In March 2026, English Wikipedia's editors voted 44 to 2 to restrict one specific tool from writing article text: large language models. The resulting guideline, WP:LLM, does not forbid using one at all — a reviewed copyedit that adds no new content still passes — but it draws a hard line at letting a model generate or rewrite the substance of an article, footnotes included.
The vote followed two years of cleanup work by a volunteer group called WikiProject AI Cleanup, formed in 2023 around one recurring failure: articles that read fluently, cite sources confidently, and cite sources that don't say what the text claims — or don't exist at all. Ask a model to write about a minor public figure, a startup, or a regional musician, and it will often produce a tidy paragraph with a number, a date, and a footnote attached to both. The paragraph is usually well formed. The footnote is sometimes invented.
If citations are generated as part of the output, editors must verify that each source is real, reliable, relevant, and suitable, and check that the source actually supports the text it's attached to.
Wikipedia's verifiability policy has never asked whether a claim is true. It asks whether a reader can check it against a source that says so. A fabricated citation fails that test more completely than a missing one: a missing citation is an honest gap a reviewer can flag and wait on, while an invented one looks closed until someone opens the source and finds a page that never mentions the subject, or a link that resolves to nothing. Since August 2025, a page suspected of being AI-generated can be speedily deleted instead of routed through the week-long discussion a contested article would normally get. A draft with invented sources doesn't receive the benefit of the doubt a thin one would.
The guideline also undercuts the obvious shortcut. It tells reviewers not to judge a text by how it reads — the smoothness, the stock phrases, the slightly-too-balanced tone are unreliable tells on their own. What holds up is whether the material complies with the core content policies, and whether the account's editing history looks like a person doing research or a script producing pages. A fluent draft earns no extra trust for being fluent. It earns the same scrutiny of its actual citations that a clumsy one would.
That is worth knowing before handing a Wikipedia submission to a model because the output looks complete. The same rule that already holds a client responsible for a ghostwritten paragraph holds an editor fully responsible for an AI-assisted one — the tool absorbs neither the disclosure nor the accountability, and it can manufacture confident prose faster than it can manufacture a source that actually exists. A model can help organize research a person has already verified. It cannot do the one thing the whole guideline is built around: produce a citation that still holds up once somebody, eventually, checks it.
