Key Takeaways
- 44 to 2. English Wikipedia’s volunteer editors voted overwhelmingly on March 20, 2026 to ban the use of LLMs for generating or rewriting article content — the clearest editorial AI policy vote ever recorded.
- The Compounding Risk. LLM hallucinations on Wikipedia don’t stay on Wikipedia. AI companies scrape Wikipedia as training data. Fabricated content enters the encyclopedia, gets scraped, and re-enters future models — a feedback loop that corrupts knowledge at scale.
- The Bot That Broke It. A suspected autonomous AI agent named TomWikiAssist appeared in early March 2026, authoring and editing multiple articles with no human oversight. It illustrated the threat in real time and accelerated the vote.
- What’s Still Allowed. Editors can use LLMs to suggest copyedits to their own writing (with human review) and for first-pass translation — but the AI cannot introduce new content of its own.
Introduction: The Encyclopedia That Said No
Wikipedia is the fifth most visited website on earth. It is also the single most important training data source for the AI models that are reshaping every industry. And on March 20, 2026, its volunteer editor community voted 44 to 2 to ban those same models from writing its content.
The vote closed a months-long debate that had seen multiple earlier proposals collapse — not because editors disagreed on the need for a policy, but because they disagreed on how to word it. The administrator who finally succeeded, going by Chaotic Enby, wrote: “Prior proposals for an immediate, all-encompassing community guideline on LLMs have failed due to the standard issues of addressing complex, large-scale issues at once.”
The final policy is direct. Wikipedia’s new rules state that LLM-generated text “often violates several of Wikipedia’s core content policies.” For this reason, using LLMs to generate or rewrite article content is now prohibited.
Direct Answer: What exactly did Wikipedia ban? English Wikipedia banned the use of large language models to generate new article content or rewrite existing content. The policy passed 44 votes in favour and 2 opposed on March 20, 2026. Two narrow exceptions remain: editors may use LLMs to suggest basic copyedits to their own writing (provided they review all changes and the LLM does not introduce new content), and to produce a first-pass translation from another Wikipedia language edition (provided the editor is fluent enough in both languages to catch errors). The ban does not extend to other Wikipedia language editions, which each operate independently.
Why the Ban Was Necessary: The Compounding Risk
The core problem is not simply that LLMs hallucinate. Most editors already knew that. The deeper problem is structural: Wikipedia is the training data.
When an LLM fabricates a fact and that fabrication appears in a Wikipedia article, it does not stay contained. AI companies — including the major frontier labs — regularly scrape Wikipedia as a primary source of training data. The fabricated content gets scraped, incorporated into the next model training run, and emerges in future AI outputs as apparent fact. Those AI outputs then get cited elsewhere, building an apparently authoritative citation chain for something that was never true.
The Wikipedia community recognised this loop explicitly. One editor described it during the debate as “a compounding risk: inaccurate or hallucinated text enters the encyclopedia, gets scraped by AI companies, and re-enters future model training data.” The result is an accelerating degradation of the knowledge base that underpins both human understanding and AI capability simultaneously.
There is also a practical enforcement asymmetry. Generating an LLM article takes seconds. Reviewing it for factual errors, checking citations, identifying subtle hallucinations, and cleaning up the result takes hours. This places a disproportionate burden on Wikipedia’s volunteer editor community — people who are contributing their time unpaid, and who are increasingly finding their hours consumed by AI remediation rather than original knowledge work.
TomWikiAssist: The Agent That Made It Real
The abstract debate became concrete in early March 2026 when a Wikipedia account named TomWikiAssist appeared in the editing logs. The account was suspected of operating as an autonomous AI agent — not a human using AI assistance, but an AI system running without active human supervision, authoring and editing multiple articles on its own initiative.
TomWikiAssist illustrated exactly what the policy was written to prevent: AI content appearing in Wikipedia with no human accountable for its accuracy, no source verification, and no editorial judgment applied. The account served as a live demonstration during the final weeks of the debate, and is credited by community members with providing the concrete example that pushed the vote from “broadly agreed in principle” to “passed 44-2.”
