Wikimedia Asks the OpenAI Copyright Court to Separate AI Training From Retrieval

The filing proposes different fair-use analysis for pre-training and retrieval-augmented generation. It is an amicus request, not a court ruling; ECF No. 2361 does not itself resolve whether leave will be granted.

The Wikimedia Foundation has asked for permission to file an amicus brief in In re: OpenAI, Inc. Copyright Infringement Litigation, the Southern District of New York copyright MDL that includes The New York Times litigation against OpenAI and Microsoft.

Wikimedia's October 29 filing, ECF No. 2361, asks to submit a proposed brief in support of neither party. The immediate procedural point matters: the filing itself does not grant leave or adopt Wikimedia's proposed fair-use analysis. It is a request to be heard, not a decision about AI training, retrieval, or copyright liability.

Still, the brief offers a useful map of a distinction likely to matter across AI copyright disputes: whether a system is using copied material during model development or retrieving and delivering material to a user at runtime.

The proposed distinction is use-by-use

Wikimedia argues that fair use should be assessed separately at different stages of an LLM-based chatbot's development and operation. Its proposed brief describes that as a use-by-use inquiry rather than a single answer for every copy made in building or operating an AI product.

On Wikimedia's account, the pre-training use of copied works is likely fair use when it helps create a model that is different from the source works. That is an argument in the proposed brief, not a holding. It does not establish that a court will treat any particular training dataset, acquisition method, or model output as fair use.

The proposed brief takes a different position on retrieval-augmented generation, or RAG. It describes RAG and similar grounding techniques as systems that can retrieve and copy material from the web to answer a user's request. Wikimedia argues that this use should receive its own fair-use analysis and is likely not fair under the four-factor test when it displaces access to the source material.

That distinction does not turn on whether the source is available online. It is about the particular use at issue. The filing argues that a court should not treat every technical step involved in an AI system as legally identical merely because the same works may be involved.

Why Wikimedia says the split matters

Wikimedia's request connects the proposed analysis to the economics and governance of open knowledge. The organization says its projects depend on Creative Commons licensing terms, including attribution and ShareAlike conditions, and on human contributors continuing to create and improve material.

The motion says Wikimedia has a distinct perspective because its services sit at the intersection of freely licensed knowledge, machine-learning tools, and the traffic demands created by automated scraping and crawling. Its proposed brief argues that an undifferentiated fair-use rule could miss how retrieval and similar grounding systems may direct users away from a source while reproducing or serving its material.

Those are Wikimedia's positions. They do not establish that an AI system has infringed copyright, that every RAG implementation uses content unlawfully, or that all model training is protected by fair use. The MDL court will decide the questions properly before it on the record and arguments it accepts.

The filing is narrower than a ruling—and broader than a technical footnote

The procedural posture is narrow. Wikimedia asked Judge Sidney H. Stein for leave under the court's amicus process and attached a proposed 5,831-word brief. The motion states that no party or party counsel authored the proposed brief and that no person other than Wikimedia funded its preparation or submission.

But the distinction it asks the court to consider reaches beyond a single feature label. AI providers increasingly combine pre-trained models with search, retrieval, citation, browsing, or grounding layers. A fair-use analysis that separates those functions could make system design and deployment details more legally significant than a general description of a model as simply “trained on” online content.

For publishers and other rights holders, the proposed split focuses attention on whether a service is substituting for access to the underlying work. For AI companies and open-knowledge projects, it raises a parallel question about how copyright analysis should account for model development, current retrieval, attribution, licensing terms, and the practical burden that automated access places on a source.

None of those questions is resolved by ECF No. 2361. A later order on the motion would change the posture; a merits ruling would be a separate development altogether.

Filed motion and proposed brief: ECF No. 2361 Public docket snapshot: In re: OpenAI, Inc. Copyright Infringement Litigation Wikimedia's announcement: Wikimedia Foundation