The First Appellate AI Fair-Use Ruling: Thomson Reuters v. ROSS and What It Leaves Open
A federal appeals court has ruled on fair use in AI training for the first time, and the AI company lost. The Third Circuit held that Westlaw's headnotes are protected by copyright and that ROSS Intelligence's use of them to train a competing legal-research tool was not fair use. The opinion is deliberately narrow: it turns on a tool that retrieved existing texts rather than generating new ones, so the larger fight over training generative AI remains open.
What happened.
Thomson Reuters operates Westlaw, the legal-research platform. Westlaw attaches editor-written headnotes to judicial opinions: short statements of a point of law, drafted under guidelines that limit each headnote to 800 characters where possible and require it to stand on its own. The judicial opinions themselves are not copyrightable. The headnotes are Thomson Reuters' editorial work product, and Thomson Reuters also uses them to train its own AI search products.
ROSS Intelligence built an AI legal search engine that answered plain-language questions with passages drawn from about ten million uncopyrighted judicial opinions. The system was not generative: it returned existing opinion text rather than creating new expression. To train it, ROSS engaged LegalEase Solutions, which prepared about 25,000 memoranda pairing a legal question with graded opinion passages. The drafters used thousands of Westlaw headnotes to frame the questions, because the headnotes offered "an easy way" to do it. ROSS marketed its platform as a substitute for Westlaw, at prices in line with Westlaw's.
Thomson Reuters sued in the District of Delaware in 2020. Judge Stephanos Bibas, a Third Circuit judge sitting by designation, granted partial summary judgment on February 11, 2025, holding that ROSS had infringed 2,243 headnotes and that the copying was not fair use. At the parties' request, he certified two questions for interlocutory appeal under 28 U.S.C. § 1292(b): whether the West headnotes and the West Key Number System are original as a matter of law, and whether ROSS's use of the headnotes was fair use. ROSS had shut down in 2021, citing the cost of the litigation, but pursued the appeal.
On September 29, 2026, a panel of Judges Restrepo, Montgomery-Reeves, and Bove affirmed. Judge Montgomery-Reeves wrote for the unanimous panel; the opinion was filed under seal and unsealed the next day. "Under ROSS's framing, this case appears to concern the future of AI legal technology," the court wrote. "In truth, this is no more than an ordinary copyright case."
The headnotes are original.
Copyright's originality bar is, in the Supreme Court's words in Feist, "extremely low": independent creation plus a modicum of creativity. The panel held that Westlaw's editors clear it. For each headnote, editors decide which points of law are important enough to include and how to word them so the headnote conveys the point accurately and stands alone. Those choices supply the "creative spark." The court read Callaghan v. Myers and Georgia v. Public.Resource.Org as supporting copyright in headnotes prepared by private reporters.
ROSS's counterarguments failed. Protecting headnotes does not give Thomson Reuters a monopoly over the law, because headnotes are not law and the underlying opinions remain free for anyone to use. The merger doctrine does not apply, because a point of law can be expressed in many ways. One part of the certified question went unanswered: ROSS never challenged the originality of the Key Number System in its briefing, so the court treated that issue as forfeited and did not address it.
Why the copying was not fair use.
Fair use under 17 U.S.C. § 107 turns on four factors; the site's copyright guide explains the framework. Only the second factor favored ROSS, and only slightly.
Purpose and character of the use. The use was "highly commercial" and "minimally transformative, at best." Both companies use headnotes for the same purpose: helping researchers find judicial opinions. ROSS's tool aimed to replace Westlaw, not to send users to it, which distinguished the search engine held transformative in Authors Guild v. Google. The intermediate-copying cases ROSS relied on, Google v. Oracle, Sega, and Connectix, turned on necessity: the copying there was needed to reach unprotected functional elements. ROSS already had the opinions and copied the headnotes because that was easier. "Unlike necessity, ease is not a justification for copying."
Nature of the work. The headnotes are published and more factual than creative, since they must accurately state the law. This factor favored fair use, slightly.
