MyTechDailyNews General: How Wikipedia Volunteers Spot AI Writing That Is Getting Harder to Detect

By Sarah Chen

When volunteers at WikiProject AI Cleanup compare notes, they are not chasing a single phrase anymore. Ellia Lebleu helped guide a response that started in 2023 and now has to judge structure, source use, editing speed, and community identity. Lebleu is pursuing a Ph.D. in AI at the Inria Center of the University of Rennes, part of France’s national research institute for digital science and technology. The project’s experience is useful for general tech readers because it shows how platforms adapt when AI text becomes less obviously artificial.

For MyTechDailyNews, the story is a practical tour through a moderation problem. It is not only “can a detector find machine writing?” It is how a volunteer system protects a reference site while facing tools that write fluently, cite plausibly, and may soon be part of the environment around reading and editing.

Obvious AI Text in 2023 Gave Way to Structural Patterns

Lebleu says editors noticed a wave of suspicious pages around December 2023. They were talking on Discord, and the same odd page habits kept appearing. Some pages used “stands as a testament” or “rich cultural heritage.” Others had formulaic conclusion paragraphs or headings such as “Why is this important?” That style looked typical of ChatGPT in 2023 and early 2024, but unusual on Wikipedia. It made the problem easy enough to name, and Lebleu pushed to formalize the cleanup effort.

By 2025, the surface cues had thinned. Lebleu describes the review as a search for recurrence. Models may favor the “rule of three,” negative parallelisms, and other repeated rhetorical habits. One page with one pattern is not proof. Many similar pages build suspicion. Editing behavior matters too. If someone posts five articles of thousands of words each in 5 minutes, the pace itself suggests machine assistance.

Citations: The Risk Shifts from Fake Sources to Misused Sources

Fabricated references were once an obvious worry. Lebleu notes that retrieval-augmented generation makes it easier for models to find real sources, so purely invented citations are less common. The problem has moved into the relationship between source and sentence. A page may cite a real document while changing what that document supports. When a citation does not back the claim written next to it, editors describe the issue as extremely widespread.

Estimates: How Much Still Appears Despite the Ban

English Wikipedia recently prohibited editors from using LLMs to generate articles. To estimate what still arrives, Lebleu worked with Pangram, an AI-detection company. She first tested Pangram’s tool on Wikipedia text created before ChatGPT and saw no false positives, so she treats later detections as a cautious lower bound rather than a precise total.

In late 2025, the detector marked perhaps 20 percent of new articles and drafts as machine-generated. In early 2026, that share dropped to 15 or 16 percent. Lebleu cautions that the decrease could come from recent policies, more capable models, or Pangram needing more time to adapt.

Timeline Snapshot

Period	What volunteers watched	Why it mattered
around December 2023	Repeated phrases such as “stands as a testament” and “rich cultural heritage.”; headings such as “Why is this important?”	The style looked typical of ChatGPT in 2023 and early 2024, so it stood out on Wikipedia.
By 2025	Structural signals, including the “rule of three,” and unusually fast production such as five articles of thousands of words each in 5 minutes.	Obvious wording faded, so editors leaned on repetition and context.
late 2025	Pangram marked perhaps 20 percent of new articles and drafts.	Estimation had to be handled carefully because the policy ban was already in force on English Wikipedia.
early 2026	The flagged share fell to 15 or 16 percent.	The change could reflect rules, stronger models, or detector adaptation.

Models as Review Aids, Not Guarantees

Volunteers have also started using AI to help review AI-assisted pages. One editor created an AI source verifier that scans the references listed on a page and tests whether they fit the claims. Because the verifier is itself a model, its answers cannot be treated as certain. Lebleu says it can still filter obvious cases at scale and provide a useful first step.

The Wikimedia Foundation is separately building Paste Check, a tool meant to flag text pasted straight from certain AI chatbots. The broader lesson is that detection is becoming a layered process: tools can reduce human work, but they still leave questions about accuracy and trust.

Refusals, Open Models, and Beginner Editors

Some commercial model behavior can help, Lebleu says. She points to a pattern Claude sometimes shows: “I can’t write a Wikipedia article; it would go against Wikipedia’s policies.” She describes such prompt refusal as useful for people who rely on commercial models and are not deeply familiar with AI technology.

The limitation is obvious. Open-weight models can be tuned to dodge limits, so refusals do not erase the problem. It is also too soon to know whether the rule barring AI writing is fully working. Editors see possible effects, including refusals and an apparent slight decline in flagged content.

Lebleu emphasizes that the vast majority of AI-generated submissions come from beginner editors. That is why the community has placed the rule in welcome messages, editing help, and advice.

From Cleanup Policy to Identity Debate

The guideline was adopted for practical reasons. Wikipedia had been flooded with AI content, editors lacked strong tools, and the community needed a way to stop it. Lebleu says the discussion has since gained a philosophical layer. Many people have described Wikipedia as “the last good place on the internet.” Keeping that name free from low-quality machine output has become part of the site’s identity.

