Thomson Reuters v. ROSS Intelligence (Third Circuit): No fair use for training a Westlaw rival

The first appellate word on AI training

On 29 September 2026, a panel of the U.S. Court of Appeals for the Third Circuit (Judges L. Felipe Restrepo, Tamika R. Montgomery-Reeves and Emil J. Bove) decided Thomson Reuters v. ROSS Intelligence. Writing for the court, Judge Montgomery-Reeves held that 2,243 headnotes written by Westlaw's editors are protected by copyright, and that ROSS's use of them to train an AI legal search engine was not fair use ( Third Circuit opinion). The decision affirmed a February 2025 summary judgment by Judge Stephanos Bibas, a federal appeals judge who had sat by designation as the trial judge in Delaware.

The ruling arrived in two steps. The court first filed the opinion under seal and gave the parties 10 days to propose redactions, so for a short time the outcome circulated while the reasoning did not ( LawSites on the sealed filing). The full text was released shortly afterward ( LawSites on the unsealed opinion). Law firms quickly described it as the first decision by a federal court of appeals on whether training an AI system on copyrighted material is fair use ( Ballard Spahr analysis).

That description is accurate and easy to overread. The opinion itself insists that, whatever ROSS's framing, this is "no more than an ordinary copyright case," and in a footnote it separates ROSS's tool, which returned passages from existing court opinions, from the generative models at the center of the book and news lawsuits. This report follows what the court decided, why its reasoning matters beyond legal research, and where its authority stops.

How ROSS built a search engine from Westlaw's notes

Westlaw is Thomson Reuters's online research service; West Publishing, one of the early private case reporters, is now part of the company. Westlaw organizes judicial opinions with its Key Number System and attaches headnotes to them. A headnote states one point of law from an opinion in the abstract; in the interface it sits above the opinion and links to the passage it summarizes. The opinion records that West's editors follow drafting rules: each headnote must be concise, carry only the essential facts, stand on its own for a reader without the opinion, and stay within about 800 characters where possible. Thomson Reuters holds registered copyrights in this material, disclaims any copyright in the government-authored opinions themselves, and had used its headnotes to train its own AI search feature, WestSearch Plus ( Third Circuit opinion).

ROSS began, according to the same opinion, with three University of Toronto computer science students who entered an IBM competition built on the Watson system and proposed using it to answer legal questions without Boolean keyword searches. The company they founded built a search engine that took plain-language questions and returned relevant passages from a bank of roughly ten million public-domain judicial opinions. It was not generative: it could only hand back text that judges had already written.

To learn which passages answer which questions, the system needed labelled examples. ROSS first asked to license Westlaw content, and Thomson Reuters refused because ROSS was a competitor, as Bibas recounted in his 2025 memorandum opinion. ROSS then hired LegalEase Solutions, which with a subcontractor wrote about 25,000 training memos. Each posed a legal question and offered four to six opinion passages graded great, good, topical or irrelevant. The drafters framed their questions from thousands of Westlaw headnotes, and the great passage was most often the one Westlaw itself linked to that headnote. LegalEase's guide told writers not to paste headnotes in directly, but the questions tracked them closely, and the memos were converted into numerical training data.

In machine-learning terms this is a supervised relevance-ranking dataset, and the mechanism explains why the headnotes were worth copying. Westlaw's editors had already decided which point of law each passage stands for and phrased it as a self-contained statement. Turned into memo questions, that editorial work supplied both the queries and, through Westlaw's own links, much of the answer key. ROSS made no secret of its aim: it ran advertisements comparing itself directly with Westlaw at prices in line with Westlaw's, and some firms switched.

Six years from complaint to appeal

Thomson Reuters sued in the District of Delaware in May 2020, alleging copyright infringement and tortious interference with contract. In December 2020, ROSS announced that it would cease operations because the lawsuit had left it unable to raise new funding, and that its platform would go offline by 31 January 2021, while it kept defending the case with litigation insurance ( ABA Journal). The product at the center of the first appellate ruling on AI training had therefore been dark for more than five years when the ruling came, and had closed nearly two years before ChatGPT made generative AI a mass product.

In 2023 Bibas largely denied summary judgment and left fair use for a jury. Studying the record before an August 2024 trial date, he concluded that his earlier ruling had not gone far enough, postponed the trial and invited new briefing. His February 2025 opinion found actual copying of 2,243 headnotes out of a contested batch of 2,830, held the headnotes original, and rejected fair use: factors one and four went to Thomson Reuters, factors two and three to ROSS. He stressed that only non-generative AI was before him ( Bibas memorandum opinion). At the parties' request he then certified two questions for immediate appeal: whether the headnotes and the Key Number System are original, and whether ROSS's use was fair. ROSS never briefed the Key Number System, so the panel treated that issue as forfeited ( Third Circuit opinion).

