Wikipedia’s operator says a bad AI copyright ruling could threaten the open web

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. Its argument is that both camps are pushing fair use toward blunt, all-or-nothing outcomes, and that either extreme would hurt the open web.

This matters because Wikipedia sits on both sides of the AI bargain. Its pages are among the most important training data in the field, and AI answers pull visitors away from the volunteer communities that write them. Yet the same models help its editors catch vandalism, copyright violations and doxxing. When the institution with the most to lose and the most to gain asks for nuance, judges should listen.

We think Wikimedia has drawn the most useful line anyone has offered in this fight. Learning from text is one act. Reproducing text in place of the original is another. Treating them as the same thing, in either direction, leads somewhere bad.

A clean win for publishers could let rights holders gatekeep genuinely transformative uses, including the open-source tools that a nonprofit like Wikipedia depends on. A clean win for the AI companies could, in Wikimedia's framing, become a green light to ignore licensing terms entirely. Wikipedia's content is free, but it isn't free of conditions: its Creative Commons license requires attribution and that adaptations stay under the same license. Wikimedia says chatbots reproducing that text aren't honoring those terms.

The proposed compromise is specific. Scraping shouldn't count as infringement. Pre-training without a license should be fair use when the result is something new. But when a model retrieves copies at query time and delivers them in a way that replaces a reader's engagement with the source, that should require a license. That targets the exact behavior that drains traffic from publishers and volunteer projects alike, without kneecapping training itself.

The downside is that the line will be hard to police. OpenAI calls regurgitation a rare bug, and Microsoft points to expert analysis saying Copilot seldom reproduces full sentences from news used to ground it. 'Rare' at the scale of hundreds of millions of queries is still a lot of substitution, and a court will have to decide how much is too much. If this standard is adopted, I'm watching for a new industry of attribution and licensing at the retrieval layer rather than the training layer.

Where this leads if nobody draws the line: the commons that feeds AI gets thinner. Fewer visitors means fewer editors, and fewer editors means worse source material for the models themselves. Protecting the people who write the knowledge is not anti-AI. It's maintenance.

The Times alleges OpenAI infringed its copyright, including by scraping its site to build training datasets. OpenAI and Microsoft argue their use of the material is fair use. Wikimedia's brief, filed Friday, is a friend-of-the-court filing, so it does not make the foundation a party to the case.