Inside Wikipedia’s Cleanup of AI-Generated Content

Ellia Lebleu is a founding member of the WikiProject AI Cleanup, a volunteer effort launched in 2023 to search out and take away AI-generated materials from Wikipedia that breaks the web site’s pointers. When the cleanup started, AI writing was usually betrayed by conspicuous clichés. At this time, the challenge’s volunteers are creating instruments to detect subtler types of machine-generated content material and debating AI’s position on the earth’s most generally used encyclopedia.

When did you understand Wikipedia had an issue with AI-generated materials?

Ellia Lebleu: It was round December 2023. A number of different editors and I have been chatting on Wikipedia’s Discord server, and we had all seen very unusual articles with the identical odd traits. They repeated phrases like “stands as a testomony” or “wealthy cultural heritage.” They generally had conclusion sections, or sections referred to as “Why is that this vital?” That was very typical of ChatGPT in 2023 and early 2024 however very odd for Wikipedia articles—and really straightforward to identify. Then I made a decision, let’s create an official challenge on Wikipedia so extra individuals can take care of this.

How do editors spot AI-generated writing in the present day?

Lebleu: By 2025, the indicators of AI writing have been changing into subtler and extra structural. For instance, AI loves the “rule of three,” adverse parallelisms, and different constructions of speech that aren’t conclusive by themselves however turn into significant as they accumulate. We additionally look past the article. If somebody writes 5 articles of 1000’s of phrases every in 5 minutes, there may be motive to suspect AI.

Does the AI drawback simply seem in the primary textual content, or additionally in citations?

Lebleu: Pretend or deceptive citations have turn into a serious fear. Typically an AI hallucinates a quotation fully, though that’s a lot rarer in the present day due to retrieval-augmented technology that permits the AI to search out sources. Extra usually, it finds an actual supply however will get its content material utterly incorrect. AI citations that don’t truly help what’s being written are extraordinarily commonplace.

English Wikipedia lately prohibited editors from utilizing LLMs to generate articles. How a lot AI-generated content material remains to be getting by?

Lebleu: I’ve been working with [AI-detection company] Pangram to estimate the proportion of AI-generated content material. I examined it on Wikipedia content material from earlier than ChatGPT and located no false positives, so its outcomes present a really conservative decrease sure. In late 2025, it flagged maybe 20 p.c of latest articles and drafts as AI-generated. In early 2026, that went down to fifteen or 16 p.c. However we don’t but know whether or not that’s as a result of our current insurance policies have had an impact or as a result of fashions are getting extra highly effective and Pangram hasn’t had time to adapt.

Is Wikipedia utilizing AI to fight AI?

Lebleu: One other Wikipedia editor has made an AI supply verifier, which matches by the sources on a web page and checks whether or not they match the claims. Since it’s itself an AI mannequin, there isn’t any assure that it will likely be right each time. However if you need one thing that works at scale and filters out apparent circumstances, it’s first step. The Wikimedia Basis can also be engaged on a software referred to as Paste Check, which might flag textual content pasted instantly from some AI chatbots.

And we’ve seen Claude generally say, “I can’t write a Wikipedia article; it will go towards Wikipedia’s insurance policies.” If all main AI firms applied that form of immediate refusal, it wouldn’t resolve the issue fully, as a result of there are nonetheless open-weight fashions that may be tweaked to keep away from restrictions. However it will assist with individuals who use industrial fashions and aren’t particularly accustomed to the expertise.

Are you able to inform if the rule prohibiting AI writing is working?

Lebleu: It’s nonetheless a little bit early to inform. We see potential results, akin to immediate refusal and the obvious slight drop in AI-generated content material. However once more, we don’t but know if that’s as a result of fashions are getting higher or as a result of inexperienced persons are studying the foundations. The overwhelming majority of AI-generated content material is submitted by newbie editors, so we’ve made many efforts to make the rule very seen in welcome messages, enhancing assist, and recommendation.

What if AI turns into able to writing really dependable Wikipedia articles?

Lebleu: That’s one of many huge questions we’re dealing with. When the rule of thumb was handed, it was for very sensible causes: We have been flooded with AI content material, we didn’t have good instruments, and we wanted to place a cease to it. However since then, it has taken on a extra philosophical dimension. Many individuals have referred to as Wikipedia “the final good place on the web.” Being a spot that was, in principle, free from AI slop has turn into a part of Wikipedia’s identification.

Quite a lot of Wikipedia’s communication is now centered on the concept that we’re a human-written encyclopedia. Even AI firms discover that attention-grabbing. It may be priceless to have a purely human-written corpus on which to [train] AI methods. The truth that Wikipedia remains to be that, and can stay that for at the very least the close to future, makes it a treasured useful resource that isn’t replicated anyplace else.

However will the individuals who learn and contribute to Wikipedia need it to turn into partially AI-generated if that makes it extra complete? We might protect Wikipedia’s identification as a human-written corpus whereas additionally providing an AI-assisted choice.

Has Wikipedia reckoned with the concept of “Google Zero”—a potential future during which web sites now not get visitors from Google search?

Lebleu: If every little thing goes by AI search and web sites plummet to zero visibility, how will we expose Wikipedia and present that it may be edited and that individuals can nonetheless contribute? Wikipedia is steadily cited in [Google’s] AI Overviews, however customers might solely see an AI reply citing Wikipedia. They might not understand that Wikipedia is a dwelling challenge they will edit and contribute to.

The Wikimedia Basis is engaged on an attribution framework that might give AI customers extra details about the article [being cited]: when it was final edited, how complete it’s, whether or not editors have flagged issues, and even perhaps a name to motion. Sooner or later, readers would possibly even contribute by AI models: “The AI cited Wikipedia, however there appears to be an error. Can I make an edit request and add this different supply I discovered?”

I can see a really sturdy symbiotic relationship between Wikipedia and AI fashions sooner or later. I don’t see Wikipedia disappearing, however I can see it taking a really completely different form. Yesterday, AI was one thing to wash up. At this time, it’s a software to assist us with the cleanup. Tomorrow, it will likely be the air we breathe.

A condensed model of this text seems within the November 2026 print difficulty as “Ellia Lebleu.”