The appeal was argued on 11 June 2026, with Mark Davies of White & Case for ROSS and Dale Cendali of Kirkland & Ellis for Thomson Reuters. The judges spent much of the argument on market harm: Judge Restrepo turned Davies to the fourth factor almost at once, and Judge Bove pressed on whether a market for AI training data exists and on why ROSS's product was more than a different kind of legal search engine ( LawSites argument report). Amicus counsel listed in the opinion include the Electronic Frontier Foundation and Stanford and Harvard scholars on ROSS's side, and the Association of American Publishers and the News Media Alliance on Thomson Reuters's.

Why a summary of the law can be property

The first question was originality. Under the Supreme Court's 1991 decision in Feist v. Rural Telephone, a work needs only independent creation and a modicum of creativity, though an alphabetical telephone listing fell below even that bar. The panel held that every one of the 2,243 headnotes cleared it, because editors chose which points of law mattered and how to phrase them so each could stand alone while accurately reflecting the opinion. It found support in Callaghan v. Myers (1888), which recognized a reporter's copyright in headnotes, and in Georgia v. Public.Resource.Org (2020), which restated that principle for explanatory material written by someone without authority to speak with the force of law ( Third Circuit opinion).

ROSS's three counterarguments failed. Protecting headnotes does not give anyone a monopoly over the law, the court said, because headnotes are not law and judicial opinions remain free. The merger doctrine, which denies protection when an idea can be expressed in only one or a few ways, did not apply: the court compared headnotes to a banana costume, an earlier Third Circuit example of a subject open to many designs, and pointed to a LexisNexis amicus brief explaining that Lexis writes different headnotes for the same opinions. And the Second Circuit's 1998 Matthew Bender decision, which denied West protection for its arrangement of party names and its parallel citations, concerned choices dictated by industry convention, not independently composed summaries.

Not everyone accepts that reasoning. Dave Hansen of Authors Alliance argued that the merger analysis was conclusory and that the very constraints West imposes on its editors (accuracy, essential facts, length, staying close to the court's language) should push toward merger rather than away from it ( Authors Alliance critique). The panel also left open a question Bibas had raised in passing, whether a headnote that quotes an opinion verbatim can be original, because none of the 2,243 did.

Ease is not necessity

On the first factor, the purpose and character of the use, the panel followed the Supreme Court's 2023 Warhol decision: a commercial use that shares the same or a highly similar purpose as the original is likely to weigh against fair use. ROSS's use was plainly commercial. Thomson Reuters uses headnotes to help researchers find relevant opinions; ROSS used them to build a platform that helps researchers find relevant opinions. The intermediate step of training an AI program offered at most a slight difference, so the use was, in the court's phrase, "minimally transformative, at best" ( Third Circuit opinion).

ROSS leaned on two lines of precedent. The Google Books case was distinguished because Google's search tool served a different function from reading books and could lead users to buy them, whereas ROSS aimed to replace Westlaw. The intermediate-copying cases, Sega v. Accolade (1992), Sony v. Connectix (2000) and Google v. Oracle (2021), were read as turning on necessity: copying code was the only way to reach unprotected functional elements and make programs interoperate. ROSS had the judicial opinions and could have built its memos from them; it used headnotes because they were an easy way to frame questions. "Unlike necessity, ease is not a justification for copying," the court wrote, the line Dennis Crouch took as the headline of his Patently-O analysis. A footnote added that, to whatever extent good faith still matters, the record cut against ROSS: it had tried to reach Westlaw through a law-firm investor's credentials, an employee had inquired about an account while posing as a solo practitioner, and another had used a student account.

The second factor slightly favored ROSS, since the headnotes were published and largely factual. The third factor is where the panel quietly parted from the trial judge. Bibas had given it to ROSS because no headnote ever reached ROSS's users, following the Google Books focus on what a copier makes available to the public, and the U.S. Copyright Office's May 2025 report on generative AI training had cited that ruling as an example of courts looking at outputs. The Third Circuit asked instead whether ROSS took more than necessary. Because the opinions were freely available, copying headnotes was not necessary at all, and because each headnote is its own copyrighted work, ROSS copied whole works. ROSS's point that it took only 0.08% of Westlaw's 28 million headnotes did not help, since quality can matter more than quantity. Aaron Moss of Copyright Lately observed that the panel never engaged Bibas's output-based formulation ( Copyright Lately), and Hansen noted that the opinion's third-factor passage refers to 25,000 copied headnotes, the number of memos, while the judgment concerns 2,243 ( Authors Alliance critique).

