Can you copy someone else’s copyrighted legal summaries to train an AI search tool — and call it fair use? Fair use is a US copyright doctrine that sometimes lets you use protected material without permission, depending on how you use it and what harm that does to the original market.
On 29 September 2026 (the opinion’s caption date; the docket header on each page also shows 30 September), the US Court of Appeals for the Third Circuit answered no in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153. Judge Montgomery-Reeves wrote the opinion for a panel of Restrepo, Montgomery-Reeves and Bove. The court affirmed partial summary judgment for Thomson Reuters [1].
Law firm Ballard Spahr calls this the first federal appellate decision on fair use in AI training. That “first” claim is the firm’s characterisation; the opinion itself does not say it [2].
What was ROSS actually building?
Per the opinion, ROSS’s search engine returned passages from about ten million uncopyrighted judicial opinions and “was not a generative AI.” LegalEase wrote about 25,000 training memos, using Westlaw headnotes to frame questions. ROSS advertised itself directly against Westlaw at comparable prices [1].
Westlaw headnotes are short editorial summaries of points of law in a case. The court held that the 2,243 headnotes at issue have “some creative spark,” and that each headnote is a copyrightable work. ROSS forfeited any challenge to Westlaw’s Key Number System, which the court did not address [1].
This was an interlocutory appeal — a mid-case appeal on two certified questions, originality and fair use — under 28 U.S.C. § 1292(b). The trial judge was Circuit Judge Stephanos Bibas, sitting by designation. The appeal was argued on 11 June 2026 [1].
How did the four fair-use factors land?
US courts weigh four fair-use factors. The Third Circuit found [1]:
- Factor 1 (purpose and character of the use): against ROSS. The use was highly commercial and “minimally transformative, at best,” because it served the same purpose as Westlaw’s: building a legal-research tool.
- Factor 2 (nature of the copyrighted work): slightly for ROSS, because headnotes are more factual.
- Factor 3 (amount and substantiality): against ROSS, because whole headnotes were copied and copying them was not necessary when the opinions were freely available. The court rejected ROSS’s argument that it took only 0.08% of 28 million headnotes.
- Factor 4 (market effect): against ROSS, on harm to the legal-research market and to a “rapidly developing” market for licensing headnotes as AI training data. Thomson Reuters already trains its own AI products on them. The court found ROSS gave no evidence for its public-benefit, AI-development and national-security arguments.
The court summed it up: “this is no more than an ordinary copyright case” [1].
What about the Justice Department’s OpenAI brief?
In footnote 7, the court notes the Department of Justice’s 1 September 2026 statement of interest in In re OpenAI and says those concerns “do not apply here.” “Unlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression.” It adds that the DOJ “knows how to assert its interests, but… notably did not do so here” [1].
What was that DOJ argument? In an advisory brief (not a ruling) filed in the consolidated OpenAI copyright MDL in the Southern District of New York, the United States asked the court to reject the view that training an LLM on copyrighted text is infringement in itself. Quoting the March 2026 White House AI framework, the Department said that, consistent with that framework, the “training of AI models on copyrighted material,” in and of itself, “does not violate copyright laws.” It separated training copies from outputs that reconstruct a work, and from a market-dilution theory [3].
The Third Circuit’s point: that generative-AI training debate is not this case.
What this does not prove
- This was a non-generative tool. The ruling does not settle fair use for chatbots that generate new text. Ballard Spahr stresses that generative-AI questions remain open [2].
- Interlocutory only. Remaining district-court issues were not reviewed. Whether ROSS will seek rehearing or Supreme Court review is unknown [1].
- An internal inconsistency in the opinion: page 22 refers to “the 25,000 Westlaw-written headnotes,” while page 5 says about 25,000 memos and the holding covers 2,243 headnotes. The research note flags this without resolving it [1].
The Bottom Line
The Third Circuit held that thousands of Westlaw headnotes are copyrightable and that ROSS’s commercial, same-purpose use of them to train a non-generative legal search tool was not fair use — including because of harm to a developing market for licensing headnotes as AI training data. Ballard Spahr calls it the first appellate AI fair-use ruling; attribute that label to the firm. For generative AI cases, the court itself pointed elsewhere: the DOJ’s OpenAI arguments, it said, do not apply here.
Sources
- US Court of Appeals for the Third Circuit opinion PDF (via Courthouse News), Thomson Reuters v. ROSS Intelligence, No. 25-2153, filed 29 September 2026. https://www.courthousenews.com/wp-content/uploads/2026/09/opinion-thomson-reuters-v-ross-intelligence-ai-third-circuit-affirm-partial-summary-judgment.pdf
- Ballard Spahr, 2 October 2026. https://www.ballardspahr.com/insights/alerts-and-articles/2026/10/third-circuit-addresses-fair-use-in-ai-training-but-leaves-generative-ai-questions-unresolved
- United States Statement of Interest, Document 1682, In re OpenAI, Inc. Copyright Infringement Litigation, No. 25-md-3143, 1 September 2026 (Copyright Alliance PDF). https://copyrightalliance.org/wp-content/uploads/2026/09/2026-09-01-Letter-dckt-1682_0.pdf
