Third Circuit rules ROSS's AI training was not fair use
The court held 2,243 Westlaw headnotes are copyrightable and found ROSS's use of them to train its legal AI minimally transformative at best.
Key takeaways
- The Third Circuit affirmed a February 2025 partial summary judgment holding that 2,243 Westlaw headnotes are copyrightable and that ROSS Intelligence's use of them to train its legal-research AI was not fair use, according to the opinion filed September 29, 2026.
- The court distinguished Bartz v. Anthropic and In re OpenAI because ROSS's AI was not generative: it only returned passages from preexisting judicial opinions, so it could not generate original expression.
- The court found a "rapidly developing" market for licensing headnotes as AI training data and said ROSS usurped Thomson Reuters's opportunity to enter it.
- ROSS shut down its platform in 2021 and says it will seek Supreme Court review; the case returns to the district court, where Thomson Reuters's tortious interference claim is still pending.
The U.S. Court of Appeals for the Third Circuit on September 29, 2026 affirmed a ruling that Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. hold valid copyrights in 2,243 Westlaw headnotes and that ROSS Intelligence Inc.'s use of those headnotes to train its legal-research AI was not fair use, the first time a federal appeals court has decided whether training an AI system on copyrighted material is fair use.
The three-judge panel affirmed Judge Stephanos Bibas's February 2025 partial summary judgment in the U.S. District Court for the District of Delaware, according to the opinion, filed September 29, 2026. The court released the full text of the opinion the following day, and LawNext reported its contents. Judges L. Felipe Restrepo, Montgomery-Reeves and Emil J. Bove sat on the panel, and Montgomery-Reeves wrote the opinion.
The court held that the headnotes are original enough for copyright protection. Under the Supreme Court's Feist standard, a work needs only "some creative spark," and the court found that Thomson Reuters's editors made creative decisions about which points of law to include in each headnote and how to word them. "All 2,243 independently created headnotes at issue in this appeal have 'some creative spark,'" the opinion states.
On fair use, the court weighed the four statutory factors. The second factor, the nature of the copyrighted work, weighed slightly in favor of fair use, because headnotes are more factual than fictional. The first, third and fourth factors weighed against it, according to the opinion. The court found ROSS's use was "highly commercial and minimally transformative": both Thomson Reuters and ROSS used the headnotes to build a legal-research platform that helps users find responsive legal material.
The court rejected ROSS's argument that it copied only 0.08 percent of Westlaw's 28 million headnotes. Because each headnote is a copyrightable work in its own right, ROSS copied an entire work with each one it took, the court held.
Central to the ruling was the court's finding that ROSS did not need the headnotes. ROSS had access to roughly ten million uncopyrighted judicial opinions and could have copied those instead. "Unlike necessity, ease is not a justification for copying," the opinion states.
The court also found that a market for licensing headnotes as AI training data is "rapidly developing," and that by copying the headnotes without authorization, ROSS "usurped Thomson Reuters's opportunity to enter that derivative market."
The court took pains to limit the decision's reach. It held that ROSS's AI was not generative: the system "would not create any new expression" and "would only return text passages from preexisting judicial opinions." For that reason, the court distinguished Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025), and In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-3143 (S.D.N.Y. Sept. 1, 2026), which addressed whether training a large language model that "can generate original responses" is transformative. "Unlike the AI models in Bartz and In re: OpenAI, ROSS's AI platform cannot generate original expression, and the evidence here supports the opposite conclusion about transformativeness," the court wrote.
The court framed the dispute, which began when Thomson Reuters sued ROSS in 2020, as unremarkable. "Under ROSS's framing, this case appears to concern the future of AI legal technology," the opinion states. "But appearances can be deceiving. In truth, this is no more than an ordinary copyright case."
ROSS, founded by three University of Toronto computer science students, built an AI search engine that answered plain-language legal questions with passages from judicial opinions. It shut down its platform in 2021, citing the cost of the litigation, as Reuters reported in an account relayed by 24-ore. The case had run about 6.5 years by the ruling, LawNext reported. Thomson Reuters owns Reuters News.
Thomson Reuters said it "firmly believes that respecting copyright is essential for fostering innovation," explainx.ai reported.
For AI video creators, the ruling signals that the market for licensing copyrighted material as training data is one courts will weigh, and that a rights holder's lost licensing opportunity can count as market harm even where no such market yet sells the material. The decision is narrow, though: it turns on the fact that ROSS's system was not generative, and it does not decide the pending fair-use cases over models that produce new output.
The case returns to the District Court for the District of Delaware, where Thomson Reuters's tortious interference claim and questions about headnotes beyond the 2,243 resolved on summary judgment remain pending. ROSS says it will ask the Supreme Court to review the ruling.