The encyclopedia anyone can edit is learning that "anyone" now includes machines pretending to be people.
The Summary
- Wikipedia's WikiProject AI Cleanup, launched in 2023, is a volunteer effort hunting AI-generated content that violates the site's human-authorship guidelines
- Early AI writing was easy to spot ("stands as a testament"), but 2025 saw subtler tells: the rule of three, negative parallelisms, structural patterns humans rarely use
- The real threat isn't slop, it's speed: editors who produce thousands of words across multiple articles in minutes, and citations that look real but lead nowhere
The Signal
Ellia Lebleu, a founding member of WikiProject AI Cleanup and PhD researcher at France's Inria Center, noticed the problem in late 2023. Strange articles kept appearing with the same tics: phrases like "rich cultural heritage," conclusion sections that read like high school essays, headers asking "Why is this important?" Classic ChatGPT tells from that era. The kind of writing that made you immediately suspicious.
The project started on Wikipedia's Discord server, a group of editors comparing notes and realizing they were all seeing the same thing. Lebleu formalized it into an official WikiProject so more volunteers could join the hunt. By 2025, the game changed. The obvious clichés faded. What remained were structural patterns, speech rhythms that felt off but weren't immediately damning.
"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."
This is the tell that matters now. Not a smoking-gun phrase, but an accumulation of small choices no human writer would consistently make. The way an AI defaults to certain rhythms because those patterns appeared most often in its training data. It's like spotting a native speaker versus someone who learned a language from textbooks. Technically correct, structurally weird.
The second tell is metadata, not content:
- Five articles written in five minutes
- Thousands of words produced simultaneously
- Contribution velocity that no human can match
These patterns force the question: are you using AI to draft, or are you just running prompts and pasting output? Wikipedia's policy is clear. The encyclopedia is human-written. Assistive tools are fine. Ghost-authorship by machines is not.
The citation problem is worse. Early AI models hallucinated sources wholesale, inventing authors and titles that sounded plausible but didn't exist. By 2025, models got better at generating references that look real but still mislead. A citation might point to a real publication with a fake article title, or a real author who never wrote that piece. Harder to catch. Harder to trust.
This matters because Wikipedia is infrastructure. It's the default knowledge layer for the internet, the source LLMs trained on, the place people go when they want ground truth. If that layer gets poisoned with machine-generated content masquerading as human work, the whole stack corrupts. AI trains on AI output. The commons degrades.
The Implication
The WikiProject AI Cleanup is a preview of the verification wars coming to every platform that relies on human contribution. As agents get better at mimicking human writing patterns, the tell won't be what they write, it's how fast they write it and whether anyone can verify the sources they cite.
For anyone building with AI, the lesson is simple: transparency beats deception every time. If your agent wrote it, say so. If it's assisted, make that clear. The alternative is a trust collapse where nothing can be taken at face value and every contribution requires forensic analysis. Wikipedia's editors are already living in that world. The rest of the internet is next.