AI Training Fair Use Rejected as Third Circuit Backs Westlaw Over ROSS

A first appellate answer, with a narrow scope

A federal appeals court has ruled for the first time on whether copying copyrighted material to train an AI system counts as fair use. Its answer, in this case, was no. On September 29, 2026, the U.S. Court of Appeals for the Third Circuit upheld a partial summary judgment for Thomson Reuters. The court found that 2,243 Westlaw headnotes are original enough to be copyrighted and that ROSS Intelligence's use of them to train a legal-research AI was not fair use.1 The opinion was first issued under seal and made public the next day.49

Press coverage summed it up as "AI training is not fair use." That framing goes further than the court did. The panel called the dispute "no more than an ordinary copyright case."2 It also said directly that ROSS's tool was a different kind of system from the generative models at the center of the biggest AI lawsuits.214 So the ruling matters, but mostly because of what it adds to how courts weigh market harm. It does not settle the generative AI question.

What ROSS actually did

ROSS set out to build an AI search engine that would take plain-English legal questions and answer with passages from court opinions.15 Thomson Reuters had refused to license its content to ROSS. ROSS then got its training material from a vendor, LegalEase. LegalEase and a subcontractor wrote about 25,000 "Bulk Memos" built from Westlaw headnotes, and ROSS turned those memos into machine-readable training data.12 Users never saw a headnote. The headnotes were used only to teach the model.1617

That detail is why the case matters to AI developers. ROSS argued that copying used only inside training, and never shown in the product, should be treated differently from republishing. The Third Circuit disagreed. It said training was an intermediate step that served the same commercial goal as Westlaw. Writing for the court, Judge Tamika Montgomery-Reeves granted that the training step "arguably presents a slight degree of difference in use." She found that this difference did not change the result. The panel was unanimous. The other judges were Restrepo and Bove.11

ROSS's own conduct also hurt it. A footnote cited evidence that people working for ROSS tried to get into Westlaw using law-firm investor credentials, a student account, and an inquiry made under the guise of a solo practitioner.11 The court said that, to whatever extent good faith still matters in fair use analysis, this weighed against ROSS.714

How the four factors fell

Reports agree on the overall tally: three of the four statutory factors went against ROSS.9 Under the first factor, the court found the use highly commercial and "minimally transformative, at best," because both companies used headnotes for the same job of helping users find relevant law.12

The line most likely to be quoted came when the court dismissed ROSS's comparison to the software intermediate-copying cases: Google v. Oracle, Sega, and Connectix. In those cases, copying was needed to reach unprotected functional code. ROSS had free access to the court opinions themselves and chose the headnotes because they were easier to work from. As the court put it, "Unlike necessity, ease is not a justification for copying."111 The panel also explained why Authors Guild v. Google did not help ROSS. Google's book search served a different purpose from the books and could even lead readers to buy them.1

The second factor leaned slightly toward ROSS because headnotes are more factual than creative.2 The third factor is where the appeals court departed from the trial judge. Judge Stephanos Bibas, a Third Circuit judge sitting by designation in Delaware, had given that factor to ROSS.1 The panel gave it to Thomson Reuters. It reasoned that each headnote is a complete copyrighted work and that ROSS took more than it needed.14 ROSS argued it had used only 0.08 percent of Westlaw's roughly 28 million headnotes. The court did not accept that argument.11 So on that factor the appellate opinion is somewhat tougher on ROSS than the ruling it upheld.7

The real precedent: a licensing market for training data

The fourth factor is the part of the ruling most likely to reach beyond legal research. The court found harm in two separate markets. The first was Westlaw's existing legal-research business, where ROSS sold itself as a replacement. The second was a developing market for licensing headnotes as AI training data.2 ROSS had marketed itself openly as a Westlaw alternative at similar prices.15

On the second market, the court relied on evidence that licensing headnotes for AI training is "rapidly developing." Thomson Reuters already uses its headnotes to train its own AI search products.116 Thomson Reuters had not licensed the headnotes to anyone else, and the court held that this did not make the market imaginary.11 Haynes Boone reads the result as appeals-court recognition that licensing text for AI training is a legitimate derivative market, even when the owner has licensed to no one.1 One commentator noted that Judge Bibas had said in the trial court that the effect on a potential training-data market was enough on its own to count against ROSS, and the Third Circuit left that view in place.10

This is the part of the opinion that should worry AI developers most, including generative AI companies. If a market for training licenses can be harmed before it fully exists, copyright owners have an argument available in almost any case. Womble Bond Dickinson makes a related point. The footnote suggests generative AI developers could do better on transformativeness, but they would still have to deal with the licensing-market reasoning under the fourth factor. That finding, though, came in a case where the use was at most minimally transformative.17

Where coverage diverges

Coverage agrees on the outcome but splits on what it means. Some analysts treat the decision as tightly limited. A Substack analysis argued that training a general-purpose model is probably still fair use and that the ruling mainly threatens products that compete directly with their own training data.10 King & Spalding stressed that competing with the source material was the riskiest position ROSS could take, and noted that the court observed generative AI cases showed more transformation and less direct competition.15

Others expect the reasoning to travel further. ArentFox Schiff wrote that the court's points about necessity and market harm could come up in disputes over fine-tuned or specialized models.7 A compliance-focused analysis predicted that every foundation-model developer will rely on the generative-AI carve-out, and said the work now is figuring out where that carve-out actually holds.6

Accounts of the case's history differ somewhat. Several reports say Judge Bibas first treated fair use as a question for a jury in 2023 and then reversed himself in February 2025.910 IPWatchdog says he originally ruled for ROSS on fair use before "new information and understanding" changed his mind.4 All accounts agree that the February 2025 ruling is the one the Third Circuit affirmed.8

One point is clear across the coverage: this decision runs against the trend in generative AI cases. Two federal judges in California, in Bartz and Kadrey, found training on books to be fair use on the records before them in 2025.13 The U.S. government has also called training "extraordinarily transformative." The Third Circuit set aside both the Justice Department's statement of interest in the OpenAI multidistrict litigation and those two rulings, on the ground that ROSS's system cannot generate original expression.14 Meanwhile, the authors' case against OpenAI in the Southern District of New York is still waiting on summary judgment before Judge Stein.13

The legal-tech angle

The legal-tech industry is most directly affected. The ruling backs incumbents that control curated legal content. The court held that editors' choices about which points of law to summarize, and how, supply enough creativity for copyright. It also dismissed ROSS's argument that this gives Westlaw a monopoly over the law, noting that the opinions themselves remain free.12

This looks like a strategic win for Thomson Reuters that goes beyond a defunct opponent. ROSS shut down in 2021, citing litigation costs.3 The case now goes back to the District of Delaware for trial on remaining issues and damages.17 Those include thousands of other headnotes and questions about whether some copyrights have expired.7 The larger prize is court recognition that Westlaw's editorial work is a licensable training asset. Thomson Reuters is already using that asset for its own AI tools.1

For startups building AI research tools, the practical message is about where data comes from. Law firm advisers are telling clients to look at whether their product competes with its sources, to respect terms of service, and to budget for licensing.11 ArentFox Schiff notes that the court treated the source of training data, the reasons for using it, and compliance with access terms as relevant to fair use.

The narrow reading of the case is probably right as a matter of doctrine. The court's own footnote limits it. But the fourth-factor holding is broader than the footnote, and copyright owners will likely invoke it well outside legal research. No federal appeals court has yet ruled on generative AI training.7 Until one does, the ROSS opinion is the only appellate ruling courts can cite on AI training and fair use, and its lesson is that AI training gets no special exemption from copyright law.15