Wikimedia: AI Training Can Be Fair Use, but Reproducing Text in Answers Must Follow Licensing Terms

The Wikimedia Foundation, which operates Wikipedia, announced on October 30 (local time) that it had filed an amicus brief with a US court on the question of how AI systems may use copyrighted works. The filing addresses 12 consolidated lawsuits, including the dispute brought by The New York Times against OpenAI and Microsoft. The foundation urged the court to treat two things separately: training an AI model on copyrighted works, and reproducing those works verbatim in an AI’s answers.

How training and answering differ

The foundation argued that when a model built by training on copyrighted works amounts to something new and different from the originals, fair use can apply. Fair use is the legal standard that permits certain uses of copyrighted material without the rights holder’s permission. By contrast, the foundation’s position is that the protection should not extend to a chatbot showing a user a copyrighted work verbatim.

It also stressed that systems which answer by consulting retrieved material must respect the licensing terms attached to the source. If an AI pulls in outside content while composing an answer and copies it directly, the conditions attached to that content should apply. The foundation proposed drawing a line between whether material was used for training and whether it was reproduced as-is in an answer.

What reusing Wikipedia requires

When Wikipedia content is reused, the foundation said, the original authors must be credited and the Creative Commons Attribution-ShareAlike 4.0 license (CC BY-SA 4.0) must be honored. That means that when content is modified and shared again, the same licensing terms must travel with it. The foundation argues that a chatbot drawing on Wikipedia content in its answers should likewise respect the source and its terms of use.

The foundation also set conditions on how content is collected in bulk. Its position is that users should respect the request limits of the official channel for obtaining data, or else use the enterprise-grade access option. According to the foundation, Wikipedia hosts more than 67 million articles written in 300 languages. Roughly 250,000 volunteers contribute to building that knowledge base each month.

Why tilting too far either way is a problem

The dispute began when The New York Times sued OpenAI and Microsoft in 2023, alleging that its works had been used without permission. OpenAI maintains that training AI on publicly available information qualifies as fair use. The Wikimedia Foundation said its brief is not intended to take the side of any party in the litigation.

The foundation warned about what stands to be lost if copyright rulings lean too far in either direction. An overly strict ruling, it explained, could chill open-source AI development and make it harder for Wikipedia editors to use AI tools that are genuinely useful to them. Conversely, allowing the use of copyrighted works too broadly could shrink the very body of source material Wikipedia relies on, such as journalism and research.

The issue at stake is a standard of use that can protect both the AI tools people rely on and the material those answers are built from. What to watch next is how the court distinguishes AI training from content reproduction, and how it applies fair use and licensing terms to each.