Wikimedia Urges Court to Separate AI Training From Content Retrieval

The Wikimedia Foundation is asking a court in an OpenAI copyright case to treat AI model training differently from systems that retrieve and present specific copyrighted works, arguing the distinction could affect both open knowledge and publishers.

The foundation filed an amicus brief Oct. 30 in In re: OpenAI, Inc. Copyright Infringement Litigation. An amicus brief comes from an organization outside the case that wants to explain how a court decision could affect people or groups beyond the parties. Wikimedia’s filing takes neither side in the litigation.

The foundation’s argument turns on how an AI system uses copyrighted material. During pretraining, a model learns patterns from a large body of text. Wikimedia argued that this process typically qualifies as fair use.

Retrieval-augmented generation, or RAG, works differently: a system retrieves particular material and includes it in a response. Wikimedia argued that this kind of use should be subject to licensing requirements. The distinction would separate learning from a broad collection of works from presenting selected content in answers.

The foundation warned of risks on both sides. A ruling that restricts AI training too broadly could curb tools that help volunteers improve Wikipedia, it said. A decision that permits too much use without licensing could lead publishers to limit access to their work.

Wikipedia contains more than 67 million articles in over 300 languages. The foundation’s warning puts the possible effects of the copyright dispute in the context of a large, freely available knowledge resource, as well as the publishers whose work AI systems may retrieve.

The filing follows the foundation’s 2023 response to the U.S. Copyright Office, which called for case-specific reviews of AI fair use.