Short answers, pulled from the story.
Wikipedia has allowed approved bots to edit its pages since 2002. Early programs like rambot converted census data into local geography articles, and the 2010 ClueBot NG applied machine learning to vandal detection. The Objective Revision Evaluation Service, known as ORES, launched in late 2015 to score the quality of individual edits.
English Wikipedia looks for fabricated citations, particular phrases such as "Here is your Wikipedia article on" or "Up to my last training update," heavy use of em dashes, and the word "breathtaking" in otherwise neutral prose. The community published a formal guide to these patterns in August 2025.
In March 2026, English Wikipedia's community voted to prohibit using large language models to add or rewrite article content. The decision followed years of concern about AI tools producing misinformation, fabricated citations, and non-encyclopedic prose. Two exceptions were preserved: copyediting one's own writing and machine translation from other language versions of Wikipedia.
As of 2023, Wikipedia's text corpus ranked among the largest well-curated data sets available for training AI models. Stephen Harrison reported that every major large language model developed to that date had trained on it. The Wikimedia Foundation expressed concern that many AI products use this text without the attribution Wikipedia's licensing terms require.
A Princeton University study published in October 2024 found that about 5 percent of roughly 3,000 newly created English Wikipedia articles were written using AI. Those articles had all been created in August 2024. Some covered ordinary topics with only minor machine assistance; others had been used to promote businesses or political interests.
Jimmy Wales proposed in 2025 that AI support the editing process rather than write article text. He described specific uses: tailored feedback on rejected drafts, spotting inconsistencies, flagging missing information, and summarizing discussions for editors joining a debate mid-stream.