Wikipedia's volunteer editors say they are holding the line on human-written articles, and they are now using AI to help. In an IEEE Spectrum interview published October 21, 2026, Ellia Lebleu, a founding member of WikiProject AI Cleanup and a PhD student in AI at the Inria Center of the University of Rennes, describes how the project spots machine-written content, what an AI detector measured, and what comes next.
For anyone who builds with language models, it is a rare look at what AI text looks like to the people who read the most of it, and at how a large community enforces a "no LLM articles" rule.
TL;DR: Wikipedia and AI-generated content
| Question | Answer, as reported |
|---|---|
| Is AI-written content allowed? | English Wikipedia prohibits using LLMs to generate articles |
| Who polices it? | WikiProject AI Cleanup, a volunteer effort started in 2023 |
| How much new content was flagged? | About 20% of new articles and drafts in late 2025, about 15 to 16% in early 2026 (Pangram) |
| Who writes most of it? | The vast majority comes from beginner editors |
| What tools exist? | Pangram, an AI source verifier, and Wikimedia's Paste Check |
| Is the policy working? | Too early to say, per Lebleu |
How did WikiProject AI Cleanup start?
Around December 2023, Lebleu and other editors noticed a pattern of odd articles while talking on Wikipedia's Discord server and decided to make it an official project. The goal was to find and remove AI-generated material that violates Wikipedia's guidelines.
Early on, the work was easy in one sense: the text carried obvious clichés, stock phrases and unusual section headings that were typical of ChatGPT in 2023 and early 2024. By 2025 the tells had become subtler and more structural, and the project's job moved from deleting obvious junk to building tools that catch the harder cases and debating what role AI should have at all.
How do editors tell if text is AI-written?
According to the interview, no single signal is enough. Lebleu points to patterns such as the "rule of three" and negative parallelisms ("not X, but Y"), each weak alone but meaningful in combination. Volume is another flag: five articles of thousands of words each, produced in about five minutes, is a reason to suspect a model.
The bigger shift is in citations. Fabricated references used to be the giveaway. Now that retrieval-augmented generation can find real sources, fake citations are rarer. The more common problem is a real source whose content does not support the sentence it is attached to. That is much harder to catch by eye and is why checking claims against sources has become the center of the work.
This matches a theme we have covered elsewhere: AI output that looks fine on the surface but needs checking, the cost we described in the AI verification tax, and the stylistic regularities in the NYU and UT Austin study of six patterns that make AI text hard to read.
What did the detector measure?
Lebleu worked with Pangram, an AI-detection company. Reported results:
- Late 2025: roughly 20% of new articles and drafts were flagged as AI-generated.
- Early 2026: roughly 15 to 16%.
- Pangram showed no false positives on pre-ChatGPT Wikipedia text, so Lebleu treats its figures as a conservative lower bound.
- The vast majority of AI-generated submissions come from beginner editors.
She cautions that it is too early to say what caused the drop. It could mean the policy is working, that newcomers are learning the rules, or that newer models write in ways the detector catches less often. Treat the numbers as a sample from one tool, not a census.
Detectors are a contested area. We looked at the book-prize dispute where a detector's verdict was challenged in the Goncourt and Pangram case, and at a research approach to making them auditable in DeepMind's DetectorScope. The Wikipedia case is a favorable one for a detector, since editors can pair its flags with human review and source checks.
What tools and policies is Wikipedia using?
A rounded sheet with a faint pattern in the corner, a picture for how Wikipedia AI cleanup tools mark machine-written text
- Policy. English Wikipedia's rule, Writing articles with large language models, prohibits using LLMs to generate articles. It is surfaced in welcome messages, editing help and advice to new editors.
- AI source verifier. Another editor built a tool that checks whether cited sources match the claims. It is itself an AI model, so it can be wrong, but it helps filter obvious cases at scale.
- Paste Check. The Wikimedia Foundation's Paste Check flags text pasted directly from some AI chatbots.
- Pangram. Used by volunteers to estimate how much AI text is arriving.
