Third Circuit affirms Thomson Reuters win over ROSS on Westlaw headnotes and AI training

The U.S. Court of Appeals for the Third Circuit held on September 29, 2026, that Westlaw headnotes are original enough for copyright protection and that ROSS Intelligence’s use of 2,243 of them to train an AI legal research tool was not fair use. The precedential opinion, by Judge Montgomery-Reeves for a panel with Judges Restrepo and Bove, affirms the partial summary judgment Judge Stephanos Bibas granted Thomson Reuters in February 2025. The opinion was first filed under temporary seal and is now posted on the court’s website. The case is Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153.

How the case reached the Third Circuit

Thomson Reuters and West Publishing Corp. sued ROSS in the District of Delaware (No. 1:20-cv-00613), alleging copyright infringement and tortious interference with contract. According to the opinion, ROSS built a search engine that answered plain-language legal questions with passages from about ten million judicial opinions. To train it, ROSS hired LegalEase Solutions, which wrote about 25,000 memos, each posing a legal question with four to six responsive passages. The opinion states that memo drafters used thousands of Westlaw headnotes to frame the questions. ROSS did not contest on appeal that the copying was attributable to it.

On February 11, 2025, Judge Bibas, a Third Circuit judge sitting by designation, granted Thomson Reuters partial summary judgment on direct infringement and fair use (D.I. 770). On April 4, 2025, he granted ROSS’s motion for interlocutory appeal and a stay, certifying two questions: whether the West headnotes and Key Number System are original, and whether ROSS’s use of the headnotes was fair use (docket, D.I. 799). The Third Circuit granted leave to appeal on June 17, 2025 (D.I. 805) and heard argument on June 11, 2026.

Originality: “Headnotes are not law”

The panel held that each of the 2,243 headnotes shows the “creative spark” that Feist Publications v. Rural Telephone Service requires, because editors chose which points of law to include and how to word them so each headnote stands on its own. It relied on Callaghan v. Myers (1888) and Georgia v. Public.Resource.Org (2020), which it read as recognizing a private reporter’s copyright in headnotes.

The court rejected three counterarguments. To the claim that protection would give Thomson Reuters a monopoly over the law, it answered: “Headnotes are not law; judicial opinions are.” (p. 11). It held the merger doctrine inapplicable because points of law can be expressed many ways, citing an amicus brief from RELX, Inc. that, as the opinion describes it, shows Lexis, a Westlaw competitor, writing different headnotes for the same opinions. It distinguished Matthew Bender & Co. v. West Publishing Co. (2d Cir. 1998) as a case about party names and parallel citations, not headnotes.

Fair use, factor by factor

On the first factor, the court found ROSS’s use commercial, since ROSS priced its platform to compete with Westlaw, and “minimally transformative, at best” (p. 17), because both companies used the headnotes to help users find responsive law. Training an AI program was an intermediate step that “arguably presents a slight degree of difference in use” (p. 16). It distinguished Authors Guild v. Google and the software interoperability cases, Google v. Oracle, Sega v. Accolade and Sony v. Connectix, on the ground that those copiers needed the copying to reach unprotected elements. ROSS had the public opinions available. “Unlike necessity, ease is not a justification for copying.” (p. 21). A footnote adds that the undisputed evidence showed ROSS at times acted in bad faith, including attempts to reach Westlaw through a law-firm investor’s credentials and a student account.

The second factor favored ROSS, because the headnotes are published and more factual than fictional. The third weighed against it: ROSS copied each headnote in full, and the court treated each headnote as a complete work, rejecting ROSS’s point that it took only 0.08 percent of 28 million headnotes.

On the fourth factor, the court held ROSS’s copying diminished the headnotes’ value as a draw for Westlaw subscriptions and harmed Westlaw in the legal research platform market. It also recognized a derivative market in licensing headnotes as AI training data, finding the evidence showed that market “is rapidly developing” (p. 26) and that Thomson Reuters uses its own headnotes to train its AI search products. It rejected ROSS’s public-benefit arguments about access to law, AI development and national security for lack of evidence.

What the opinion does not decide

The panel drew limits in its footnotes. It did not decide whether headnotes that quote an opinion verbatim are original, because the judgment covered only headnotes that do not. It treated the originality of the Key Number System as forfeited, because ROSS never briefed it. And it distinguished a September 1, 2026 statement of interest the Justice Department filed in In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-3143 (S.D.N.Y.). There, relying on a 2025 Northern District of California decision, the department argued that training a large language model that can generate original responses is transformative. The panel said those concerns do not apply because ROSS’s platform “cannot generate original expression” (p. 17, n.7) and was built as a commercial substitute for Westlaw. It also noted that the department did not file in this case.

Who it applies to

The ruling binds the parties and, as a precedential opinion, the district courts of the Third Circuit, which covers Delaware, New Jersey, Pennsylvania and the U.S. Virgin Islands (28 U.S.C. § 41). Its fair-use analysis addresses a non-generative AI tool trained to compete with the source it copied. The opinion describes its own scope narrowly, calling this “no more than an ordinary copyright case” (p. 2). By contrast, Bloomberg Law and Copyright Lately described it as the first federal appellate ruling on fair use in AI training.

What happens next

Under Federal Rule of Appellate Procedure 40(d)(1), a petition for panel rehearing or rehearing en banc is due within 14 days after entry of judgment. Under Rule 41(b), the mandate issues seven days after that time expires, or seven days after an order denying a timely petition. Under Supreme Court Rule 13, a petition for certiorari is due within 90 days after entry of judgment, or after the denial of a timely rehearing petition. When the mandate issues, the case returns to the District of Delaware, where Judge Bibas stayed proceedings, including the pretrial conference and trial, pending the appeal (D.I. 799).

What is not yet known

The sources reviewed do not say whether either party will seek rehearing or Supreme Court review. The opinion does not address the complaint’s tortious interference claim or damages, and no schedule for further district court proceedings appears in the sources reviewed. The opinion’s own footnotes leave open whether verbatim headnotes are protected and whether the Key Number System is original. How courts will treat training of generative models remains before other courts, including in the OpenAI multidistrict litigation the panel distinguished.

Sources

The Third Circuit’s opinion of September 29, 2026; the district court’s February 11, 2025 memorandum opinion and docket, No. 1:20-cv-00613; reporting by Bloomberg Law, Copyright Lately and LawSites.