Thomson Reuters v. ROSS: First Appeals Court Rejects AI Fair Use

A first appellate answer on AI training

On September 29, 2026, the U.S. Court of Appeals for the Third Circuit upheld Thomson Reuters' copyright win against ROSS Intelligence. The court rejected ROSS's argument that it made fair use of Westlaw material when it trained a competing AI legal search engine.45 Every outlet covering the case calls this the first time a federal appeals court has decided whether training an AI system on copyrighted works can be fair use.215 That makes it the most important precedent so far in the wave of AI training cases, though it may matter less than the headlines suggest.

The panel was Judges L. Felipe Restrepo, Tamika Montgomery-Reeves and Emil Bove, and Judge Montgomery-Reeves wrote the opinion.13 The court decided two questions. Westlaw's headnotes are original enough to be protected by copyright, and ROSS's use of them to train its tool was not fair use.1314 At first the public saw only a judgment reading "AFFIRMED." The full opinion stayed sealed while the parties had a chance to propose redactions.718 Neither side asked for any, and the reasoning was released soon afterward.119

How a defunct startup ended up making law

The case is older than the generative AI boom. Thomson Reuters sued in 2020, before ChatGPT existed, and Nieman Lab describes it as the first known AI copyright lawsuit in the United States.2 The company alleged that ROSS copied Westlaw headnotes, the short summaries of legal points in court opinions, to train a rival legal search engine.5

ROSS did not take the headnotes directly. It hired a contractor, LegalEase, which produced about 25,000 "Bulk Memos" pairing legal questions with passages from court opinions. ROSS used nearly all of them for training and validation.18 Thomson Reuters says the LegalEase drafters copied headnotes into those memos even though they had been told not to copy and paste from Westlaw.9 ROSS shut down its platform in early 2021 and blamed the cost of the litigation.4 Its antitrust counterclaim, which accused Thomson Reuters of illegally tying Westlaw's search tools to its case-law database, was dismissed in September 2024.4

The trial court's path was unusual. In 2023, Judge Stephanos Bibas, a Third Circuit judge sitting by designation in Delaware, sent fair use toward a jury. In February 2025 he reversed himself, finding that ROSS infringed 2,243 headnotes and could not claim fair use.7 He said he had studied the record more closely and concluded that his earlier ruling had not gone far enough.1 He then sent the originality and fair use questions to the appeals court.14

Why the court ruled against ROSS

The panel decided fair use mainly on the first and fourth factors, which courts usually treat as the most important.15 On purpose and character, it found ROSS's use highly commercial and only minimally transformative. ROSS and Westlaw used the headnotes for the same thing: improving a service that helps lawyers find relevant law.15 Training was an extra step, and the court acknowledged it as a slight difference. Even so, it held that the evidence showed ROSS trained its system to benefit a competing research platform.3

The opinion also rejected ROSS's comparison to the software "intermediate copying" cases such as Sega, Connectix and Google v. Oracle. ROSS could have worked from the judicial opinions themselves, which are free, and the court said convenience does not justify copying.14 On originality, the panel applied the low bar set in Feist. It held that headnotes are not "law": the opinions are the law and remain free for anyone to use, while the summaries reflect editorial choices.14

The market-harm analysis may have the most effect going forward. The court found harm in two markets. One is the existing legal research market, where ROSS was built to compete directly with Westlaw. The other is a developing market for licensing headnotes as AI training data.15 Thomson Reuters' use of headnotes to train its own AI search products counted as evidence that this licensing market exists. The court also said that the company's failure to license headnotes to others does not show the market is absent. The opinion also pointed to evidence that ROSS sometimes acted in bad faith, including by getting into Westlaw with other users' credentials.14

Where the reasoning diverges from the trial court

The two courts reached the same result, but not by identical routes. Judge Bibas had weighed the third factor, the amount used, in ROSS's favor because users never saw any headnotes in ROSS's output.18 The appeals panel did not consider output under that factor. It asked only whether ROSS took more than its purpose required, and because the opinions were freely available, it concluded the copying was unnecessary. That change matters. A test that ignores output takes away an argument generative AI defendants have relied on: that training copies are only an intermediate step and the finished model does not reproduce the works.7