- Chatbot refusals. Lebleu says Claude sometimes declines to write Wikipedia articles, citing the site's policies, though open-weight models can be modified to bypass such restrictions.
Lebleu's summary of the change in attitude: "Yesterday, AI was something to clean up. Today, it's a tool to help us with the cleanup."
Why does this matter beyond Wikipedia?
Training and retrieval. Wikipedia is a large, curated, human-written corpus. Lebleu describes it as valuable for AI training. If machine text leaks in at scale, future models risk learning from their own output, and retrieval systems that cite Wikipedia inherit its errors.
The "Google Zero" problem. If AI search sends less traffic to websites, readers may see Wikipedia facts through a chatbot without knowing the page is editable or how reliable it is. The Wikimedia Foundation is developing an attribution framework that would show users of AI products details such as last edit date, comprehensiveness, flagged problems and a call to action to contribute. Compare this with the licensing and priority questions in ChatGPT's reported partner-publisher lane.
A pattern across communities. Wikipedia is one of many projects drawing lines on AI contributions: see QGIS banning AI-assisted contributions, arXiv's cap on submissions, and Spotify removing 50 million slop tracks. The common thread is that cheap generation shifts the cost to reviewers, so communities respond with rules, filters and friction.
What are readers and editors likely to ask next?
Can a detector alone justify deleting an article? Not on the evidence in the interview. The volunteers describe detector output as one input, alongside stylistic signals, volume and a check of whether the sources support the text. A flag raises suspicion; it does not prove authorship. That is a sensible stance given how often detectors are disputed elsewhere.
Why do beginners account for most of it? The interview says the vast majority of AI-generated submissions come from new editors. A plausible reading, which is our inference and not a finding in the interview, is that experienced editors already know the policy and the sourcing standard, while newcomers see a chatbot as a shortcut to a first article. That is why the project has pushed the rule into welcome messages and onboarding advice rather than relying only on cleanup after the fact.
Is a falling flag rate good news? Possibly, but the interview is careful. A drop from about 20% to about 15 or 16% could reflect better rule awareness, or newer models whose output the detector flags less. Without a control, it is not possible to tell which. For anyone tracking AI content on their own platform, the lesson is to record the detector version, the sampling method and the date, so a trend can be interpreted later.
What does "human-written" buy a reader? It buys accountability. A human editor can be asked why a claim is there, a revision history shows who changed what, and the community can reverse mistakes. Those are the properties that make Wikipedia useful as a reference for people and for models.
How does this compare with other communities? Open-source projects and platforms are drawing different lines: some ban AI contributions outright, some require disclosure, some filter at the platform level. Wikipedia sits near the strict end for article text, while still welcoming AI as a helper for cleanup and verification. That split between "AI writes the content" and "AI helps check the content" is likely to become the common pattern.
Where might Wikipedia go next?
Lebleu suggests Wikipedia could keep its human-written identity while also offering an optional AI-assisted version, though that is her idea, not policy. She expects a strong symbiosis between Wikipedia and AI, with Wikipedia changing shape rather than disappearing. Whether the volunteer base can keep pace as models improve is the open question.
What can you take from this if you publish or build?
- Verify claims against sources, not style. As fabricated citations fade, the failure that remains is a real source that does not say what the text claims. Build or use checks for that.
- Do not rely on one detector. Combine tool flags with human review, as the volunteers do.
- If you contribute to community projects, read their AI policy before submitting. More are banning or limiting machine-written contributions.
- If you build products that reuse reference content, show provenance and freshness. Wikimedia's attribution framework is a model.
What we could not check
This post draws on one interview. We did not independently test Pangram, and the detection percentages are as reported by the interviewee. We also could not confirm exact release timing for Paste Check or the attribution framework beyond what the interview says.
Related reading
- The AI verification tax: Sage, Workday and Foxit surveys
- Six patterns that make AI text hard to read
- QGIS bans AI-assisted contributions
- arXiv caps submissions at two per month
- Goncourt, Pangram and the AI detector dispute
- Spotify removes 50 million slop tracks
Figures are as reported by IEEE Spectrum on October 21, 2026, and may change.