A market for training data, recognized from the bench

The fourth factor, effect on the market, produced the holding most likely to travel. ROSS argued that nobody sells headnotes on their own. The panel answered with its 2003 Video Pipeline decision about Disney film trailers: the statute protects a work's value, not only a standalone market, and headnotes are a feature that draws subscribers to Westlaw. In the legal research market itself, unrestricted copying to build substitute platforms would plainly harm Westlaw. Then the court turned to derivative markets. It found that the market for licensing headnotes as AI training text is "rapidly developing," that Thomson Reuters uses them to train its own AI search products, and that its choice not to license them to others does not show that no market exists. By copying without permission, ROSS usurped that opportunity, and as the party asserting fair use it offered no contrary evidence ( Third Circuit opinion).

The commercial context explains why the point matters to Thomson Reuters. In August 2023 it closed a 650 million dollar cash acquisition of Casetext, whose CoCounsel legal assistant was powered by GPT-4 ( Thomson Reuters announcement). The court did not rely on that deal, but legal publishers now have their own AI products to feed and a reason to price their editorial corpora as training data.

ROSS's public-benefit arguments fared no better. The court said copying headnotes did little for access to the law when opinions were free and ROSS charged Westlaw-level prices, and it found no evidence that the ruling would halt AI development or that legal research had anything to do with national security. The Association of American Publishers, which had filed an amicus brief, welcomed the ruling for recognizing a developing market for licensing works as AI training data, and noted that roughly 150 similar AI copyright cases are pending in the United States ( AAP statement).

The footnote that fences off generative AI

Footnote 7 is the opinion's answer to the question everyone outside legal research was asking. On 1 September 2026, the Justice Department had filed a statement of interest in In re OpenAI, the consolidated copyright litigation in the Southern District of New York, relying on Bartz v. Anthropic to argue that training a large language model able to generate original responses is transformative and does not cause substitutive competition. The panel said those concerns did not apply: unlike the models in Bartz and In re OpenAI, ROSS's platform could not generate original expression, and ROSS trained it to create a commercial substitute for Westlaw. It added that the government had not weighed in on this case ( Third Circuit opinion).

The distinction runs in both directions. In June 2025, Judge William Alsup had already set Bibas's ruling aside in Bartz, noting that Bibas twice emphasized the tool was not generative and suggesting that a better analogue would be an AI tool trained on legal materials to produce fresh legal writing ( Bartz Order on Fair Use). Yet Alsup's reasoning on necessity differs in kind. He accepted that Anthropic could have trained on other books or none, held that reasonably necessary does not mean strictly necessary, and concluded that because language models need an enormous volume of text, using any one book was about as reasonable as the next, especially as no infringing output reached the public. Two days later, Judge Vince Chhabria, ruling for Meta over the training of its Llama models, held that lost fees for licensing a work for a transformative purpose are not a cognizable harm, precisely to keep the fourth factor from becoming circular ( Kadrey v. Meta order).

Divergence or reconciliation

Commentators disagree about whether these rulings can coexist. Chad Rutkowski of BakerHostetler argued that the Third Circuit applied a strict necessity standard where Warhol speaks of reasonable necessity, that it treated competition with Westlaw as enough where Bartz said copyright does not protect authors against competition, and that it has therefore diverged from Bartz in a way that sets up a possible circuit split. His larger worry was not frontier models but smaller enterprise models trained on competitors' content to learn unprotected information ( BakerHostetler analysis).

Moss read the decisions as reconcilable. Alsup and Chhabria both assumed a training-license market could exist and held only that lost fees for transformative uses do not count; ROSS lost because its use was not transformative, so the real dispute remains whether generative training is. He also noted that footnote 7 offers two grounds, the absence of generation and the purpose of substitution, which point toward different rules: AI developers will stress the first, rights holders the second. Plaintiffs had already cited the ruling in Concord Music v. Anthropic ahead of a summary judgment hearing on 21 October 2026 ( Copyright Lately).

A formal split needs two courts of appeals to disagree. Bartz and Kadrey are district court rulings, and Bartz ended in a settlement without appellate review, so for now the tension is between a binding appellate opinion and persuasive trial-level ones.

The honest limits

Four limits keep the ruling narrower than its billing. First, it concerns a tool that retrieved passages rather than writing text. Second, it is interlocutory. The case returns to Delaware, where whether some copyrights have expired, the claims on the remaining headnotes and damages are still for trial ( Patently-O analysis), along with the tortious interference claim ( LawSites on the unsealed opinion).