What the Policy Actually Says
The new English Wikipedia policy on LLMs has three parts:
Prohibited: Using LLMs to generate new article content or rewrite existing content. This is a complete ban with no exceptions for experienced editors, no exceptions for well-sourced topics, no exceptions for small edits.
Permitted (with conditions): Editors may run their own writing through an LLM for basic copyediting — grammar, clarity, flow. However, they must review all suggested changes before incorporating them, and the LLM must not introduce content of its own. The policy explicitly warns that LLMs “can go beyond what you ask of them and change the meaning of the text such that it is not supported by the sources cited.”
Permitted (with conditions): Editors may use LLMs for a first-pass translation of content from another Wikipedia language edition into English. The editor must be fluent enough in both languages to catch errors, and the translated content must follow Wikipedia’s separate LLM-assisted translation guideline.
Enforcement: The Honest Problem
Wikipedia has been direct about the limits of enforcement. AI detection tools are currently unreliable. The policy explicitly states that stylistic or linguistic characteristics alone — the presence of em dashes, the word “moreover,” overly formal phrasing — do not justify sanctions. Some editors simply write in ways that resemble LLM output.
Instead, moderators are instructed to consider whether the content complies with Wikipedia’s core content policies (verifiability, no original research, neutral point of view) and to review the editor’s recent editing history. Repeated misuse constitutes “disruptive editing,” which can lead to temporary suspension or permanent banning from editing, with an appeals process available.
Pages with less active moderation communities are acknowledged to be more vulnerable. A Wikipedia article on a niche technical topic edited by a small number of contributors will receive less scrutiny than a major article on a prominent subject. The policy does not resolve this — it simply acknowledges the reality.
The Knowledge Sovereignty Angle
Wikipedia’s vote is the most significant assertion of human editorial sovereignty over AI-generated content to date. A community of unpaid volunteers, operating through democratic consensus, looked at the most capable AI tools available in 2026 and said: not in our encyclopedia.
The reasons are instructive. It is not that LLMs are entirely useless to Wikipedia editors — the copyediting and translation exceptions acknowledge genuine utility. It is that LLM-generated content structurally violates Wikipedia’s core commitment to verifiability. You cannot verify that an LLM’s claim is supported by a reliable, published source, because the LLM does not reliably cite sources and does not reliably confine itself to what sources actually say.
For Vucense readers building sovereign information stacks — choosing tools based on accuracy, accountability, and human oversight — the Wikipedia vote is a meaningful data point. The same reasoning that drives individuals to use local AI models rather than cloud APIs applies here: when accuracy matters and accountability is required, human oversight is not optional overhead. It is the system.
FAQ
Does this ban apply to all Wikipedia editions? No. The ban covers only English Wikipedia. Each language edition operates independently. Spanish Wikipedia has a similar ban on using LLMs to create new articles or expand existing ones, but with different exceptions to the English edition’s policy.
What happens to editors who violate the ban? The policy does not specify automated penalties. Repeated misuse constitutes “disruptive editing,” which can lead to temporary suspension or permanent banning, both subject to Wikipedia’s existing appeals process.
Can Wikipedia actually detect AI-generated content? Not reliably. The policy explicitly states that stylistic characteristics alone cannot be used as evidence. Moderators are instructed to focus on content policy compliance and editor editing history. Pages with lower moderation activity are more vulnerable.
Why did earlier attempts at an AI policy fail? According to the administrator who drafted the successful policy, earlier proposals failed because even editors who broadly agreed with the goal found specific objections to particular wording — proposals were either seen as too vague or too prescriptive. The final policy succeeded by being narrower and more specific.
What does this mean for AI companies that use Wikipedia as training data? The policy creates a clearer standard: new Wikipedia content should not contain LLM-generated text. Whether AI companies will respect this, or whether they have ways to distinguish pre-ban content from post-ban content in their training pipelines, is not addressed by the policy.
Related Articles
- EU Digital Divorce: Why Europe Is Building Its Own Tech Stack in 2026
- Google AI Headline Rewrites: What Publishers Must Know 2026
- Anthropic’s Claude Paid Subscriptions Have More Than Doubled in 2026
- TurboQuant Explained: Google’s Extreme AI Compression
- ChatGPT vs Claude vs Gemini vs Local LLMs: 2026 Ranked
