In Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir. Sept. 29, 2026), the United States Court of Appeals for the Third Circuit became the first federal appellate court to decide whether copying copyrighted material to train an artificial intelligence system qualifies as fair use, and it held that it did not where a would-be competitor copied 2,243 Westlaw headnotes to build a rival legal research platform. Writing for a unanimous panel in a precedential opinion (Restrepo, Montgomery-Reeves, and Bove, JJ.), Judge Tamika R. Montgomery-Reeves affirmed partial summary judgment for Thomson Reuters and declined ROSS’s invitation to treat the appeal as a referendum on the future of legal AI, characterizing the dispute instead as “no more than an ordinary copyright case.” The decision supplies the first circuit-level framework for AI-training fair use, while expressly reserving the questions presented by generative models.
Background
Thomson Reuters and West Publishing Corporation sued ROSS Intelligence in the District of Delaware in 2020, alleging copyright infringement and tortious interference with contract. ROSS, unable to obtain a Westlaw license, had engaged a third party, LegalEase Solutions, to prepare “Bulk Memos” — question-and-answer training data derived from Westlaw headnotes — which ROSS used to train a natural-language legal search tool that returned passages from judicial opinions.
Judge Stephanos Bibas, a Third Circuit judge sitting by designation, initially sent fair use to a jury in 2023, then reversed course. In a revised opinion, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025), he granted partial summary judgment on 2,243 headnotes, finding them original and copied, and rejected fair use, assigning the first and fourth statutory factors to Thomson Reuters and the second and third to ROSS. He then certified two questions for interlocutory appeal under 28 U.S.C. § 1292(b): whether the headnotes and the West Key Number System are original as a matter of law, and whether ROSS’s use of the headnotes was fair. The panel heard argument on June 11, 2026.
Originality of the Headnotes
Applying the “extremely low” threshold of Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991), the court held that each of the 2,243 headnotes possesses the requisite creative spark. The editorial process mattered: Thomson Reuters attorney-editors draft headnotes under detailed guidelines requiring that each be concise and capable of standing on its own — conveying enough factual background and reasoning to be understood without the opinion or any other headnote — while accurately reflecting the underlying passage. Those choices of selection, distillation, and expression, rather than the public-domain text of the opinions themselves, supply the protectable authorship.
The panel located this result within a long line of authority. It read Callaghan v. Myers, 128 U.S. 617 (1888), and Georgia v. Public.Resource.Org, Inc., 590 U.S. 255 (2020), as consistent with copyright in reporter-prepared headnotes, and it distinguished Matthew Bender & Co. v. West Publishing Co., 158 F.3d 674 (2d Cir. 1998), which denied protection to West’s star pagination and parallel citations because industry convention dictated them — while describing headnotes as independently composed. The court further held that each headnote is a separately copyrightable work, a conclusion that proved consequential under the third fair use factor, and it left open whether headnotes that quote an opinion verbatim are protected.
Fair Use
The court’s fair use analysis rests principally on Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023), and finds that three of the four statutory factors weigh against ROSS.
Purpose and character. ROSS’s use was commercial and, in the court’s words, “minimally transformative, at best.” The panel acknowledged that using headnotes as an intermediate step to train a model arguably reflected a slight difference in use, but concluded that both parties ultimately employed the headnotes for the same end: building and improving a legal research platform. The court rejected ROSS’s reliance on Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015), where the search function served a different purpose and could drive book sales, and it distinguished the intermediate-copying line of Sega Enterprises Ltd. v. Accolade, Inc., 977 F.2d 1510 (9th Cir. 1992), and Google LLC v. Oracle America, Inc., 593 U.S. 1 (2021), on the ground that ROSS faced no comparable necessity. ROSS had the judicial opinions themselves and could have prepared its own training memoranda from them. As the court put it, “Unlike necessity, ease is not a justification for copying.”
Nature of the work. Consistent with its originality holding, the panel treated the headnotes as more factual than creative for purposes of the second factor, which it weighed slightly in ROSS’s favor.
Amount and substantiality. Here the panel departed from the district court. Because each headnote is a separate work, and ROSS’s training data incorporated the substance of each headnote in full, ROSS took entire copyrighted works, and the third factor weighed against fair use.
Market effect. The fourth factor produced the decision’s most consequential holding. Crediting evidence that a market for licensing headnotes as AI training data is rapidly developing — and that Thomson Reuters already trains its own AI research products on that content — the court held that ROSS’s unlicensed copying usurped a derivative market Thomson Reuters was entitled to exploit. Drawing on Video Pipeline, Inc. v. Buena Vista Home Entertainment, Inc., 342 F.3d 191 (3d Cir. 2003), the panel emphasized that the inquiry reaches the effect of the use on the value of the copyrighted work, not merely on a narrowly defined existing market. The court summarized its holding directly: Thomson Reuters’ materials are copyrightable, and ROSS’s use was not fair.
The Boundaries of the Holding
The opinion is careful about its own reach. In a footnote, the panel distinguished generative AI systems that produce new expression from ROSS’s tool, which retrieved and returned existing opinion passages. The decision therefore does not resolve the questions pending in the large-language-model cases, and it sits in tension, at least in tone, with the 2025 district court rulings in Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal. June 23, 2025), and Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. June 25, 2025), each of which found training on lawfully acquired books to be fair use on the records presented. Industry participants on both sides have characterized the Third Circuit’s narrow framing as favorable to their positions.
What the opinion does establish, as binding precedent within the Third Circuit and persuasive authority elsewhere, is that the presence of AI does not alter the ordinary fair use calculus; that intermediate training copies are judged by the ultimate purpose they serve; and that an emerging market for licensing content as training data is a cognizable market under the fourth factor. The latter point is likely to feature prominently in every pending training-data dispute, generative or not.
The case now returns to the District of Delaware, where questions concerning thousands of additional headnotes, the expiration of certain copyrights, and remaining defenses are set for trial.
Practical Significance
For publishers, media companies, and other owners of editorial content, the decision rewards the disciplined documentation of authorship. Thomson Reuters’ registrations and written editorial guidelines were central to establishing originality in material that summarizes public-domain text. Content owners that license, or intend to license, their catalogs for AI training should also document those programs; the existence of a developing licensing market carried decisive weight under the fourth factor.
For developers and acquirers of AI tools, ROSS underscores that diligence over the provenance of training data is a substantive legal question, not an operational one. Where a developer trains on a competitor’s protected content to build a substitute product, particularly after being refused a license, the Third Circuit has now made clear that the label “AI training” will not carry a fair use defense on its own.