Wikipedia has an AI problem. Approximately 270,000 active volunteers contribute to the online encyclopedia, and their work now includes finding material that machines generated but presented as human writing. The challenge has moved beyond obvious chatbot phrasing into structure, citations, and patterns that only become suspicious when they accumulate.
Ellia Lebleu, a cofounder of the WikiProject AI Cleanup effort, realized Wikipedia had a problem with AI-generated material around December 2023. At the time, Lebleu was pursuing a Ph.D. in AI at the Inria Center of the University of Rennes, while also working with other editors to understand what they were seeing across Wikipedia.
“It was around December 2023. A few other editors and I were chatting on Wikipedia’s Discord server, and we had all seen very strange articles with the same odd characteristics,” Lebleu said. “They repeated phrases like ‘stands as a testament’ or ‘rich cultural heritage.’ 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.”
From Clichés to Structure
The WikiProject AI Cleanup launched in 2023 to help keep Wikipedia human-written. When the cleanup began, AI writing often betrayed itself through conspicuous clichés, repeated phrases, and sections that sounded more like ChatGPT than encyclopedia material. Those clues gave volunteers a practical starting point, even if the machine had not left a label saying it wrote the article.
That advantage did not last. “By 2025, the signs of AI writing were becoming subtler and more structural,” Lebleu said. Volunteers are now developing tools to detect those less obvious forms of machine-generated content, because familiar phrases alone no longer provide a reliable test.
AI often uses the “rule of three,” negative parallelisms, and other speech structures. None of these features proves that a machine wrote an article; people use them too. But when those patterns appear together across a body of work, they become meaningful evidence rather than isolated quirks.
Timing can also raise questions. If someone writes five articles of thousands of words each in five minutes, there is reason to suspect AI. That does not identify the exact system involved, but it gives editors a signal that the content needs closer attention. Human productivity has limits, inconveniently enough.
The Citation Problem
Fake or misleading citations have become a major worry in AI-generated content on Wikipedia. AI sometimes hallucinates a citation entirely, although that is rarer today because of retrieval-augmented generation. The harder problem is a citation that exists but does not support the content attached to it.
AI citations that do not support the content are extremely commonplace. That creates a problem for editors who must check not only whether a source exists, but whether it actually backs the claim. A polished paragraph with a real citation can still mislead readers if the source says something else.
English Wikipedia recently prohibited editors from using certain sources or methods. The rule reflects the central problem facing the cleanup effort: detecting AI-generated text is only one part of maintaining accurate information. Editors must also judge the evidence behind the prose, including on pages covering arms trade.
Wikipedia’s volunteer system now faces a moving target. In 2023 and early 2024, ChatGPT’s habits made suspicious articles easy to spot; by 2025, the clues had become subtler, more structural, and harder to separate from ordinary writing. The WikiProject AI Cleanup is responding with new detection tools, but the final standard remains human judgment—especially when a citation looks convincing and turns out to support nothing.
Based on
