Wikipedia's volunteer editors are fighting a quiet war against machine-generated content, and for the first time they have hard numbers. According to a recent IEEE Spectrum interview with Ellia Lebleu, a founding member of the WikiProject AI Cleanup, an AI-detection tool estimated that as many as 20 percent of new articles and drafts on English Wikipedia were machine-generated in late 2025 — and 15 to 16 percent in early 2026. The interview drew fresh attention this week after topping Hacker News.
Lebleu, a Ph.D. student in AI at the Inria Center of the University of Rennes, co-founded the cleanup project in December 2023, when she and other editors on Wikipedia's Discord server noticed a wave of strange new articles sharing the same odd characteristics. For more context on this story, see our ongoing AI trends.
From Obvious Cliches to Subtle Tells
In the early days, the tells were conspicuous. AI-written articles repeated stock phrases like "stands as a testament" or "rich cultural heritage," and often included formal conclusion sections or odd headers like "Why is this important?" — trademarks of ChatGPT-era output that made detection relatively easy.
By 2025, the signals had grown subtler and more structural, Lebleu told IEEE Spectrum. AI writing loves the "rule of three" and negative parallelisms — patterns of speech that mean nothing on their own but become meaningful in accumulation. Editors also look beyond the single article: if an account produces five articles of thousands of words each in five minutes, something is wrong.
The Citation Problem
The most serious worry is not style but sourcing. Fully hallucinated citations — references to papers that do not exist — have become much rarer as retrieval-augmented generation lets models find real sources. But Lebleu said the more common failure is subtler and more dangerous: the AI finds a real source and then mischaracterizes it, citing support that the source does not actually provide. She called AI citations that do not support what is being written "extremely commonplace."
The community is fighting back with tools of its own. One Wikipedia editor built an AI source verifier that reads the sources on a page and checks whether they match the claims — an AI model policing AI text, which Lebleu acknowledged is not guaranteed to be correct but works at scale as a first filter. The Wikimedia Foundation, meanwhile, is developing a tool called Paste Check that can flag text pasted directly from some AI chatbots.
How Much Wikipedia Content Is AI-Written?
The 15-to-20 percent estimates come from Lebleu's collaboration with Pangram, an AI-detection company. She stress-tested the detector on Wikipedia content from before ChatGPT and found no false positives, meaning the figures should be read as a conservative lower bound rather than a ceiling.
What explains the decline from 20 percent to 15 or 16 percent is unresolved. It could be that Wikipedia's recent policies are working — or that newer models produce less detectable text while detection tools have not caught up. Lebleu told IEEE Spectrum that most AI-generated content comes from beginner editors, and the project has worked to make the rules highly visible to newcomers.
The Ban, Refusals, and a Philosophical Question
English Wikipedia has already acted, prohibiting editors from using LLMs to generate articles — a ban The Guardian reported on in March 2026 after the community approved the guideline. The policy's effects are still being measured. By Lebleu's account, the change was a matter of necessity rather than ideology: "We were flooded with AI content, we didn't have good tools, and we needed to put a stop to it," she told IEEE Spectrum.
One measurable shift is already visible on the model side: Lebleu noted that Claude sometimes refuses requests outright, responding that it cannot write a Wikipedia article because doing so would violate Wikipedia's policies. She argued that if all major AI companies implemented similar refusals, it would meaningfully help — though open-weight models that can be tweaked to bypass restrictions would remain a loophole.
The deeper question dividing the project is philosophical. What began as a practical measure has, in Lebleu's words, taken on a more philosophical dimension since then, as many people have come to call Wikipedia "the last good place on the internet" — one of the few major information sources still, in principle, written by humans.
Why It Matters Beyond Wikipedia
The stakes extend well past the encyclopedia itself. Wikipedia is a top reference result for search engines, a common source for AI training data, and a frequent first stop for fact-checking. If a meaningful share of its new pages are machine-written with mischaracterized citations, the errors propagate: into search snippets, into model training corpora, and back into future AI outputs.
For now, the cleanup continues one flagged article at a time — a volunteer effort using pattern-spotting, edit-velocity checks, and increasingly, AI itself, to keep a human-written encyclopedia human.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
