Up to 20 Percent of New Wikipedia Articles Are AI-Generated, Cleanup Editor Says

Between 15 and 20 percent of new Wikipedia articles and drafts are being flagged as AI-generated, according to estimates produced by the volunteer WikiProject AI Cleanup together with the AI-detection company Pangram. Ellia Lebleu, a founding member of the project, disclosed the figures in an interview with IEEE Spectrum, offering the most concrete picture yet of how much machine-written content is reaching the world's most widely used encyclopedia.

The numbers come with caveats. In late 2025, Pangram flagged roughly 20 percent of new articles and drafts as AI-generated. By early 2026, the figure had fallen to 15 or 16 percent. "But we don't yet know whether that's because our recent policies have had an effect or because models are getting more powerful and Pangram hasn't had time to adapt," Lebleu said.

The stakes are considerable. Wikipedia, which turned 25 recently, supplies training data and citations to AI systems across the industry, and its volunteer editors have staked the site's identity on remaining human-written. Lebleu, who is pursuing a Ph.D. in AI at the Inria Center of the University of Rennes, part of France's national research institute for digital science and technology, sits at the intersection of both worlds.

How did the cleanup project start?

The project began around December 2023, when Lebleu and other editors noticed a wave of strange articles sharing the same characteristics. "They repeated phrases like 'stands as a testament' or 'rich cultural heritage,'" she said. "They sometimes had conclusion sections, or sections called 'Why is this important?' That was very typical of ChatGPT in 2023 and early 2024 but very odd for Wikipedia articles—and very easy to spot."

The editors had been comparing notes on Wikipedia's Discord server. Lebleu then decided, in her words, to "create an official project on Wikipedia so more people can deal with this." WikiProject AI Cleanup, launched in 2023, now coordinates volunteers hunting for AI-generated material that violates the site's guidelines.

What does AI writing on Wikipedia look like now?

The obvious clichés of 2023 have given way to subtler tells. "By 2025, the signs of AI writing were becoming subtler and more structural," Lebleu said. "For example, AI loves the 'rule of three,' negative parallelisms, and other structures of speech that aren't conclusive by themselves but become meaningful as they accumulate."

Editors also look beyond the text itself. "If someone writes five articles of thousands of words each in 5 minutes, there is reason to suspect AI," she said.

The structural patterns—such as the "rule of three" (grouping ideas into triads) and negative parallelism (contrasting what something is not)—are not proof on their own. They accumulate into evidence.

Is the citation problem worse than the prose problem?

Possibly. Fake or misleading citations have become a major worry for the project. Fully hallucinated citations—sources the AI invented entirely—have become rarer, Lebleu said, because retrieval-augmented generation (systems that let an AI model search for and cite real sources) now anchors outputs to existing documents.

The remaining problem is subtler and more persistent. "More often, it finds a real source but gets its content completely wrong," Lebleu said. "AI citations that don't actually support what is being written are extremely commonplace." A citation that exists but misrepresents its source is harder to catch than an invented one.

How reliable are the detection numbers?

Lebleu's estimates rest on Pangram's classifier, which she validated against history. She tested it on Wikipedia content written before ChatGPT existed and found no false positives, meaning the tool never mislabeled pre-ChatGPT human writing as machine-generated. That makes its results, in her assessment, "a very conservative lower bound" on the true volume of AI content.

The measurement question matters because English Wikipedia has recently prohibited editors from using LLMs (large language models, the systems behind tools like ChatGPT and Claude) to generate articles. Whether the policy is working remains unsettled. "It's still a little early to tell," Lebleu said. "We see possible effects, such as prompt refusal and the apparent slight drop in AI-generated content. But again, we don't yet know if that is because models are getting better or because beginners are learning the rules."

The vast majority of AI-generated content comes from beginner editors, she said, so the project has worked to make the rule visible in welcome messages, editing help, and advice pages.

Is Wikipedia fighting AI with AI?

Increasingly, yes. A fellow editor on the project has built an AI source verifier that walks through the sources cited on a page and checks whether they match the article's claims. Lebleu is candid about its limits: "Since it is itself an AI model, there is no guarantee that it will be correct every time. But if you want something that works at scale and filters out obvious cases, it's a good first step."

The Wikimedia Foundation, the nonprofit that operates Wikipedia, is developing a tool called Paste Check, which can flag text pasted directly from some AI chatbots.

Model builders are also involved, unevenly. "We've seen Claude sometimes say, 'I can't write a Wikipedia article; it would go against Wikipedia's policies,'" Lebleu said. Universal prompt refusal of that kind would not solve the problem—open-weight models (models whose parameters anyone can download and modify) can be tweaked to bypass restrictions—but, she noted, "it would help with people who use commercial models and aren't especially familiar with the technology."

Could AI ever write acceptable Wikipedia articles?

This is the philosophical fault line under the practical policy. The ban on LLM-generated articles passed for operational reasons: "We were flooded with AI content, we didn't have good tools, and we needed to put a stop to it," Lebleu said. Since then, the rule has taken on identity-defining weight.

"Many people have called Wikipedia 'the last good place on the internet,'" she said. "Being a place that was, in theory, free from AI slop has become part of Wikipedia's identity." A lot of the encyclopedia's communication now centers on the idea that it is a human-written project—and even AI companies see value in that, she added, because a purely human-written corpus is valuable for training AI systems. "The fact that Wikipedia is still that, and will remain that for at least the near future, makes it a precious resource that isn't replicated anywhere else."

She leaves the door open to hybrid models: "Will the people who read and contribute to Wikipedia want it to become partially AI-generated if that makes it more comprehensive? We could preserve Wikipedia's identity as a human-written corpus while also offering an AI-assisted option."

What about "Google Zero" and AI search?

Lebleu also addressed the scenario in which AI answer engines stop sending traffic to websites at all. "If everything goes through AI search and websites plummet to zero visibility, how do we expose Wikipedia and show that it can be edited and that people can still contribute?" she asked. Wikipedia is frequently cited in Google's AI Overviews, but users may only see an AI-generated answer and never realize a living, editable project stands behind it.

The Wikimedia Foundation is building an attribution framework that would give AI users more information about the cited article: when it was last edited, how comprehensive it is, whether editors have flagged problems, and perhaps a call to action. In the future, Lebleu suggested, readers might contribute through AI models—asking an assistant to file an edit request when an AI answer citing Wikipedia contains an error.

Her closing framing captures the project's trajectory: "Yesterday, AI was something to clean up. Today, it's a tool to help us with the cleanup. Tomorrow, it will be the air we breathe."

The near-term test is measurable: whether the 15-to-16-percent detection rate of early 2026 keeps falling, and whether tools like Paste Check and the source verifier can hold the line as models grow harder to detect.