Wikimedia’s OpenAI Copyright Brief Draws a Line Between Training and RAG

The Wikimedia Foundation filed a brief on October 30 in the consolidated copyright litigation involving OpenAI and Microsoft. It argues that transformative model training can qualify as fair use, while retrieving and reproducing copyrighted material in chatbot answers should respect licensing requirements. This is an argument to the court, not a ruling.

For teams building AI search and knowledge products, the useful distinction is between what happens during model development and what happens each time someone asks a question.

What Wikimedia is asking the court to consider

The Foundation’s announcement describes an amicus brief, a submission from an interested party rather than one of the litigants. The consolidated case includes The New York Times’ lawsuit. Wikimedia says it supports neither side.

Its concern runs in two directions. Making training dependent on expensive licensing could disadvantage open-source development. But allowing every AI use under a blanket fair-use interpretation could weaken the publishers and contributors whose work makes the open web useful.

Training permission and answer-generation permission are separate questions in Wikimedia’s argument. That does not settle either question under existing law, but it makes the proposed boundary easier to understand.

Why retrieval changes the discussion

Retrieval-augmented generation, or RAG, brings external material into an AI system when it answers a question. Unlike training, this can involve fetching current information for a particular request.

In his explanation of the filing, Wikimedia General Counsel Stephen LaPorte says products that retrieve and reproduce copyrighted content should honor the source’s license. Wikipedia uses CC BY-SA 4.0, with attribution and share-alike conditions.

This should not be flattened into a claim that all RAG is unlawful. Wikimedia is presenting a position about the use of copyrighted material, not announcing a technical ban on retrieval or a judicial outcome.

What this means for AI product design

A useful product review would trace the whole path: where information comes from, what gets retained, what reaches the user and how the original source is represented. A model’s training-data policy alone cannot describe that entire workflow.

For example, a company might use the same model in two products: an internal writing assistant and a public answer engine that retrieves publisher articles. The underlying model may be identical, while the material handled and the experience delivered are quite different. Treating both as one generic “AI use” hides the part that needs examination.

The practical question to watch is whether courts develop distinctions that map onto these stages. If they do, source handling will deserve the same product attention as retrieval quality and answer accuracy.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams explain complex AI products and the decisions behind them through clear technical stories.

Book a 30-minute call.