A federal appeals court has held that training an AI legal-research tool on Westlaw editorial material is not fair use. In Thomson Reuters Enterprise Centre GmbH and West Publishing Corp v. ROSS Intelligence Inc., No. 25-2153, decided on 29 September 2026, the US Court of Appeals for the Third Circuit affirmed partial summary judgment for Thomson Reuters on both questions certified for interlocutory review: that 2,243 Westlaw headnotes are original enough to be copyrightable, and that ROSS’s use of them was not fair. Judge Montgomery-Reeves wrote for a panel with Judges Restrepo and Bove, after argument on 11 June 2026. The appeal came from the District of Delaware, No. 1:20-cv-00613, where Judge Stephanos Bibas was sitting by designation.
The facts were narrow and the court kept them that way. Denied a Westlaw licence, ROSS commissioned a third party to build roughly 25,000 legal memos used to train its search tool, and the district court found that for 2,243 of them the questions tracked the headnote text rather than the judicial opinion underneath. On fair use the panel found the second factor favoured ROSS and the other three did not. The use was commercial and “minimally transformative, at best”, because both companies used the same material for the same end — helping a researcher find responsive law. The court distinguished Authors Guild v. Google, and distinguished the intermediate-copying line of Sega, Connectix and Oracle on a single point: ROSS already had the judicial opinions and could have written its own memos from them. It took the headnotes because that was easier, and, in the panel’s words, “ease is not a justification for copying”. On amount, ROSS’s argument that it took only 0.08% of Thomson Reuters’s 28 million headnotes failed because each headnote is itself a copyrightable work — so ROSS took all of 25,000 works, not a sliver of one. On market effect, the court accepted both that ROSS damaged the value of headnotes as a draw for Westlaw subscriptions and that it usurped a derivative market for licensing headnotes as AI training data.
What this changes
The most consequential holding is the one about a market that the plaintiff had never actually entered. Thomson Reuters has not licensed its headnotes to anyone for AI training. The panel held that did not matter: the evidence showed the market for licensing text as training data is “rapidly developing”, Thomson Reuters was already using the headnotes to train its own AI search products, and under Castle Rock a potential derivative market is not illusory merely because the owner has shown little interest in exploiting it. That reasoning converts the AI-licensing market itself into the harm. Any publisher, database owner or annotation-heavy business can now plead factor four without ever having sold a licence — provided it can put evidence of the market in the record. That evidentiary burden is the practical takeaway: the win here rested on proof that comparable licensing is happening and that the owner uses the material the same way internally. Content owners who expect to make this argument should be gathering that proof now, not at summary judgment.
Equally useful is the court’s treatment of the original market. Following Video Pipeline, it held there need be no standalone market for the copied material at all; it was enough that headnotes are a feature Thomson Reuters uses to sell subscriptions, so diminishing their pull diminishes the value of the work. For subscription publishers, research databases and anyone whose protected material is a hook rather than a product, that removes the most common defence to damages — “nobody buys this separately, so nothing was lost.”
Defendants should read footnote 7 closely, because the panel fenced its own holding. It noted that the Department of Justice had filed a statement of interest on 1 September 2026 in the consolidated OpenAI copyright litigation in the Southern District of New York, No. 1:25-md-3143, arguing on the strength of Bartz v. Anthropic that training a model able to generate original responses is transformative — and the panel said those concerns did not apply, because ROSS’s platform cannot generate original expression and was trained for the purpose of substituting for Westlaw. It also pointed out, pointedly, that the DOJ knew how to assert its interests and had not done so in this case. So the decision does not resolve generative training. What it does is make the generative/retrieval distinction load-bearing: if a tool ranks, retrieves or classifies rather than produces new expression, the strongest transformativeness arguments in the current case law are unavailable to it.
Three operational consequences follow for anyone assembling training data. First, the denominator changes. Dataset risk is routinely measured as a share of a corpus; this panel measured it per work, and 0.08% of a database became 100% of 25,000 separate works. Audit datasets by the number of discrete protected works ingested, not by volume share. Second, necessity now has to be documented. The intermediate-copying cases survive, but only where copying is the sole route to unprotected material — so the record should show why a protected source was required rather than merely convenient, and where a lawful alternative existed, why it was unusable. Third, the question of what the trained system does has moved upstream into the acquisition decision, because the output’s character is what decides the first factor.
On the litigation map, the ruling binds Delaware, New Jersey, Pennsylvania and the Virgin Islands. Delaware is where a large share of technology companies are incorporated and sued, which gives this a reach beyond its circuit. The posture is still interlocutory — partial summary judgment affirmed, with the case returning to the district court — and ROSS, through White & Case partner Yar Chaikovsky, has said it will petition the Supreme Court, telling LawNext the decision leaves continued uncertainty over copyright and AI model training. Treat the holding as binding law in the Third Circuit and persuasive elsewhere, not as the last word.
If you are the owner of the content, the work that matters now is evidentiary: establishing that a licensing market for your category of material exists and what it pays, which is a market research exercise before it is a litigation one. If you are the one acquiring the data, the exposure sits in sourcing terms and provenance records, where contract management decides whether you can show a licence, a lawful alternative, or neither.
Reported by PerspireIP from the Third Circuit’s opinion in No. 25-2153, with commentary from The IPKat and LawNext.