U.S. Court of Appeals Rejects Fair Use Defense for Copyrighted AI Training Data

A U.S. appeals court has ruled that AI startup ROSS Intelligence could not rely on fair use to justify using copyrighted Thomson Reuters content to train its legal research technology.

The Third Circuit Court of Appeals affirmed an earlier ruling for Thomson Reuters, which accused ROSS of using thousands of editorial headnotes from its Westlaw legal research platform to develop a competing AI-powered service.

The decision is significant because it is the first U.S. federal appellate ruling to address fair use in the context of AI training. But it does not establish that training AI on copyrighted material is always infringement. Instead, the court focused closely on how ROSS obtained and used the material and the commercial market in which the two companies competed.

Why ROSS Lost Its Fair Use Argument

Westlaw’s headnotes summarize important legal points contained in judicial opinions. While the underlying court decisions are not protected by copyright, Thomson Reuters argued that its editors make creative decisions when selecting and writing the headnotes.

The Third Circuit agreed, finding that the 2,243 headnotes at issue contained enough original editorial judgment to qualify for copyright protection.

ROSS used those headnotes in training materials for an AI system designed to help users find relevant legal passages. According to the court, using AI as an intermediate step did not fundamentally change the material’s purpose.

Both products ultimately helped users conduct legal research, and ROSS intended to compete directly with Westlaw. The court therefore described the use as “minimally transformative at best.”

The judges also considered the potential market for licensing Westlaw’s content for AI training, strengthening Thomson Reuters’ argument that unauthorized use could affect the commercial value of its material.

Not Every AI Training Case Is the Same

This distinction matters because U.S. courts have reached different conclusions in other AI copyright disputes.

In 2025, federal judges considering cases involving Anthropic and Meta found that certain uses of copyrighted books to train generative AI could qualify as transformative fair use, although separate questions remained about the acquisition and storage of pirated copies.

The Third Circuit specifically distinguished the ROSS case from generative AI litigation. ROSS’s system did not generate original expression like a large language model, and its product was intended as a commercial substitute in the same legal-research market.

AI training therefore remains a highly fact-specific copyright question, not one with a universal legal answer.

What This Means for E-commerce Content

The case also raises a relevant question for e-commerce businesses adopting generative AI: where does the information their AI systems use come from?

Retailers, marketplaces, brands, and technology providers increasingly use AI to create descriptions, translate catalogs, classify products, generate marketing materials, and enrich existing product information.

Those workflows can involve enormous quantities of content, including descriptions, specifications, manuals, images, and other materials created by brands and content providers.

The ROSS ruling shows why provenance and licensing cannot be treated as an afterthought.

For Icecat, this connects directly with the importance of established product-content relationships. Icecat distributes structured product information through defined licensing models, including content supplied and authorized by participating brands.

As AI becomes more deeply involved in catalog enrichment and e-commerce automation, knowing where product information originated and how it may be used becomes increasingly valuable.

AI Needs More Than Data

The AI industry has often focused on acquiring larger datasets and building increasingly capable models. Copyright litigation is adding another requirement: understanding the rights attached to the information inside those datasets.

The Third Circuit ruling will not settle the wider debate around generative AI training. ROSS has said it intends to seek further review, and other major copyright cases involving AI companies continue to move through U.S. courts.

But for companies deploying AI commercially, the lesson is already practical.

Good AI systems need high-quality data. Increasingly, businesses may also need to demonstrate that they have the right to use it.