Wikipedia Bans AI-Generated Content With Exceptions

Wikipedia has adopted a strict new policy that prohibits the use of large language models to generate or rewrite article content on the English Wikipedia, which contains 7.1 million articles. The prohibition is categorical for drafting and rephrasing article text, with two narrowly defined exceptions: AI-assisted translations and AI-suggested basic copyedits, both requiring human verification. The community vote backing the change was decisive and reflects widespread editor concern about factual errors and fabricated sources produced by models like ChatGPT.

The policy targets core failure modes of current LLMs: hallucinations, invented citations, and opaque provenance. Editors may use LLMs only in two scenarios: translations where a bilingual human verifies fidelity to the original sources, and non-substantive copyedits to an author s own prose, where the LLM must not introduce new factual material. 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," and it disallows relying on stylistic detection as a primary enforcement mechanism because style signals are unreliable.

• •Permitted uses: AI-assisted translation (with bilingual human review), and basic grammar or format suggestions to an editor s own draft.
• •Prohibited uses: generating new articles, rewording existing articles, synthesizing content that cannot be linked to reliable sources, or inserting material that amounts to original research.
• •Verification requirements: humans must apply suggested edits manually, confirm citations, and ensure neutrality and verifiability before publishing.

This policy is a defensive response to two industry-level problems: the proliferation of machine-generated text across the web and the risk of an "AI ouroboros" in which models ingest other models output, amplifying errors. Wikipedia occupies a special role in the web information ecosystem and in training data pipelines for many models, so a human-first mandate is a purposeful attempt to keep that canonical corpus grounded in verifiable sources. The decision aligns with similar moves by publishers and some platforms to restrict or label AI-generated content, but it is stronger because it removes drafting privileges entirely for LLMs rather than merely requiring disclosure.

Dataset curators, search engineers, and model trainers rely on Wikipedia as high-quality ground truth. The ban clarifies that English Wikipedia intends to remain a human-curated source, which affects downstream data hygiene and sampling decisions. For developers building tools for editors, the permitted use cases define the narrow, defensible APIs that could be offered: translation assistants with traceable source alignment and copyedit suggestion UIs that require click-to-apply workflows. For researchers, the policy highlights the continuing gap between LLM fluency and reliable citation-level provenance.

Jimmy Wales has publicly warned that current models are "nowhere near good enough" for drafting encyclopedia content, a sentiment echoed in the community s vote to prioritize verifiability and neutrality over automation gains.

Implementation details and enforcement will determine practical impact. Key variables include whether other language editions adopt the same rule, how the Wikimedia Foundation supports detection and audit tooling, and whether dataset curators tag or filter AI-origin content to prevent contamination of future training corpora.