404 Media reported.

The project doesn't ban AI tools on its own; it works under Wikipedia's content policies, which require verifiable sources and a neutral point of view. AI text tends to fail those policies, and that, not the use of AI itself, was the original basis for removing it.

The project also helped draft the March 2026 guideline, which Lebleu proposed under the username Chaotic Enby, according to Quartz. "More and more administrative reports centered on LLM-related issues, and editors were being overwhelmed," Lebleu told 404 Media, as Quartz reported.

What is G15, the speedy-deletion rule for AI articles?

Normally, deleting a Wikipedia article goes through a discussion that runs about a week. Speedy deletion is a short list of cases an administrator can act on immediately, which before August 2025 mostly meant obvious nonsense and thinly disguised advertising, PCWorld reported.

On August 4, 2025, editors added criterion G15, then titled "LLM-generated pages without human review," to Wikipedia's speedy-deletion policy after a request for comment. Now titled "Unambiguously LLM-generated pages," it applies to any page showing one of three signs that it "could only plausibly have been generated by" an LLM:

• Communication intended for the user, such as "Here is your Wikipedia article on…", knowledge-cutoff disclaimers or unfilled placeholders.
• Non-existent or nonsensical references, such as links dead on arrival, ISBNs with invalid checksums or DOIs that point to unrelated papers.
• Technical indicators, such as chatbot citation markup like oaicite or turn0search0.

Finding these suggests nobody checked the text before submitting it. Lebleu called the rule a "band-aid" for the most egregious cases, PCWorld reported, citing 404 Media.

The rule targets output nobody reviewed, not every article that might involve AI. The policy says other common signs of LLM writing, even an author admitting to using an LLM, aren't enough for G15; those pages go to a proposed-deletion route for AI content, the normal deletion discussion or the AI noticeboard.

What signs of AI writing do Wikipedia editors look for?

The project maintains a field guide, "Signs of AI writing." TechCrunch described it in November 2025 as cataloging habits, not magic words. Examples it reported include:

• Inflated importance: generic claims such as "a pivotal moment" or "a broader movement."
• Trailing clauses that assert significance: "emphasizing the significance," "reflecting the continued relevance."
• Marketing tone: phrasing that sounds "more like the transcript of a TV commercial."
• Padding with minor media mentions to make a subject look notable.

Lebleu added others in the IEEE Spectrum interview: stock phrases like "stands as a testament" and "rich cultural heritage," common in ChatGPT output in 2023 and early 2024; conclusion sections and headings like "Why is this important?"; the "rule of three"; and an implausible output rate, such as five long articles in five minutes.

The hardest problem now is citations. Outright invented sources are rarer than they were, Lebleu said, but real sources often don't support what the text attributes to them, which takes a human reading the source to catch.

Two caveats run through this. TechCrunch reported in November 2025 that the guide called automated detectors "basically useless" for this job and treated single words such as "delve" as weak evidence; the current version says not to rely solely on detection tools such as GPTZero and Pangram, which "have non-trivial error rates." And the March 2026 guideline as adopted said "more evidence than just stylistic or linguistic signs is needed to justify sanctions." In practice, editors weigh an account's contribution history and whether the text meets content policy, not prose style alone.

How much of Wikipedia is AI-generated?

There's no official figure. The best-known numbers come from Lebleu's work with Pangram, a company that sells an AI-text detector. Lebleu told IEEE Spectrum that the detector "flagged perhaps 20 percent of new articles and drafts as AI-generated" in late 2025, and "15 or 16 percent" in early 2026.

Those numbers need context. They come from a commercial detector, cover new articles and drafts rather than the whole encyclopedia, and Lebleu describes them as a conservative lower bound. Lebleu said Pangram showed no false positives on content written before ChatGPT existed, but that is the vendor's tool assessed by the people using it, not an independent audit. Lebleu also said it's unclear whether the drop reflects the new rules or changes in the models. The project's own "Signs of AI writing" guide warns editors not to rely solely on detectors, Pangram included.

What did English Wikipedia ban in March 2026?

The guideline, as adopted on March 20, read: "Text generated by large language models (LLMs) often violates several of Wikipedia's core content policies. For this reason, the use of LLMs to generate or rewrite article content is prohibited, save for the exceptions given below." News reports, including Decrypt's and SiliconANGLE's, called it a policy; on Wikipedia it is a content guideline. The exceptions are suggesting basic copyedits to your own writing, as long as the tool adds no new content and you check it, and producing a first-pass translation from another language edition, under a separate translation guideline.

The request for comment closed on March 20, 2026, with 44 votes in favor and two opposed, according to its closing statement (Quartz had reported 40 to 2). A separate request for comment, run from March to May 2026, added that persistent LLM use outside the exceptions "is considered disruptive and will result in an editor being blocked," the current guideline says. The rule applies to English Wikipedia; other language editions set their own, and German Wikipedia has adopted separate AI guidelines, heise reported.

What does the Wikimedia Foundation say?

The Foundation, which runs Wikipedia's servers, says the rules aren't its call. "The Wikimedia Foundation does not determine editorial policies and guidelines on Wikipedia; volunteer editors do," a spokesperson told Decrypt in March 2026.

Its own AI strategy, announced April 30, 2025 by machine learning director Chris Albon and research head Leila Zia, commits to using AI to support volunteers over the next three years, for moderation, translation and onboarding, and to "take a human-centered approach," TechCrunch reported. IEEE Spectrum says the Foundation is also developing Paste Check, which can flag text pasted in from some AI chatbots, and is building an attribution framework to tell AI users when a cited article was last edited and whether editors have flagged problems.

The Foundation's bigger complaints are about traffic. On April 1, 2025, it said bandwidth for multimedia downloads had grown 50% since January 2024, mainly from scrapers collecting data for AI models, and that "at least 65%" of its most resource-intensive traffic came from bots, which made up "about 35%" of pageviews. "Our content is free, our infrastructure is not," the post said. In October 2025, after reclassifying traffic that turned out to be evasive bots, it reported human pageviews down "roughly 8%" from the same months of 2024, and said AI and search companies using Wikipedia content "must encourage more visitors" to the site. For how sites can identify and limit those crawlers, see our guide to detecting and blocking AI agents. For why being a trusted source for AI answers matters, see how AI answers get gamed.

What's confirmed, reported and uncertain

• Confirmed by the Wikimedia Foundation's own posts: the 50%, 65% and 35% crawler figures; the roughly 8% human pageview decline and the bot reclassification.
• Confirmed by Wikipedia's own pages: the WikiProject's December 2023 start; G15's August 2025 adoption and current text; the March 2026 guideline's text, its 44–2 vote and March 20 close; the blocking rule; the "Signs of AI writing" caveat on detectors.
• Reported: Lebleu's account (IEEE Spectrum, 404 Media, Quartz); the "band-aid" remark (PCWorld, citing 404 Media); the Foundation's strategy (TechCrunch) and statement (Decrypt); the Pangram figures (Lebleu, via IEEE Spectrum).
• Uncertain: how many pages have been deleted under G15, and how much AI text remains in older articles.