The US Court of Appeals for the Third Circuit ruled on 29 September that ROSS Intelligence could not claim fair use for building a legal research tool out of material derived from Thomson Reuters’ Westlaw headnotes. It is the first time a US appellate court has decided whether training an AI system on copyrighted material is fair use, and the answer went against the company that did the training.
The panel affirmed the February 2025 summary judgment of Judge Stephanos Bibas, sitting in Delaware, who found that ROSS infringed 2,243 headnotes — the editorial summaries Thomson Reuters writes to mark the points of law in a judicial opinion. ROSS obtained material derived from those headnotes through a Westlaw licensee and used it to train a product sold to the same customers. The company shut its platform down in 2021, citing the cost of the litigation.
The opinion arrived a day after the judgment
On 29 September the court released only a one-page order reading “AFFIRMED”, with the reasoning under seal pending proposed redactions. The opinion, written by Judge Tamika Montgomery-Reeves, became available on 30 September.
Its core is the first fair use factor. The panel found that ROSS’s system did not produce new expression: it “quoted from judicial opinions relevant to a query” rather than writing anything of its own. On that reading the copying served the same purpose as the original and competed with it directly, which is close to the opposite of a transformative use.
The court drew a line around generative systems
The most consequential sentence for everyone else may be a footnote. At page 17, footnote 7, the panel expressly distinguished this case from the generative AI litigation, naming Bartz and Kadrey — the two district court decisions that have so far framed the American argument about training large language models on books.
That is a deliberate narrowing. ROSS built a retrieval tool that surfaced existing judicial text; a model that writes new prose from what it has read is a different object, and the Third Circuit declined to decide it here. Copyright holders will nonetheless cite the ruling for its structure: a commercial competitor copying editorial work to compete in the same market loses on factors one and four.
What it changes now
Dozens of AI copyright suits are pending in US district courts, and until this week none of them had appellate authority to work from in either direction. They do now, and it favours rightsholders on the facts it covers. Defendants will argue that the facts it covers are unusually narrow — a headnote is editorial writing, ROSS sold into Westlaw’s own market, and nothing the system produced was new.
What the opinion does not do is settle whether ingesting copyrighted text to train a generative model is fair use. It says, in a footnote, that this was not that question.
What to watch
Whether ROSS seeks review by the full Third Circuit or the Supreme Court, and how quickly plaintiffs in the generative cases start citing footnote 7 — either to contain it or to argue it proves the court’s reasoning reaches further than the panel admitted.
