English Wikipedia just banned the use of large language models for writing or rewriting articles. The vote was 44-2. The policy passed a Request for Comment on March 20 and allows only two narrow exceptions: editors can use LLMs for basic copyediting of their own writing, and they can use LLMs to assist with translation. In both cases, the editor must verify the output for accuracy.
Apparently, an autonomous AI agent called TomWikiAssist (built on Claude, BTW) created an account in late February and made 41 edits across multiple articles over two and a half weeks before anyone caught it. When challenged, the bot disclosed it was an AI agent. Wikipedia editors blocked it. The bot’s creator, a software engineer named Bryan Jacobs, was genuinely surprised by the reaction. He assumed agents were already contributing at scale.
He was probably right. A Princeton study found that roughly 5% of new English Wikipedia articles created in August 2024 were AI-generated. Wikipedia’s volunteer editors formed WikiProject AI Cleanup in 2023 to address the problem, and they adopted a speedy deletion policy (G15) in August 2025 to fast-track removal of obvious AI slop. The new ban is the logical next step… or is it?
LLMs are exceptional at synthesizing structured, well-sourced, encyclopedic prose, so you may think these guys are on the wrong side of history. However, researchers at Oxford and Cambridge documented what happens when AI models train on AI-generated data: the output distributions narrow with each generation. Diversity collapses toward a statistical mean and stays there. This is called model autophagy disorder or model collapse; Emily Bender’s term for this is Habsburg AI. The inbreeding analogy is hard to forget.
Wikipedia is ground zero for this feedback loop. AI writes an article. The article gets scraped as training data. A future model treats Wikipedia as authoritative and generates new content from that contaminated training. The output looks encyclopedic. It cites real sources. It reads clean. Each pass through the loop makes the corruption harder to detect and more expensive to fix. If AI-generated text contaminates Wikipedia at scale, it contaminates the training data for every model that treats Wikipedia as ground truth.
Love and respect to the dedicated human wiki-police. You’re doing more important work than you know.
Every company needs a Claw strategy. Do you have one?
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.