Third Circuit Rules in Thomson Reuters v. ROSS on 29 September 2026: Training a Legal AI on 2,243 Westlaw Headnotes Was Not Fair Use
In 1888 the Supreme Court decided Callaghan v. Myers, a dispute over the Illinois law reports. The judges' opinions, it held, belonged to no one. The reporter who had read those opinions and condensed them into headnotes, syllabi and tables of contents could own that work, because it was the product of his own intellectual labor. One hundred and thirty-eight years later the Third Circuit relied on that case to decide one of the first appellate disputes over training artificial intelligence on someone else's annotations.
Practical takeaway. The Third Circuit's opinion of 29 September 2026 holds that copying editorial annotations to train a competing, non-generative AI product is not fair use where the underlying public material was freely available. Necessity can justify copying; convenience cannot. The court also treats licensing for AI training as a derivative market that belongs to the copyright owner, even where the owner has not yet licensed anyone.
What the panel decided in No. 25-2153, filed 29 September 2026
The precedential opinion resolves an interlocutory appeal from the District of Delaware, where Circuit Judge Stephanos Bibas, sitting by designation, had granted Thomson Reuters partial summary judgment. He certified two questions under 28 U.S.C. 1292(b): whether the headnotes and the West Key Number System are original, and whether ROSS's use of the headnotes was fair use. The panel heard argument on 11 June 2026. Judge Montgomery-Reeves wrote the opinion of the court for a panel with Judges Restrepo and Bove, and the court affirmed on both points it reached. ROSS never briefed the originality of the Key Number System, so the court treated that question as forfeited.
The facts are unusually clean for an AI case. ROSS was founded by three University of Toronto computer science students after an IBM Watson competition. Its product answered plain-language legal questions by returning passages from roughly ten million judicial opinions; the court stresses that it generated no new text. To teach the system which passages answer which questions, ROSS hired LegalEase Solutions, which wrote about 25,000 training memos. Each memo posed a legal question and offered four to six opinion passages labeled great, good, topical or irrelevant. The memo writers used Westlaw headnotes to frame the questions, because headnotes, in the words quoted from the trial record, were "an easy way" to do it. The district court found that 2,243 of the memo questions tracked the headnote text so closely that no reasonable juror could find they were not copied.
Why the court held Westlaw headnotes are copyrightable
The originality bar in Feist v. Rural Telephone (1991) requires only a "modicum of creativity". The court found it in the editors' choices about which points of law matter and how to state them in about 800 characters so that each headnote stands on its own. It rejected the argument that this gives Thomson Reuters a monopoly over the law: headnotes are not law, and the opinions remain free to all, as the Supreme Court held in Banks v. Manchester, also in 1888. The merger doctrine failed because there are many ways to express a point of law; the brief from Lexis, which writes different headnotes for the same opinions, made that concrete. One question stays open. The trial judge had suggested that even headnotes quoting a court verbatim might be protected, and the panel declined to decide it, because the 2,243 headnotes at issue do not copy opinion text word for word.
How the four fair-use factors came out
The court applied the four factors of 17 U.S.C. 107 through the Supreme Court's 2023 decision in Andy Warhol Foundation v. Goldsmith. Three went against ROSS and one leaned slightly its way.
Purpose and character. ROSS's use was commercial, and its purpose matched Westlaw's: helping researchers find responsive case law. The panel allowed that training an AI program "arguably presents a slight degree of difference" and still called the use "minimally transformative, at best". It distinguished Authors Guild v. Google, where book search pointed readers back to the books; ROSS, by its own admission, aimed to replace Westlaw. In a footnote the court added that ROSS at times acted in bad faith, including attempts to reach Westlaw through investor and student accounts.
Nature of the work. Headnotes are published and largely factual, so this factor favored fair use, though the court noted it rarely decides cases.
Amount used. ROSS pointed out that it used about 0.08 percent of Westlaw's 28 million headnotes. The court answered that each headnote is its own copyrighted work, so every headnote ROSS copied was a complete work, and that copying more than necessary is unreasonable without a transformative purpose.
