Skip to content
Neural Daily

PolicyNews

Wikimedia in NYT v OpenAI: training yes, RAG needs a license

🇭🇺 Olvasd el magyarul →

The Wikimedia Foundation, which runs Wikipedia, filed an amicus brief in the New York Times copyright case against OpenAI. It argues LLM training can be fair use, but RAG that copies sources back into answers should need a license.

By: Mikola Máté Patrik · · 2 min read

Close-up of a bronze Lady Justice statue holding scales
Image: Wikimedia Foundation · Press kit · Image from the Wikimedia Foundation announcement · Source

What happened

On October 30, the Wikimedia Foundation filed an amicus brief with the Southern District of New York in In re: OpenAI, Inc. Copyright Infringement Litigation, which combines 12 lawsuits including The New York Times' case against Microsoft and OpenAI (Wikimedia Foundation). The court will rule on whether AI companies' use of copyrighted content to develop large language models is fair use. Wikimedia takes neither side.

What Wikimedia proposes

  • Model training: using copyrighted content to train LLMs that produce something new and different from the sources should be fair use, with no licensing requirement.
  • RAG and similar grounding techniques: copying web content and delivering it back through a chatbot should not be fair use. Companies should respect each source's license; for Wikipedia that is CC BY-SA 4.0, which requires attribution and share-alike.

According to The Verge, Wikimedia worries that a black-and-white ruling could hurt either way (The Verge). A clear publisher win could let rights holders gatekeep transformative uses, including AI tools that help Wikipedia editors fight vandalism and detect copyright violations or doxxing. A clear AI company win could be a green light to ignore licensing terms.

Why it matters

With more than 67 million articles, Wikipedia is one of the most important LLM training sources, while AI answers pull visitors away from the community that maintains it. Few courts have ruled on this question, so this decision will carry weight in later cases. Separating training from RAG would also have practical consequences: AI products built on search and retrieval would need to license what they return in their answers.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles