Third Circuit Rules Westlaw Headnotes Are Protected and ROSS's AI Training Was Not Fair Use

Published October 9, 2026

Score

Why it matters

The U.S. Court of Appeals for the Third Circuit ruled on September 29, 2026, that Thomson Reuters' Westlaw headnotes are protected by copyright and that ROSS Intelligence infringed them by using about 2,243 of them to train a legal-research AI tool. Judge Tamika Montgomery-Reeves wrote the opinion, joined by Judges Restrepo and Bove. The panel affirmed a February 2025 partial summary judgment by Judge Stephanos Bibas in the District of Delaware. The opinion was issued under seal and made public shortly afterward. The court called the case "no more than an ordinary copyright case." It is the first federal appellate decision on fair use in AI training. The court found that each headnote meets the low originality threshold. It weighed the four fair-use factors against ROSS, calling its use commercial and only minimally transformative because the tool served the same purpose as Westlaw: helping lawyers find relevant case law. The court also found market harm in two areas: the existing legal-research market, where ROSS sold a direct substitute, and the emerging market for licensing headnotes as AI training data. The dispute began in 2020, when Thomson Reuters sued ROSS. ROSS had hired third parties to write training memoranda that drew on Westlaw headnotes to frame legal questions. ROSS shut down in 2021 due to litigation costs but pursued an interlocutory appeal, with oral argument held June 11, 2026. Ten amicus briefs supported ROSS.

The court stressed that ROSS's product was non-generative, so the holding does not resolve how generative models trained on copyrighted text should be treated. Commentators say the ruling's reach may be limited for that reason. ROSS has said it will ask the Supreme Court to review the decision. The timing and outcome of any petition are not yet known.

Attorneys handling AI training disputes should read the opinion's fair-use analysis closely, particularly the market-harm finding that treats a licensing market for training data as cognizable. That reasoning could affect how courts assess pending generative AI suits, though the non-generative limitation gives defendants a clear distinction to argue. Clients who train models on licensed or scraped legal content should review their data sourcing and licensing terms now. Watch for a Supreme Court petition from ROSS and for how district courts cite the decision in the dozens of generative AI cases already pending.