Market effect. Here the opinion speaks most directly to the AI economy. Even assuming no one buys headnotes on their own, they draw subscribers to Westlaw, and the court applied its 2003 Video Pipeline decision on movie trailers to protect that value. It then recognized a potential derivative market for licensing headnotes as AI training data, which it described as rapidly developing, and noted that Thomson Reuters already trains its own AI search product on the headnotes. That Thomson Reuters had licensed them to no one else did not defeat the market. ROSS's claimed public benefits, wider access to the law, unhindered AI development and national security, failed for lack of supporting proof.
Where the intermediate-copying argument ran out
ROSS's strongest argument came from software law. In Google v. Oracle (2021), and in the Ninth Circuit's Sega (1992) and Connectix (2000) decisions, copying code as an intermediate step was fair because it was the only way to reach functional elements the law leaves free, such as the interfaces that let programs work together. The Third Circuit accepted that reading of the cases and applied it against ROSS. The unprotected material here was the judicial opinions, and ROSS had all of them. It could have written its training questions from the opinions. It used the headnotes because they were easier, and the court wrote that "ease is not a justification for copying."
That sentence moves the debate from the abstract question whether training transforms a work to a practical one: whether the developer could have built its training set from material it was free to use. Engineers will recognize the distinction. In supervised learning, the labels carry the expertise. ROSS took the editorial layer, the part where Westlaw's editors had already judged what each opinion stands for, and it did so to build a competing research service. For anyone who studies law, practices alone or teaches, the decision changes little about access to the law itself: opinions remain free, and the court said so twice.
What the opinion leaves to the generative-AI cases
Footnote 7 marks the limit of the holding. On 1 September 2026 the Department of Justice filed a statement of interest in In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-3143 in the Southern District of New York. The government relied on Bartz v. Anthropic (N.D. Cal. 2025) to argue that training a large language model that can "generate original responses" is transformative, and that such training did not produce "substitutive competition". The panel answered that ROSS's system cannot generate original expression and was trained to be a commercial substitute, so those concerns "do not apply here". It also noted that the Department of Justice filed nothing in this appeal. The decision binds federal courts in Delaware, New Jersey, Pennsylvania and the Virgin Islands; for model developers sued elsewhere it is persuasive authority, and its reasoning on necessity and on training-license markets will be cited against them.
The decision lands in the same week California set its own rules for lawyers who use these tools. As our reading of California SB 574, signed on 30 September 2026, set out, from 1 January 2027 an attorney there must take reasonable steps to verify every output and every case and statutory citation a generative AI tool produces. Taken together, the two instruments of that week address both ends of legal AI: the data a tool is trained on, and the output a lawyer files.
How Quentir Reads It
The most durable part of the opinion is its treatment of necessity. In the interoperability cases the court discussed, copying was fair because it was the only route to unprotected functional elements, and in Oracle the court also stressed the special character of declaring code. ROSS copied an expert's annotations because they saved time, and that weighed against it on the first and third factors. Fair use remains a four-factor judgment, and this one turned on protected headnotes taken by a direct competitor. Within those limits, the reasoning favors developers who can show how their training sets were assembled, from which sources, and why any protected portions were needed. It also gives weight to curated annotation: on these facts the court treated the market for licensing protected headnotes as AI training data as the owner's to exploit, and owners of comparable editorial work will cite that finding.
The open question is whether the generative cases follow Bartz or ROSS on transformation. Contract drafting cannot wait for the answer. Content licenses that grant access "as permitted by the functionality" of a platform say nothing about training, and after this opinion that silence has a price. Readers following this thread across copyright, state legal-practice rules and federal AI policy can read the full set of Quentir's paid analyses alongside this archive through one All-access membership. The next decision to watch is the Southern District of New York's ruling on fair use in the OpenAI litigation, where the government has already taken a side.
Sources: U.S. Court of Appeals for the Third Circuit, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153, precedential opinion (Montgomery-Reeves, J.), argued 11 June 2026, filed 29 September 2026, on appeal from D. Del. No. 1:20-cv-00613; 17 U.S.C. 107; Supreme Court of the United States, Google LLC v. Oracle America, Inc., 593 U.S. 1 (2021); Callaghan v. Myers, 128 U.S. 617 (1888). Other decisions and the Department of Justice statement of interest of 1 September 2026 in No. 1:25-md-3143 (S.D.N.Y.) are cited as described in the Third Circuit opinion. Sources checked 5 October 2026.
Published intelligence, built to inform your own decisions. Published: October 5, 2026.