Amount used. ROSS copied the entire text of the headnotes it used, and each headnote is itself a complete copyrightable work. Because the opinions were freely available, the copying was not necessary, so ROSS took more than it needed. The district court had scored this factor for ROSS; the panel disagreed.
Market effect. The panel found two kinds of harm. First, value: following its Video Pipeline decision, the court looked to the effect on the work's value, not only on a narrowly defined market. Thomson Reuters promotes its headnotes as a draw for Westlaw subscriptions, and copying them to build a competing platform diminished that value; widespread copying to build substitutes would also harm Westlaw in the legal-research market itself. Second, the derivative market: "the market for licensing headnotes as text to train AI is rapidly developing," and Thomson Reuters already trains its own AI search products on them. "That Thomson Reuters did not license its headnotes to others does not disprove that a market exists to do so." By copying without authorization, ROSS "usurped Thomson Reuters's opportunity to enter that derivative market." ROSS's public-benefit arguments failed for want of evidence: the opinions are freely available to everyone, and ROSS priced its service in line with Westlaw.
What the decision does not decide.
The panel bounded its own ruling. ROSS's system could not generate original expression, and in a footnote the court set aside the Justice Department's statement of interest in the OpenAI multidistrict litigation and the district-court decisions in Bartz and Kadrey, which concern generative systems. Whether training a generative model on copyrighted books, articles, or music is fair use remains undecided at the appellate level. Commentators critical of the decision, including the Authors Alliance, argue that it gets the merger doctrine wrong and accepts a thinly evidenced training market; the Association of American Publishers welcomed it. Both sides caution against treating the case as a template for the generative-AI suits.
The decision is binding on the district courts within the Third Circuit, which includes Delaware. For Texas courts it is persuasive only: the Fifth Circuit has not addressed fair use in AI training.
What it means for Texas businesses.
- Training data is a licensable market. The ruling treats licensing content as AI training data as a derivative market the copyright owner may enter or reserve, even an owner that has not yet licensed anyone. Businesses that license content in either direction should expect training rights to be an express term.
- Substitutes face the hardest road. The facts that hurt ROSS were a direct competitor, the same purpose as the original, pricing set against the incumbent, and copying chosen for convenience. Uses further from substitution present a different case.
- Necessity is the argument that works. The intermediate-copying decisions the court credited involved copying that was necessary to reach unprotected material. A record showing why copying was needed is different from a record showing it was easier.
- Generative AI is still open. Businesses training or fine-tuning generative models should not treat this ruling as settling their position in either direction. It decides the retrieval-tool case in front of it.
What to watch.
Whether ROSS seeks Supreme Court review; it has said it intends to, according to reports, and no petition had been filed in the reporting checked for this article. The remainder of the Delaware case, including damages and the issues outside the 2,243 headnotes covered by the summary judgment. And the first appellate decisions in the generative-AI cases, which will test how far the Third Circuit's market analysis travels.
Sources.
- Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d Cir. Sept. 29, 2026) (precedential opinion; filed under seal Sept. 29, 2026 and unsealed Sept. 30, 2026). ca3.uscourts.gov (retrieved October 20, 2026).
- Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025) (Bibas, J.) (partial summary judgment for Thomson Reuters; two questions certified under 28 U.S.C. § 1292(b)). govinfo.gov (retrieved October 20, 2026).
- Reuters, "Unsealed opinion shows why Thomson Reuters won landmark AI fair-use ruling" (Sept. 30, 2026). reuters.com (retrieved October 20, 2026).
- Dennis Crouch, "Ease Is Not Necessity: Third Circuit Affirms No Fair Use in Thomson Reuters v. ROSS," Patently-O (Oct. 9, 2026). patentlyo.com (retrieved October 20, 2026).
- Authors Alliance, "Third Circuit Affirms in Thomson Reuters v. ROSS: The Court Gets Merger Wrong and Does a Disservice to Fair Use" (Oct. 2, 2026). authorsalliance.org (retrieved October 20, 2026).
- Association of American Publishers, statement welcoming the Third Circuit decision (Oct. 1, 2026). publishers.org (retrieved October 20, 2026).
This article is general information, not legal advice. Law checked through October 20, 2026. See the Disclaimer.