Third Circuit rejects fair-use defense in Westlaw AI training dispute

The US Court of Appeals for the Third Circuit has affirmed a partial summary judgment holding that ROSS Intelligence infringed copyrights in Westlaw headnotes and could not establish a fair-use defense for its use of those materials. Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir. filed Sept. 29, 2026). Although the dispute involved artificial intelligence (AI), the court declined to treat the technology as presenting an entirely new legal problem. Instead, it described the appeal as an “ordinary copyright case” requiring application of established principles of originality and fair use.

The case involved Westlaw, Thomson Reuters’ legal-research platform. Westlaw organizes judicial opinions through editorial features that include headnotes identifying points of law addressed in the opinions. According to the court, Thomson Reuters’ editors prepare those headnotes under detailed guidelines requiring clarity, concision, accuracy, sufficient context and attention to the most important points in an opinion. Although editors generally follow the language of the underlying decision, each headnote must provide enough context to stand on its own.

ROSS developed an AI-powered legal search engine that responded to natural-language questions by retrieving relevant passages from judicial opinions. The system was not generative: it did not create new expression but instead returned passages from a collection of approximately 10 million uncopyrighted judicial opinions. To train the system to recognize which passages were responsive to particular legal questions, ROSS retained LegalEase Solutions to prepare approximately 25,000 training memoranda. The memoranda paired legal questions with judicial-opinion passages ranked according to their responsiveness.

The record established that LegalEase and a subcontractor used thousands of Westlaw headnotes to formulate the memoranda’s legal questions. The responsive passages were frequently the same opinion passages that Westlaw linked to the corresponding headnotes. The memoranda were subsequently converted into a machine-readable format and used to train the ROSS system. For purposes of the appeal, ROSS did not dispute that the conduct of LegalEase and its subcontractor was attributable to ROSS or that they had copied Westlaw headnotes.

The district court granted Thomson Reuters partial summary judgment as to 2,243 headnotes. It found the corresponding memorandum questions sufficiently similar to the headnotes, and sufficiently dissimilar from the underlying judicial opinions, that no reasonable jury could conclude the headnotes had not been copied. The Third Circuit affirmed the district court’s determination that the headnotes possessed the “creative spark” required for copyright protection. It also affirmed the conclusion that ROSS’s use was not fair.

Two facts were central to the appellate court’s framing of the case. First, Thomson Reuters’ editorial materials contained protectable original expression even though they were based on judicial opinions that they were not protected. Second, ROSS intended to compete directly with Westlaw while using the copied materials for a highly similar purpose. The record included advertisements comparing ROSS with Westlaw, evidence that ROSS priced its service in line with Westlaw, and evidence that ROSS executives hoped their product would serve as a Westlaw substitute. Some law firms switched from Westlaw to ROSS.

The court did not decide whether Westlaw’s Key Number System was independently original. Although that issue appeared in the questions certified for interlocutory review, the court concluded that ROSS had forfeited it by failing to challenge the system’s originality in its briefing. The appellate holding therefore should not be described as a separate determination that the Key Number System was copyrightable.

The decision is significant, but its factual boundaries matter. ROSS used copied editorial material to train a non-generative legal search product intended to compete directly with the platform from which the material originated. The opinion therefore does not establish that every use of copyrighted material in AI training is infringing or that every AI-training fair-use defense must fail. It confirms more specifically that the involvement of machine learning does not displace ordinary copyright analysis where protectable expression is copied and used for a closely similar competitive purpose.