That identity can also matter to AI companies, Lebleu adds, because they may see value in a corpus written only by people for training systems. For at least the near future, she says, Wikipedia remains a human-written resource not replicated anywhere else.

A Possible Middle Path: Human Identity, AI-Assisted Option

Lebleu raises a harder choice. If readers and contributors want a more comprehensive encyclopedia, would they accept partial AI generation? She suggests one path: keep Wikipedia’s identity as a human-written corpus while also offering an AI-assisted option.

This is not a detection question, but a design and communication question. It asks whether a platform can preserve one brand of quality while separating another workflow.

Google Zero, Attribution, and a Future Relationship

The cleanup conversation also touches Google Zero, a possible world where websites receive no traffic from Google search. If visibility collapses, Lebleu asks how Wikipedia can still show that it is editable and that people can contribute. Wikipedia is cited often in Google’s AI Overviews, but readers may see only an AI answer that mentions Wikipedia and not understand the living project behind it.

The Wikimedia Foundation is developing an attribution framework that could tell AI users when an article was last edited, how complete it is, whether editors have marked problems, and perhaps prompt them to act. In that setting, a reader might ask an assistant for a change based on a new source. Lebleu imagines: “The AI cited Wikipedia, but there seems to be an error. Can I make an edit request and add this other source I found?”

She sees a strong symbiosis between Wikipedia and AI models, not a disappearance. The project may take a very different shape. In her telling, yesterday AI was content to clean up; today it is a tool for cleanup; tomorrow it could be the surrounding environment.

Checklist: What Readers Should Understand

• Obvious phrase traces mattered in 2023, especially “stands as a testament,” “rich cultural heritage.”, and headings such as “Why is this important?”
• By 2025, structure became a stronger clue: repeated habits, the “rule of three,”, and output pace can raise suspicion.
• A source can be real and still not support the sentence next to it.
• Pangram numbers are estimates, not proof. In late 2025, the tool marked perhaps 20 percent of new articles and drafts; in early 2026, the share was 15 or 16 percent.
• Most suspected submissions appear to come from beginner editors, so onboarding guidance is part of enforcement.

FAQ: Quick Questions About the Cleanup Project

Why does WikiProject AI Cleanup matter to ordinary readers?

It shows how one of the most widely used reference sites is trying to keep its pages reliable when AI text can be fast, fluent, and plausible. The project’s concerns also connect to how Wikipedia appears through AI answers rather than through direct traffic.

Has English Wikipedia banned AI article writing?

Lebleu notes that English Wikipedia recently prohibited editors from using LLMs to generate articles. The cleanup project still tries to estimate what slips through.

Are fake citations the main citation problem now?

No. Lebleu says fabricated references are less common because retrieval-augmented generation can surface real sources, but misleading use of real sources remains a major worry.

Can AI tools help without fully solving detection?

Yes. Lebleu says an AI source verifier can filter obvious cases at scale, but because it is a model, its answers are not guaranteed. Paste Check and Pangram can assist, but they do not turn cleanup into a certainty.

Could Wikipedia use AI and still present itself as human-written?

One idea raised by Lebleu is to keep the human-written identity while offering a separate AI-assisted option, making the distinction clearer for readers and contributors.

Common Misconceptions

1. AI writing is always easy to spot. The Wikipedia cleanup story describes a shift from obvious ChatGPT-era phrases to structural patterns.
2. A real citation proves accuracy. Editors now worry more about sources being misused than purely invented.
3. Detection counts are complete counts. Pangram readings are treated as a cautious lower bound, and 15 or 16 percent is not a finished verdict.
4. Model refusals end the problem. Lebleu points out that open-weight models can be tuned to dodge limits.
5. Cleanup is only about bad text. It has become a debate about Wikipedia’s identity, beginner onboarding, attribution, and Google search visibility.

Action Notes

• Read fluency as a signal, not evidence. Repeated structure may be AI, but it may also be style.
• Check the link between citation and claim before trusting a polished paragraph.
• If you contribute to an encyclopedia-style project, learn the AI writing rules before starting a page; many suspected cases come from beginners.
• When an AI answer mentions Wikipedia, look for context about last edit, completeness, and flagged problems if available.
• Treat automated verification tools as triage, not final judgment, because verifier models can be wrong.

Closing Take for MyTechDailyNews Readers

The cleanup project is best understood as a small preview of a larger governance problem: platforms must define what counts as human, what counts as assisted, and how readers can tell the difference when information is consumed through AI answers. Lebleu’s arc, from Discord chats to Pangram estimates to Wikimedia Foundation tools, shows that the question is no longer only whether a machine wrote a page. It is how a human-led encyclopedia can survive the changes it may still be forced to review.