How Does Wikipedia Defend Its Human-Written Bottom Line? Inside the AI Cleanup Team

In late 2023, several Wikipedia editors noticed a batch of entries that had a "distinctive smell" on Discord: repeatedly appearing stock phrases like "stands as a testament" and "rich cultural heritage," often with paragraphs at the end like "Why is this important?" that didn't read like encyclopedia entries. These features pointed to the same source—ChatGPT. Thus, WikiProject AI Cleanup was born. This volunteer-initiated project has a clear goal: to identify and remove AI-generated content that violates Wikipedia's policies.

One of the founding members, Ellia Lebleu, is currently pursuing a PhD in AI at the Inria Rennes research center near Toulouse. She collaborates with AI detection company Pangram to try to quantify how much AI content has actually infiltrated. Her assessment: by the end of 2025, about 20% of new entries and drafts were flagged as AI-generated; by early 2026, this number dropped to 15% to 16%. But she herself is cautious—whether the decline is due to community policies taking effect or because models have become stronger and detection tools haven't kept up remains undetermined.

From "Obvious Fake" to "Increasingly Realistic": The Evolution of AI Writing

Lebleu recalls that by the end of 2023, AI traces were almost blatant: concluding paragraphs, formulaic rhetoric, mechanical three-part structures. But by 2025, the signals have become structural and more subtle. AI favors "triple parallelism" and negative parallel structures; these features alone mean nothing, but when densely stacked in long articles, they form a pattern worth vigilance.

Editors also look at signals beyond the text. For example, an account writing five entries of thousands of words in 5 minutes is more telling than any textual feature. In other words, detecting AI writing is not just reading text; it's also reading behavior.

Retrieval-Augmented Generation (RAG) refers to letting large language models retrieve relevant documents from external knowledge bases or search engines before generating answers, and then organizing content based on these documents. It significantly reduces the probability of the model fabricating facts out of thin air, but also brings new problems: the model may correctly find a source but misinterpret its meaning, leading to a mismatch between citation and claim. In the Wikipedia context, RAG reduces completely fabricated citations, but "real source, wrong interpretation" citations are more hidden and harder to investigate.

Citation Fabrication: A Trickier Problem Than the Body Text

AI contamination doesn't only occur in the body text. Lebleu points out that false or misleading citations have become a major concern. Completely fabricated citations are relatively rare today because retrieval-augmented generation allows models to find real sources. More commonly, the source is real, but the AI misunderstands its content, and the citation simply doesn't support the claim in the text. This type of "citation-claim mismatch" is extremely common and the hardest to check manually one by one.

Because of this, another editor in the project created an AI source verification tool to check whether citations and claims on pages match one by one. Lebleu's attitude toward this tool is pragmatic: it is itself an AI model and cannot guarantee correctness every time; but if the goal is to filter obvious problems at scale, it's a good starting point. The Wikimedia Foundation is also developing the Paste Check tool to flag text pasted directly from certain AI chatbots.

After the Ban: Refusals from Commercial Models and Loopholes in Open-Source Models

English Wikipedia has explicitly prohibited editors from using large language models to generate entries. Lebleu observes that Claude sometimes outright refuses: "I cannot write Wikipedia entries; this violates Wikipedia's policies." She believes that if all mainstream AI companies added such prompt refusals, it would be practically helpful—at least it would stop ordinary users who use commercial models and are not tech-savvy. But this is far from a panacea: open-weight models can be adjusted to bypass restrictions, and the problem persists.

Whether the ban is effective is still too early to judge. Lebleu mentions some possible effects: prompt refusals, a slight decline in the proportion of AI-generated content. But she also admits it's impossible to distinguish whether this is due to policy taking effect, models becoming stronger, or newcomers gradually learning the rules. The vast majority of AI-generated content comes from novice editors, so the community puts a lot of effort into making the rules "visible"—repeatedly emphasized in welcome messages, editing help, and various prompts.

Open-weight models refer to large language models whose parameters are publicly downloadable, locally deployable, and fine-tunable, such as the Llama series, Mistral, etc. Unlike commercial closed-source models that can only be accessed via API, open-weight models allow users to modify system prompts, fine-tune, or even remove safety alignment layers. Therefore, prompt constraints like "refuse to generate Wikipedia entries" built into commercial models are easily bypassed on open-weight models. This means the ban has some interception effect on ordinary users but poses almost no obstacle to technically capable users.

"The Last Good Place": Identity and the Possibility of Future Compromise

When the ban was passed, the reasons were very practical: AI content was proliferating, tools were insufficient, and brakes had to be applied. But since then, this rule has been given a more philosophical meaning. Many people call Wikipedia "the last good place on the internet," and "human-written" has gradually become part of the project's identity.

Lebleu offers an intriguing observation: AI companies themselves are interested in this purely human-written corpus, which has unique value for training AI systems. Wikipedia will at least in the foreseeable future maintain this attribute, making it a precious resource that cannot be replicated elsewhere.

But she also leaves open questions: if AI can make Wikipedia more comprehensive, would readers and contributors be willing to accept some AI-generated content? She envisions a compromise—retaining the core identity of "human-written" while offering an AI-assisted option. This is more of a topic for discussion than a determined path.

Wikipedia Under the Shadow of Google Zero

The interview also touches on a broader anxiety: if search traffic overall goes to zero, how can the website be seen? Lebleu points out that Wikipedia is often cited by Google's AI Overviews, but users may only see an AI answer that cites Wikipedia; they may not realize that Wikipedia is a living project that they themselves can edit and contribute to.

The Wikimedia Foundation is developing an attribution framework to try to let AI users see more information: the last edit time of an entry, content completeness, whether editors have flagged issues, and possibly even a call to action. Lebleu imagines that in the future, readers might be able to participate in contributions through AI models—"The AI cited Wikipedia, but there seems to be an error here. Can I submit an edit request with the source I found?"

She offers a phased assessment of the relationship between Wikipedia and AI models: yesterday, AI was something to be cleaned up; today, it is a tool to assist cleaning; tomorrow, it will be the air we breathe. Wikipedia will not disappear, but its form will become very different.

(This article is based on an IEEE Spectrum interview with Ellia Lebleu.)

Google Zero refers to a hypothetical scenario: search engines' AI summaries directly answer users' questions, causing users to stop clicking through to original websites, driving independent websites' search traffic to near zero. For Wikipedia, this means it may be heavily cited by AI Overviews but rarely receive direct visits and editorial participation. The Wikimedia Foundation is therefore trying to use an attribution framework to attach metadata such as entry edit time and content completeness to AI answers, and exploring paths for readers to submit edit requests directly from the AI interface, in order to maintain the project's visibility and contributor base.