When Wikipedia editors began noticing AI-generated text in 2023, the job was relatively simple: spotting stock phrases and the clichés typical of chatbots. Today it is no longer so. WikiProject AI Cleanup, a volunteer effort to remove content that breaks the encyclopedia's rules, must now deal with machine-written text far harder to detect.
Cofounder Ellia Lebleu says the problem was already evident at the end of 2023, when they were working with other editors to understand what was happening on Wikipedia. Over time the signals have changed: by 2025 the main clues were more structural, such as models' fondness for the "rule of three", negative parallelisms and other rhetorical constructions that conclude nothing precise.
Image: IEEE Spectrum
Wikipedia's guidelines largely prohibit content generated by language models, because these systems tend to invent information, a problem that directly clashes with the requirement of verifiable sources. The cleanup is carried out mainly at the AI noticeboard, but involves editors across the site.
According to Lebleu's estimates, in early 2026 around 15-16% of articles and new drafts still showed traces of AI-generated text. That is down from about 20% at the end of 2025, though it is unclear whether the drop reflects new rules or other factors.
The work is therefore shifting towards more refined tools. Volunteers are developing systems to identify subtler forms of machine-generated content, while the community debates what role AI should play in the world's most widely used encyclopedia. The issue is all the more delicate because Wikipedia is also one of the main sources used to train many of the models themselves.
No comments yet. Be the first!