US Third Circuit rules ROSS's use of 2,243 Westlaw headnotes to train a non-generative legal AI was not fair use, while distinguishing generative LLM cases

Why it matters

【Key points】

On Sept 29 the US Court of Appeals for the Third Circuit (Thomson Reuters v. ROSS Intelligence, No. 25-2153) affirmed partial summary judgment: Westlaw's editorial headnotes are copyrightable, and ROSS's use of 2,243 of them to train its AI legal search engine was not fair use. Judge Montgomery-Reeves wrote for a panel with Judges Restrepo and Bove.

ROSS had LegalEase write about 25,000 training memos that used headnotes to frame legal questions, paired with labeled opinion passages. The court stressed ROSS's AI "was not a generative AI" — it only returned existing opinion text — and competed with Westlaw at comparable prices.

The second factor leaned slightly toward fair use; the first, third and fourth weighed against. The use was commercial and "minimally transformative, at best." On market harm, the court found the market for licensing headnotes as AI training data "rapidly developing," noted Thomson Reuters trains its own AI on them, and held ROSS usurped that derivative market. Halt-AI and national-security arguments failed for lack of evidence.

【Verifiable】

Opinion PDF: https://www2.ca3.uscourts.gov/opinarch/252153p.pdf

Justia: https://law.justia.com/cases/federal/appellate-courts/ca3/25-2153/25-2153-2026-09-30.html

【Our read】

1. A federal appeals court has squarely held that training AI on copyrighted material can fail fair use, and recognized AI-training licensing as a protected derivative market. Plaintiffs elsewhere will cite that.

2. It is deliberately narrow. Footnote 7 notes the DOJ's Sept 1 statement in the In re: OpenAI MDL, which relied on Bartz v. Anthropic to call training LLMs that "generate original responses" transformative, and says those concerns "do not apply here." Do not read it as "training LLMs is not fair use."

3. The closer a training use sits to the source's market, the weaker the defense; vertical AI tools that compete with their data suppliers carry the most risk.