Facts
Thomson Reuters owns Westlaw, whose editors write headnotes summarizing the points of law in judicial opinions. ROSS Intelligence built an AI legal search engine that answers plain-language questions with passages from existing opinions and does not generate new text. After its request to license Westlaw content was refused, ROSS had LegalEase prepare about 25,000 training memos, whose drafters used thousands of Westlaw headnotes to frame the legal questions. ROSS marketed its platform as a substitute for Westlaw at comparable prices. Thomson Reuters sued in the U.S. District Court for the District of Delaware. In February 2025 that court found 2,243 headnotes copied, rejected fair use on partial summary judgment and certified both questions for interlocutory appeal.
What was the legal question in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc.?
The Third Circuit had two certified questions before it. The first was whether the 2,243 headnotes were original enough for copyright protection, given that they restate points of law from judicial opinions that anyone may copy. ROSS argued that protecting them would give Thomson Reuters a monopoly over the law and that the merger doctrine applied. The second was whether ROSS’s use of the headnotes to train its AI tool was fair use under Section 107 of the Copyright Act, weighing the purpose and character of the use, the nature of the work, the amount used and the effect on the market. ROSS relied on Authors Guild v. Google and on intermediate-copying cases such as Google v. Oracle.
What did the U.S. Court of Appeals for the Third Circuit decide?
Judge Montgomery-Reeves, writing for the panel, affirmed. The headnotes reflected editorial choices about which points of law to include and how to state them, enough for the required creative spark, and merger did not apply because points of law can be stated in many ways. ROSS’s use was commercial and at best minimally transformative, serving the same purpose as Westlaw. Unlike the intermediate-copying cases, copying was unnecessary because the opinions were freely available, and ease is no justification. Only the second factor favored ROSS; the third weighed against fair use, unlike in the District Court’s analysis. Under the fourth, ROSS diminished the headnotes’ value as a draw for Westlaw subscribers and usurped Thomson Reuters’s opportunity to license AI training data.
Why does Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. matter?
The decision is regarded as the first ruling by a U.S. federal court of appeals on fair use in AI training, and it went against the AI developer. The court described the dispute as an ordinary copyright case and tied its reasoning to the facts. ROSS’s tool generated no new expression, was built to substitute for the plaintiff’s product, and did not need the headnotes to reach the unprotected opinions. The court distinguished generative models, citing Bartz v. Anthropic and the Department of Justice’s statement of interest in the OpenAI litigation, so its reach for large language models remains open.
Related stages
First instance: U.S. District Court for the District of Delaware, Judge Stephanos Bibas (a Third Circuit judge sitting by designation), memorandum opinion of February 11, 2025 (No. 1:20-cv-613-SB), revising his 2023 ruling that had left most questions for a jury. The Third Circuit heard argument on June 11, 2026 before Judges Restrepo, Montgomery-Reeves and Bove. The opinion was first filed under seal on September 29, 2026; the public version is posted on the court’s website. Because the appeal concerned a partial summary judgment, the remaining issues, including whether some headnote copyrights have expired and the headnotes left for trial, stay with the District Court. In early October 2026 ROSS’s counsel said it would seek Supreme Court review. For generative models, see Kadrey v. Meta (N.D. Cal., June 2025).
The same question elsewhere
May AI be trained on copyright works?
Building the dataset fell within the text and data mining exceptions.
Kneschke v LAION e.V.Germany · 2025NoMemorising lyrics in the model was reproduction, outside the exception.
GEMA v OpenAIUsing headnotes to build a competing product was not fair use (non-generative AI).
— this decisionStoring news content to train the model was held prima facie to be fair dealing; an appeal is pending.
ANI Media Pvt. Ltd. v. Open AI OpCo LLC