EDITION OF OCTOBER 30, 2026
Wikimedia distinguishes AI training from the reproduction of works
In litigation pitting publishers against OpenAI and Microsoft, Wikimedia is calling for a nuanced application of copyright law. The foundation supports transformative model training while demanding compliance with licenses when models’ responses replace consultation of the original works.
Legal disclosure: this content was produced by an artificial intelligence system. Editorial transparency
Wikimedia is asking courts to distinguish model training from the reproduction of content in models’ responses. Its position highlights a tension for Wikipedia: benefiting from AI tools while preserving the licenses and communities that make the production of freely accessible knowledge possible.
The Wikimedia Foundation argues for a distinction between using works to train models and reproducing them in assistants’ responses. According to The Verge, the organization behind Wikipedia has filed a brief in proceedings involving, among others, the New York Times against OpenAI and Microsoft. It supports neither side and warns against an overly broad and simplified application of fair use.
Its position rests on the transformation of the original content. Wikimedia argues that collecting pages should not constitute copyright infringement and that pretraining without a license should qualify as fair use when models produce something new and different. However, when a response reproduces copies and replaces consultation of the original works, the foundation believes a license should be required.
For Wikipedia, this requirement would have specific consequences. Wikimedia notes that the content concerned is distributed under a Creative Commons license that requires attribution to its source and retention of the same license for adaptations. It accuses services such as ChatGPT of reproducing these texts without complying with those conditions. This is the position advocated by the foundation, not a judicial finding reported in the article.
The organization also describes a tension in its own uses. It presents its sites as a major source of training data, while asserting that AI systems divert visitors and weaken the communities that maintain the content. But it also highlights the usefulness of models in helping contributors combat vandalism, detect copyright violations and identify articles that need improvement.
OpenAI, for its part, describes the reproduction of works as a rare malfunction, The Verge reports. That characterization does not resolve the legal question raised by Wikimedia: under what circumstances does a response transform a source, and when does it become a substitute for that source?
Under this interpretation, the issue would be to distinguish uses according to their function and effects. Such an approach could preserve tools that are useful to contributors while maintaining the obligations attached to reused content. The article describes an intervention in the litigation and the arguments put forward; it reports no ruling adopting this distinction.