English Wikipedia Prohibits Editors From Using AI To Protect a 'Purely Human-Written Corpus'

English Wikipedia's community has adopted a formal ban on using generative AI to write articles, a rule volunteer editor Ellia Lebleu told IEEE Spectrum is meant to preserve a "purely human-written corpus." Lebleu, a founding member of the 2023-launched WikiProject AI Cleanup, said AI-detection company Pangram flagged about 20 percent of new articles and drafts as AI-generated in late 2025, falling to 15 or 16 percent in early 2026, a figure she called a "very conservative lower bound." The Wikimedia Foundation is developing a tool called Paste Check to flag text pasted from some AI chatbots, while a volunteer-built AI source verifier checks whether cited sources support article claims.

English Wikipedia Prohibits Editors From Using AI To Protect a 'Purely Human-Written Corpus' The Wikipedia AI ban stops English Wikipedia editors using large language models to write articles, and a volunteer explains what is still slipping through English Wikipedia has introduced a formal ban on using generative AI https://www.ibtimes.co.uk/openai-contractors-removed-using-ai-1822365 to write its articles, in a move supporters say is vital to preserve what one volunteer calls a ' purely human-written corpus https://www.ibtimes.co.uk/australia-ai-firms-creative-work-opt-out-plan-1826606 '. The prohibition, recently adopted by the site's English-language community, follows growing concern among editors about machine-written entries slipping into the online encyclopaedia. Volunteer editor Ellia Lebleu told IEEE Spectrum that keeping Wikipedia human-written is worth protecting. Lebleu is a founding member of WikiProject AI Cleanup, a volunteer effort launched in 2023 to find and remove AI-generated material that breaks the site's guidelines. How the Wikipedia AI Ban Came About The cleanup project began after Lebleu and other editors compared notes on Wikipedia's Discord server around December 2023. They reported seeing multiple articles that shared similar stylistic patterns. Some repeated phrases such as 'stands as a testament' or 'rich cultural heritage'. Others contained conclusion sections or headings such as 'Why is this important?'. Lebleu said this was typical of ChatGPT output in 2023 and early 2024, and described it as easy to identify. She then proposed creating an official project. By 2025, Lebleu said, the indicators of AI involvement were 'subtler and more structural', including the 'rule of three' and negative parallelisms. She said no single feature is conclusive, but that they can be taken together. She added that behaviour is also taken into account, citing the example of an editor creating five articles of several thousand words each within five minutes as giving 'reason to suspect AI'. According to Lebleu, outright invented references are now much rarer because models can retrieve real sources. She said, however, that content based on real sources can still be 'completely wrong', and that citations which do not support the text are 'extremely commonplace'. Is the AI Ban Actually Working? Lebleu has worked with AI-detection company Pangram to estimate how much new material on English Wikipedia is machine-written. She said Pangram's system produced no false positives when tested on pre-ChatGPT Wikipedia text and described its figures as a 'very conservative lower bound'. According to Lebleu, Pangram flagged about 20 per cent of new articles and drafts as AI-generated in late 2025. She said that in early 2026 this proportion fell to 15 or 16 per cent. Lebleu said the decrease could be linked to the new rules, or could indicate that AI models are improving faster than detection tools. She said most submissions identified as AI-generated come from new editors, and that information about the rule now appears in welcome messages, editing help and advice. Lebleu said the Wikimedia Foundation is developing a tool called Paste Check, which is intended to flag text pasted directly from some AI chatbots. She said another project editor has built an AI source verifier https://www.ibtimes.co.uk/ai-researcher-challenges-black-box-view-language-model-audit-1826618 designed to check whether cited sources support the claims made in an article. She said that because the verifier is itself an AI model, 'there is no guarantee that it will be correct every time', but described it as a first step towards filtering out clear cases at scale. According to Lebleu, the AI system Claude https://www.ibtimes.co.uk/hidden-claude-code-ao3-ai-dispute-1822141 has in some instances responded that it cannot write a Wikipedia article because this would go against Wikipedia's policies. She said open-weight models, which can be adjusted to bypass such restrictions, could still be used, but that refusals from systems like Claude may deter people who 'aren't especially familiar with the technology'. Why a Human-Written Wikipedia Matters Beyond the Ban Lebleu said the rule was initially introduced as a practical measure. She said editors were receiving large volumes of AI-generated submissions, did not have suitable tools to manage them and sought to limit this material. She said the issue has since taken on a more philosophical dimension. According to Lebleu, some commentators have described Wikipedia as 'the last good place on the internet', and she said avoiding what she termed ' AI slop https://www.ibtimes.co.uk/ai-doom-loop-workplace-productivity-concerns-1826482 ' has become part of how some editors see the project. She said some AI companies regard a purely human-written corpus as valuable for training their systems https://www.ibtimes.co.uk/microsoft-openai-ai-training-creators-copyright-case-1821309 , and said she expects Wikipedia to remain human-written 'for at least the near future'. Lebleu also raised the question of whether readers and contributors would accept a partly AI-generated Wikipedia if it provided more comprehensive coverage. She suggested one possible model would be to retain a human-written core and offer an AI-assisted option alongside it. Lebleu said Wikipedia content is frequently cited in Google's AI Overviews, but that users may not always realise Wikipedia is a project they can edit. She said the Wikimedia Foundation is developing an attribution framework intended to show when an article was last edited, how comprehensive it is and whether editors have highlighted any issues. She said readers might in future be able to submit edit requests through AI systems. 'I don't see Wikipedia disappearing, but I can see it taking a very different shape,' she said. 'Yesterday, AI was something to clean up. Today, it's a tool to help us with the cleanup. Tomorrow, it will be the air we breathe.' © Copyright IBTimes 2026. All rights reserved.