AI Training and Copyright: Lessons Learned From the Third Circuit’s Decision in Thomson Reuters v. ROSS

The court held that Westlaw’s editorial headnotes are protected by copyright and that ROSS Intelligence’s use of those headnotes to train a competing, non-generative AI legal research tool was not fair use. The opinion was careful to distinguish generative AI, and the core questions in the pending generative AI cases remain open. Even so, the decision offers lessons for both AI developers and content owners.

The Dispute 

Thomson Reuters, the owner of the Westlaw legal research platform, alleged that ROSS Intelligence, a developer of an AI-based legal search tool, used Westlaw editorial headnotes to create training materials for ROSS’s platform. Westlaw’s headnotes are editorial summaries of discrete points of law from judicial opinions. ROSS obtained the training materials from a third party, LegalEase, after Thomson Reuters declined to license its content to ROSS.

The district court granted partial summary judgment to Thomson Reuters, finding that 2,243 headnotes satisfied copyright’s originality requirement and that ROSS’s copying was not protected by fair use. The court left other issues, including thousands of additional headnotes and whether some copyrights had expired, for trial.

The Third Circuit’s Opinion

The Third Circuit affirmed in an opinion by Judge Tamika Montgomery-Reeves. The court reviewed the two questions certified for interlocutory appeal, originality and fair use, and characterized the dispute as “no more than an ordinary copyright case.” Its analysis of the fair use factors differed in part from the district court’s: The Third Circuit found that the first, third, and fourth factors weighed against fair use and that only the second factor weighed slightly in ROSS’s favor.

The ruling turns on two issues likely to recur in future AI copyright disputes: protectability and fair use.

On protectability, the Third Circuit held that the 2,243 headnotes at issue had the “creative spark” copyright requires, because editors decided which points of law to include and how to express them so each headnote could stand on its own. The court rejected ROSS’s argument that protecting headnotes would create a monopoly over the law, noting that judicial opinions remain freely available. The court did not decide whether headnotes that quote opinions verbatim are protectable, and it did not reach the originality of the West Key Number System, which ROSS did not challenge on appeal.

On fair use, the Third Circuit found ROSS’s use commercial and “minimally transformative, at best,” because both companies used the headnotes to build and improve a legal research platform. It rejected ROSS’s reliance on the intermediate-copying cases involving computer code, reasoning that ROSS had access to the underlying opinions and did not need to copy the headnotes: “Unlike necessity, ease is not a justification for copying.” On the third factor, the court found that ROSS copied entire headnotes, each a copyrightable work, without a transformative purpose. On the fourth factor, the court identified harm both to the legal research market and to what it called a “rapidly developing” market for licensing headnotes as AI training data. In a footnote, the court also noted evidence that ROSS accessed Westlaw in violation of its terms of service and concluded that, to the extent good faith remains relevant, it weighed against fair use.

Why It Matters

The Third Circuit framed its decision narrowly, emphasizing that ROSS’s tool could not generate original expression and was built to substitute for Westlaw. How far the opinion reaches remains to be seen. It may be read as limited to non-generative tools built to compete directly with the source, but its reasoning on necessity and market harm may also be invoked in future disputes involving fine-tuned or specialized models. No federal appellate court has yet resolved how the fair use factors apply to generative AI training.

Key Takeaways

• Curated content may be protectable. The Third Circuit confirmed that editorial headnotes clear copyright’s low originality threshold even though they summarize public-domain judicial opinions, while also treating them as more factual than creative for fair use purposes.
• Fair use for AI training remains case by case. Courts will continue to apply the traditional fair use factors, including commercial purpose, transformativeness, amount used, and market effect, and the result may vary with the facts.
• Purpose and substitution matter. The court’s analysis turned on its finding that ROSS used the headnotes for the same ultimate purpose as Westlaw and aimed to replace it. Using copyrighted material with a distinct purpose, or that do not substitute for the original, present different questions.
• How training data is sourced matters. The court emphasized that ROSS could have worked from freely available opinions and chose headnotes because they were easier, and it noted evidence of access in violation of Westlaw’s terms of service. Under the court’s analysis, the source of training data, the reasons for using it, and compliance with applicable license and access terms were all relevant to fair use. The court also recognized a developing market for licensing curated content as AI training data.
• Generative AI questions remain open. The Third Circuit expressly distinguished generative AI, leaving open how courts will weigh transformativeness and market harm for models that create new content.