The generative AI carve-out

Nearly all the coverage stresses one limit. ROSS's tool retrieved existing judicial text and did not generate new content.215 The court opened by calling the dispute "no more than an ordinary copyright case."16 In a footnote, it said the generative AI training arguments raised in the OpenAI litigation, which draw on Bartz v. Anthropic, do not apply here.14 The opinion also contrasted ROSS with the models in Bartz, noting that ROSS's platform cannot generate original expression.2

The distinction matters because trial courts have so far split depending on the facts. In Bartz, a judge found that training on books was fair use but that keeping a central library of pirated copies was not. The case led to a $1.5 billion class settlement that received final approval in July 2026.18 In Kadrey v. Meta, the court granted summary judgment to Meta on fair use because the plaintiffs failed to prove market harm.18 Courts abroad have also differed. A UK High Court ruling went against Getty Images in its case against Stability AI, while a Munich court found that OpenAI infringed song lyrics.18

On this point the coverage agrees more than it might seem. Ballard Spahr says the central generative AI questions are still unresolved.15 The Legal Layer newsletter argues that training a general-purpose model is probably still fair use, and that only the line between substitutive and generative uses has become clearer.9 The Authors Alliance expects limited effects on the largest AI cases. It still warns that plaintiffs in AI suits will cite the decision often, and that it binds federal courts in Pennsylvania, New Jersey, Delaware and the Virgin Islands.13

The critics' case

The ruling has drawn real criticism. The Authors Alliance argues that the panel got the merger doctrine wrong, dismissing in one brief paragraph ROSS's argument that headnotes merge with the legal points they state.13 Copyright Lately notes that commentators had already criticized Judge Bibas for handling merger in three sentences. It also cites ROSS's objection that the licensing-market theory is circular, because any copyright owner can claim it would have licensed the very use it is suing over.7

The circularity objection is the strongest point against the ruling. Treating a "developing market for AI training data" as harm protects whatever a rights holder says it might license in the future. Bibas wrote that harm to a potential training-data market was "enough."2 The appeals court did not back away from that view.9 Here the facts made the argument easy, since Thomson Reuters actually trains its own products on headnotes. Plaintiffs with weaker evidence will still try to stretch the reasoning.

What it means for legal tech and AI policy

The effect on legal technology is direct. Thomson Reuters said the ruling confirms that Westlaw's editorial content, written by its attorney editors, is protected by copyright.3 The company has framed copyright as protecting the intellectual property behind "fiduciary-grade AI solutions."4 In practice, the editorial layer of the big legal research platforms is now legally protected training data. A legal AI startup that wants that layer has to license it, or build its own from public opinions.

The case also contains practical lessons. Hiring a contractor did not shift the risk away from ROSS. An incumbent's own internal AI use can count as evidence of a licensing market. A compliance wiki also observes that the decision is now the clearest judicial statement on how Thomson Reuters content may be used to train AI, which is also addressed in the company's contracts.17

Music industry groups were watching as well. A brief from the RIAA and the National Music Publishers' Association argued that training a model to build a competing product cannot be fair use.3 Outside the courtroom, politics is also involved: Nieman Lab reported that the Trump administration has warned that a New York Times win against OpenAI could hurt national security.2

The case is not over. In 2025, Reuters reported that a jury would still hear the underlying case.1 The appellate opinion does not say where the remaining claims in the district court stand.17

The verdict

The ruling is narrow, and it is meant to be. It does not say AI training is unlawful. It says that copying a competitor's editorial work to build a near-identical substitute is not fair use just because a model sits in the middle. Generative AI companies can point to the court's own carve-out. Rights holders get something more durable: appellate recognition that licensing content for AI training is a real market, and that skipping it can count as harm. For the major generative AI cases still pending, that second point is likely to matter most.