On June 11, 2026, a three-judge panel of the U.S. Court of Appeals for the Third Circuit — Judges L. Felipe Restrepo, Tamika R. Montgomery-Reeves, and Emil J. Bove — heard oral argument in Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153, the first case to put a federal appeals court squarely in front of the question that has hung over the AI industry since ChatGPT launched: is training a model on copyrighted material fair use?
The stakes are larger than one legal-research startup. District courts have split. In February 2025, Judge Stephanos Bibas of the U.S. District Court for the District of Delaware granted Thomson Reuters partial summary judgment, ruling that Ross's use of 2,243 Westlaw headnotes to train its legal-search AI was not fair use on any of the four statutory factors. Four months later, two Northern District of California judges reached the opposite conclusion for generative AI: Judge William Alsup called Anthropic's book-training "exceedingly transformative" in Bartz v. Anthropic (June 23, 2025), and Judge Vince Chhabria found Meta's Llama training protected in Kadrey v. Meta (June 25, 2025). The Third Circuit is now positioned to either harmonize or deepen that split — and its answer will shape how every AI company litigating training-data claims frames its defense.
What Ross actually did
Ross Intelligence built a natural-language legal search tool designed to compete directly with Westlaw. Before training its model, Ross asked Thomson Reuters to license Westlaw's data; Thomson Reuters refused, reportedly because Ross was a competitor. Ross then paid a third-party vendor, LegalEase Solutions, to produce "Bulk Memos" — legal questions and answers derived from Westlaw headnotes — which Ross used as training data. Judge Bibas found this was not transformative: the purpose was commercial, the output competed head-to-head with Westlaw, and it harmed an actual and potential market for Thomson Reuters's headnotes.
At oral argument, ROSS's counsel, Mark Davies of White & Case, argued the company copied only about 0.08% of Westlaw's roughly 28 million headnotes and transformed them into training pairs that teach a machine "how to think like a lawyer," not a substitute product. Thomson Reuters's counsel, Dale Cendali of Kirkland & Ellis, called it "a classic case of substitution." Judge Bove pressed Davies on what made Ross's tool "so transformative from what happens when I log into Westlaw" — framing the core tension as whether Ross built a materially different research method or merely a faster "cheat sheet" to the same underlying law.
Steelmanning the case against Ross
Thomson Reuters has the stronger version of the anti-fair-use argument that AI critics have been making since 2023, and it deserves a fair hearing. Ross didn't scrape the open judicial opinions underlying the headnotes — those are public domain — it paid for a workaround to reproduce Thomson Reuters's editorial labor: the synthesis, categorization, and cross-referencing that Westlaw's Key Number System represents. Ross asked to license that labor, was refused, and took it anyway through an intermediary instructed to avoid verbatim copying while preserving the substance. And the resulting product was not a chatbot with diffuse, unpredictable outputs — it was a search tool built to displace Westlaw subscriptions one law firm at a time. If fair use exists to protect transformative uses that don't cannibalize the market for the original, this is close to the paradigm case for denying it: a direct commercial substitute, built from a refused license, sold to the same customers.
Why the line still matters for innovation
But the Third Circuit shouldn't let the facts of one bad-faith licensing dispute set the interpretive rule for all AI training. As the Electronic Frontier Foundation argued in its amicus brief, joined by the American Library Association, the Internet Archive, and Public Knowledge, the headnotes at issue are themselves "entirely factual statements about what the law is," and treating a tool that extracts and reorganizes legal facts as inherently infringing risks sweeping in search engines, citation checkers, and research assistants that never reproduce anyone's expressive prose. Twelve amici — including Disney, Paramount, and Ross's own rival LexisNexis — back Thomson Reuters, while a similarly sized coalition of libraries, law professors, and open-access advocates backs Ross, which tells you the doctrine genuinely is unsettled, not that either side is obviously right.
The better outcome preserves Judge Bibas's market-substitution finding on these specific facts — Ross's refused-license, direct-competitor posture is a poor test case for AI training generally — while the Third Circuit should resist writing an opinion that treats any commercial AI training as per se unfair. The workable distinction, consistent with Bartz and Kadrey, is between models that generate new expression from ingested material and tools engineered to reproduce a competitor's specific market function using a paid-for substitute for a refused license. Congress has not updated the Copyright Act for AI training, and it shouldn't have to: fair use's four-factor test was built for exactly this kind of case-by-case line-drawing. The Third Circuit's ruling, expected by late 2026, will be the first appellate word on where that line sits — and every AI company now defending a training-data suit is watching to see whether it gets read narrowly or broadly.