Wikimedia’s brief in the Times-OpenAI case: training can be fair use, substitution shouldn’t be

The Wikimedia Foundation filed an amicus brief in the cluster of cases where publishers including The New York Times are suing OpenAI and Microsoft. It backs neither side, arguing both camps push fair use toward all-or-nothing outcomes that would hurt the open web. As a friend-of-the-court filing, it doesn’t make the foundation a party. Wikipedia sits on both sides of the AI bargain: its pages are among the most important training data, and AI answers pull visitors from the volunteers who write them, yet models help editors catch vandalism, copyright violations and doxxing. The Times alleges OpenAI infringed, including by scraping to build training datasets; OpenAI and Microsoft argue fair use, backed by the administration (administration brief) and allied briefs (fair use briefs).

Learning from text and reproducing it in place of the original are different acts. A clean publisher win could let rights holders gatekeep transformative uses, including open-source tools a nonprofit like Wikipedia depends on, and could chill open models. A clean AI-company win could, in Wikimedia’s framing, greenlight ignoring licensing terms entirely. Wikipedia’s content is free but not condition-free: its Creative Commons license requires attribution and that adaptations stay under the same license, and Wikimedia says chatbots reproducing its text aren’t honoring that. The compromise targets the behavior that drains traffic from publishers and volunteer projects without kneecapping training, the dynamic unsealed filings described as a “doom loop” (doom loop).

OpenAI calls regurgitation a rare bug, and Microsoft points to expert analysis that Copilot seldom reproduces full sentences from grounding news (Copilot matches). “Rare” at hundreds of millions of queries is still a lot of substitution, and a court must decide how much is too much. If adopted, expect attribution and licensing to move to the retrieval layer. Without a line, the commons thins: fewer visitors, fewer editors, worse source material for models. Protecting the people who write the knowledge isn’t anti-AI; it is maintenance. More in our OpenAI coverage, Microsoft coverage and regulation coverage.