Appeals court rules training a rival AI on Westlaw headnotes was not fair use

The US Court of Appeals for the Third Circuit affirmed that ROSS Intelligence infringed Thomson Reuters’s copyright when it used Westlaw’s headnotes to train an AI legal-research tool, and that the copying was not fair use. It was the first ruling by a federal appeals court on whether training an AI system on copyrighted material can be fair use. The three-judge panel, in an opinion by Judge Montgomery-Reeves, upheld the February 2025 summary judgment granted by Judge Stephanos Bibas, sitting by designation in Delaware. The panel was unanimous. The opinion was initially filed under seal pending proposed redactions.

ROSS had hired a contractor to write about 25,000 training memos built around Westlaw headnotes, the editors’ summaries of points of law that sit above each judicial opinion. It used them to teach a search engine to return passages from judicial opinions in answer to plain-language questions. The court held that 2,243 headnotes were original enough for copyright protection. Three of the four fair-use factors went against ROSS. The use was “highly commercial and minimally transformative” because it served the same purpose as Westlaw. Without a transformative purpose, the amount copied could not be justified. And it harmed Thomson Reuters both in its own market and in what the panel called a “rapidly developing” market for licensing headnotes as AI training data. Only the nature of the work, factual legal summaries, weighed slightly the other way.

The court said repeatedly that this was “no more than an ordinary copyright case”. It stressed that ROSS’s system “was not a generative AI” and could only return existing text. In a footnote it separated the case from the Justice Department’s filing backing fair use for large-language-model training in the litigation against OpenAI. It also distinguished the Bartz v. Anthropic and Kadrey v. Meta rulings, saying ROSS “cannot generate original expression” and had trained its system to be “a commercial substitute for Westlaw”. It rejected ROSS’s argument that the ruling would halt AI development as unsupported by evidence.

Ballard Spahr’s analysis said the opinion did not answer the larger questions about generative AI, because the technology before the court generated no new expression; whether training a generative model is transformative, and whether its output causes market harm, remain open. Authors Alliance, which had filed an amicus brief supporting ROSS, expected limited effect on the larger cases. It argued that the panel had hollowed out the merger doctrine and disfavoured the non-expressive copying that fair use had long protected. The appeal was interlocutory: at the parties’ request, the district court had certified the copyrightability and fair-use questions for review before the case went any further.