Wikimedia asks court to draw a line between AI training and retrieval

The Wikimedia Foundation has filed a brief in the New York copyright case involving OpenAI and Microsoft, arguing that AI training and systems that retrieve and reproduce source material should be treated differently. Its position puts the sustainability of open knowledge squarely into a major fight over how AI companies may use copyrighted work.

OpenAI Watch analysis

What happened

In an amicus brief filed in the Southern District of New York, Wikimedia argues that using copyrighted works to train a language model should qualify as fair use when the model creates something new and different from its sources. It says that use should not require a licence.

For retrieval-augmented generation (RAG) and similar grounding techniques, Wikimedia takes the opposite view: when a service finds and delivers source material to users, it should respect publishers’ licensing terms and provide proper attribution. The foundation also calls for scraping to follow its public API rate limits or use its Enterprise API. Wikimedia Foundation’s account of the brief

Why it matters

That distinction is practical as well as legal. Training uses material to build a model; RAG can bring material from the web into a chatbot answer. Wikimedia’s proposal would allow some training without a licence while requiring permission and credit when content is retrieved and presented back to users.

The court has not adopted this position. The foundation is asking it to consider a middle ground: broad room for AI development, but clearer obligations when systems serve source material directly. The case could influence how other courts approach copyright and AI.

Our read

Wikimedia is trying to avoid two unsatisfying extremes: rules so restrictive they choke off useful AI development, and rules so permissive they weaken the sources those systems depend on. That is a consequential argument, not a ruling. The important question is whether the proposed distinction can be made clearly enough to work in real products.

What to watch

• How the court treats the distinction between model training and RAG.
• Whether the parties address Wikimedia’s proposed approach in later filings.
• How attribution, licensing and API limits are handled as the case proceeds.

Discussion spark: Should AI companies be able to train on copyrighted material without a licence if retrieval systems must license and credit what they return, or is that distinction too hard to enforce?

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.