A ruling unsealed on September 30 by the US Third Circuit Court of Appeals has drawn the first appellate line in the copyright fight over AI training data. In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc (No. 25-2153), the court held that defunct legal-research company Ross Intelligence did not make fair use of the "headnotes" (case-law summaries) from Thomson Reuters' Westlaw platform when it used them to train its AI-based legal search engine. It is the first time a US appeals court has ruled on the fair-use defense in an AI context, and Reuters called it a first of its kind.
What the ruling says
The timeline stretches back years: Thomson Reuters sued in 2020, accusing Ross of copying thousands of headnotes to train a competing AI search product; Ross shut down its platform in 2021, citing the cost of the litigation; a Delaware district court first ruled for Thomson Reuters; and the appeals court, after filing its opinion under seal on September 29, unsealed it on September 30 and affirmed. The three-judge panel unanimously found that the headnotes, as distillations of points of law, possess a "creative spark" and are protected by copyright; that Ross used them to build a direct competitor for a "highly similar purpose"; and that the copying was not transformative, so the fair-use defense failed. Writing for the panel, Judge Montgomery-Reeves put it this way: the case "appears to concern the future of AI legal technology," but "in truth, this is no more than an ordinary copyright case."
The boundaries the court drew itself
The weight of this ruling lies precisely in what the court said it does not cover. First, Ross built a retrieval search engine, not generative AI — the court noted in a footnote that concerns raised by the US Department of Justice in the copyright lawsuit against OpenAI "do not apply here." Second, the court resolved only the pattern of "copying copyrighted material to build a competing product"; it did not answer whether training generative AI qualifies as fair use. Third, this is a holding under US federal copyright law, binding on analogous conduct within US jurisdiction — not a global rule.
Who is affected
The first group is teams building retrieval or RAG-style AI products: if training data comes straight from a copyrighted database and the output substitutes for that database, the "transformative use" shield will hardly hold. The second group is AI companies already in litigation — authors, news outlets and music labels have filed dozens of training-data lawsuits, courts have split, and more cases are pending in 2026; this case hands rights holders an appellate-level win to cite, but generative-training cases must still be argued on their own facts. The third group is rights holders themselves: the enforcement path is clearer now, but this ruling is no blank check for "all AI training must be paid for."
What to do on compliance
Do not treat the words "AI training" as a talisman for fair use. The court looked at three things: what was copied, what it was used for, and whether it displaced the market. Before building retrieval or competing AI products, run copyright due diligence on the corpus and get licenses where you can; keep records of data provenance, licensing chains and output controls — they matter in discovery; and track case law in the jurisdictions where you operate. What the Third Circuit says today is not necessarily what the Ninth Circuit will say tomorrow.
Three misreadings to avoid
Misreading one: treating this case as "all AI training must be paid for." The court expressly excluded generative AI, so rights holders cannot simply wave it at every model trainer. Misreading two: assuming "we build generative AI, so this is not about us." The reasoning — copying copyrighted material to build a competing product is not transformative use — is equally a warning for retrieval-augmented and AI search products; the Ninth Circuit drew the same generative-versus-retrieval line in its recent DMCA ruling. Misreading three: underestimating the cost of the lawsuit itself. Ross shut down in 2021 because it could not afford the fight, long before any final judgment — the path by which copyright litigation kills a small company deserves more attention than the outcome itself.
