The first federal appellate court to decide whether copying copyrighted material to train an artificial intelligence system can constitute fair use ruled against the AI company. That sounds like a major setback for AI developers. But the Third Circuit’s decision in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. may tell us considerably less about the coming wave of generative AI copyright cases than that headline suggests.

In fact, the court practically told us not to read its decision that broadly. It called the dispute “no more than an ordinary copyright case” and expressly distinguished the generative AI litigation involving Anthropic, Meta and OpenAI. ROSS did use copyrighted material to train an AI system. But what ROSS built, what it copied, and what it was trying to compete with made this an unusual stand-in for the much larger fight over large language models.

What ROSS did

Thomson Reuters owns Westlaw and claims copyright protection in the editorial headnotes its lawyers and editors prepare to summarize points of law appearing in judicial opinions. ROSS Intelligence wanted to build a competing legal research platform. To train its system, ROSS relied on thousands of training questions created from Westlaw headnotes, paired with passages from judicial opinions that supplied the answers.

That intermediate training step mattered, but the Third Circuit focused on the ultimate purpose of the copying. Thomson Reuters used its headnotes to help researchers locate and understand relevant judicial opinions. ROSS used those headnotes to train a competing platform that helped researchers locate relevant passages from judicial opinions. According to the court, those purposes were highly similar.

That distinction drove much of the fair use analysis. ROSS was commercial. Its use was, at best, only minimally transformative. It copied entire headnotes even though the underlying judicial opinions were freely available. And it intended to compete directly with Westlaw.

The Third Circuit therefore affirmed the district court’s ruling that the headnotes were copyrightable and that ROSS had not made fair use of them.

Why this does not map neatly onto generative AI

At a sufficiently high level of abstraction, ROSS looks familiar. A company acquired copyrighted material, copied it during an AI training process, and used what the system learned to make a commercial product.

But that level of abstraction leaves out much of what mattered to the court.

ROSS did not train a general-purpose large language model capable of generating new expression. Its system retrieved passages from existing judicial opinions in response to legal research questions. The Third Circuit emphasized precisely this point when distinguishing Bartz v. Anthropic and the pending OpenAI litigation. Generative models can produce new responses. ROSS’s system could not.

ROSS also presented an unusually straightforward substitution story. It wanted to compete with Westlaw in legal research. It charged comparable prices. And the copied material helped ROSS perform substantially the same function that the material performed for Thomson Reuters.

That is different from many of the generative AI cases now moving through the courts. An author whose novel entered an LLM training corpus may argue that the model threatens a market for the author’s work, but the relationship is not necessarily as direct as one legal research platform using a competitor’s editorial material to build another legal research platform.

The distinction is worth keeping in mind because copyright litigation against generative AI companies has already raised a much broader collection of questions about training inputs, model outputs, substantial similarity and derivative works. I wrote about some of those issues earlier in connection with authors’ claims against OpenAI. See my earlier internetcases post on the authors’ OpenAI copyright case

So while ROSS and the LLM cases may look isomorphic from a distance because copyrighted works enter an AI training process, the copyright-relevant relationships among the input, the output, the purpose of the copying and the allegedly displaced market can be quite different.

The part of the decision that may travel

That does not mean ROSS will have little effect outside legal research. One part of the Third Circuit’s reasoning may prove much more important in the generative AI cases: its treatment of the market for licensing copyrighted works as training data.

Thomson Reuters did not have an established business licensing its headnotes to third parties for AI training. ROSS argued that this undercut Thomson Reuters’s claim of market harm. The Third Circuit disagreed. It observed that Thomson Reuters used headnotes in developing its own AI products and reasoned that the absence of existing third-party licenses did not mean a licensing market could not exist. By taking the headnotes without permission, the court said, ROSS had usurped Thomson Reuters’s opportunity to enter that market.

That reasoning has potentially much wider implications.

Fair use has long presented a difficult question about hypothetical licensing markets. If a copyright owner can defeat fair use merely by saying, “I would have charged you for permission to make that use,” the analysis risks becoming circular. The claimed licensing market exists because users need licenses, and users need licenses because the claimed licensing market counts against fair use.

AI training puts unusual pressure on that problem because markets for licensed training data are developing at the same time courts are deciding whether licenses are legally necessary in the first place. The Third Circuit did not have to solve that problem across the entire AI ecosystem. It considered a much narrower situation involving copyrighted editorial material used by a direct competitor to build a competing product.

But future litigants will surely cite its language.

“AI training” does not decide the fair use question

Perhaps the most useful lesson from ROSS is that “AI training” is not itself a category that resolves the copyright analysis.

A court can ask what the defendant actually copied, why it copied it, whether the copying was necessary to achieve a different purpose, what the resulting system does, and how that system affects the market for the copyrighted work. Those questions produced a relatively easy answer for the Third Circuit because ROSS copied a competitor’s editorial work to help build a product serving much the same function in the same market. Change those facts and the fair use analysis may change with them.

That makes ROSS significant, but not necessarily for the reason suggested by the simplest headline. The first federal appellate decision about copyrighted works used in AI training does not establish that AI training generally is infringing. If anything, the court’s careful effort to distinguish generative AI shows how much work remains before the appellate courts answer that question.

Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir. Sept. 29, 2026)