Third Circuit: ROSS's use was not fair use
On 29 September 2026 the US Court of Appeals for the Third Circuit, in Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. ROSS Intelligence Inc. (No. 25-2153), affirmed partial summary judgment for Thomson Reuters: 2,243 Westlaw headnotes are original enough for copyright, and ROSS's copying of them into training memos for a legal-research program was not fair use.
Why it matters
The court draws the limit of its own ruling: ROSS's program was not generative, it only returned passages of existing judicial opinions, and the court contrasts it with the models in Bartz and the OpenAI litigation, which it cites in a footnote. So this is a ruling on training a search program on a direct competitor's material, not on training generative models.
Who and what. The panel was Judges Restrepo, Montgomery-Reeves (who wrote the opinion) and Bove; District Judge Stephanos Bibas (Delaware) sat on the appeal by designation. It is an interlocutory appeal under 28 U.S.C. § 1292(b) on two questions the district court certified itself: whether the Westlaw headnotes and the Key Number System are original, and whether the use was fair. The court did not address the Key Number System because ROSS never challenged its originality. The dispute concerns 2,243 headnotes of a batch of 2,830; LegalEase, hired by ROSS, wrote about 25,000 training memos whose questions drew on headnotes. What was held. Originality: Feist's low bar ('creative spark'), the editors choose what to include and how to word it; the court rejected ROSS's arguments about a monopoly over the law, merger of idea and expression, and the Second Circuit decision. Fair use, factor by factor: the first against (commercial use, 'minimally transformative, at best', since the purpose is the same, helping find case law; the court distinguishes Authors Guild, Oracle, Sega and Sony; ROSS's bad faith noted); the second slightly in favour of fair use (the headnotes are more factual); the third against (the headnotes were copied whole and copying them was not necessary because the opinions are free); the fourth against (harm to the legal-research platform market and to a market for licensing headnotes as AI training data that the court calls rapidly developing; public benefit not shown). The result: ROSS's use was not fair; the district court's ruling is affirmed. What the record does not claim. The court did not decide whether training generative models is fair use and expressly set itself apart from that. It did not answer whether headnotes that quote opinions verbatim are original (footnote 5). This is an interlocutory appeal; the opinion does not say how the rest of the case will go. Press reports of a possible further appeal were not used.