Third, it binds only the federal courts of Pennsylvania, New Jersey, Delaware and the Virgin Islands, although plaintiffs elsewhere will cite it as the first appellate word on AI training ( Authors Alliance critique). Fourth, its facts are unusually bad for the defendant: a direct competitor, refused a license, building a substitute from the plaintiff's own editorial work, with evidence of evasive Westlaw access. ROSS said it would ask the full Third Circuit to rehear the case and would ultimately seek Supreme Court review ( LawSites on the unsealed opinion). On 3 November 2026 it filed a petition for panel rehearing or rehearing en banc, signed by a Gibson Dunn appellate team that includes former Acting Solicitor General Jeffrey Wall. The petition argues that the panel judged the purpose of ROSS's business rather than its use of the headnotes, ignored that no headnote ever reached a user, and found market harm in the lost chance to license the very use at issue, and it says the decision conflicts with Warhol, Google v. Oracle and the Ninth Circuit's reverse-engineering cases Sega and Connectix. A timely petition holds back the court's mandate, so the case stays in the appeals court until the petition is decided, and the ruling changes only if the panel or a majority of the circuit's active judges agrees to hear it again. Moss doubts it is a good vehicle, given its interlocutory posture, a defendant that no longer operates and a non-generative tool, and Hansen expects the Supreme Court's likelier path to run through generative cases in the Second and Ninth Circuits.

How to place the ROSS ruling in AI history

Lasting influence is provisional barely five weeks on, but the first federal appeals court to rule on AI training did so on a defunct, non-generative product, wrote deliberately narrowly, and still established three propositions within its circuit. An intermediate training step does not make a use transformative when the finished product serves the original's purpose. Convenience is not the kind of necessity that excuses copying expression to reach unprotected material. And a market for licensing works as training data is real enough to count, even before the owner has licensed anyone.

Set beside Bartz, the case completes a picture of data provenance in the courts. Bartz put a price on acquiring books from pirate libraries; ROSS says that a competitor's editorial work cannot be borrowed as an answer key simply because it is convenient. Both leave untouched the central question of whether training a general generative model on expressive works is fair use. ROSS itself grew out of a student entry in an IBM Watson competition, and the most lasting product of that start may be this precedent rather than the search engine it was meant to produce.

References

• Tamika R. Montgomery-Reeves, U.S. Court of Appeals for the Third Circuit (2026). Opinion of the Court, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153.
• Gibson, Dunn and Crutcher for ROSS Intelligence Inc. (2026). Appellant ROSS Intelligence Inc.'s Petition for Panel Rehearing or Rehearing En Banc, No. 25-2153.
• Stephanos Bibas, U.S. District Court for the District of Delaware (2025). Memorandum Opinion, Thomson Reuters v. Ross Intelligence.
• William Alsup, U.S. District Court for the Northern District of California (2025). Order on Fair Use, Bartz v. Anthropic PBC.
• Vince Chhabria, U.S. District Court for the Northern District of California (2025). Order on cross-motions for partial summary judgment, Kadrey v. Meta Platforms.
• U.S. Copyright Office (2025). Copyright and Artificial Intelligence, Part 3: Generative AI Training, Pre-Publication Version.
• Legal Information Institute, Cornell Law School (1991). Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340.
• Bob Ambrogi, LawSites (2026). 3rd Circuit Issues Opinion in Thomson Reuters v. ROSS Case, But For Now It Is Sealed.
• Bob Ambrogi, LawSites (2026). No More Than an Ordinary Copyright Case: 3rd Circuit Unseals Its Opinion In Thomson Reuters Case Against ROSS.
• Bob Ambrogi, LawSites (2026). At 3rd Circuit, Judges Press ROSS and Thomson Reuters on Fair Use, AI Training and Market Harm.
• Lyle Moran, ABA Journal (2020). ROSS Intelligence will shut down amid lawsuit from Thomson Reuters.
• Harlan Mechling, Kenneth R. Davis II and Catherine I. Seibel Sinitsa, Ballard Spahr (2026). Third Circuit Addresses Fair Use in AI Training, But Leaves Generative AI Questions Unresolved.
• Chad A. Rutkowski, BakerHostetler (2026). What Thomson Reuters v. Ross Does and Doesn't Say NOW About Fair Use and Generative AI.
• Aaron Moss, Copyright Lately (2026). Everyone Won the Ross Fair Use Appeal. Except Ross..
• Dave Hansen, Authors Alliance (2026). Third Circuit affirms in Thomson Reuters v. ROSS: The Court Gets Merger Wrong and Does a Disservice to Fair Use.
• Dennis Crouch, Patently-O (2026). Ease Is Not Necessity: Third Circuit Affirms No Fair Use in Thomson Reuters v. ROSS.
• Association of American Publishers (2026). Association of American Publishers Welcomes the Third Circuit Decision in Favor of Plaintiff in Thomson Reuters v. ROSS Intelligence.
• Thomson Reuters (2023). Thomson Reuters completes acquisition of Casetext, Inc..