Third Circuit affirms ROSS Intelligence infringed Westlaw copyrights — first appellate ruling that AI training can fail fair use

What happened

On September 29, 2026, the U.S. Court of Appeals for the Third Circuit affirmed that ROSS Intelligence infringed Thomson Reuters' copyrights in 2,243 Westlaw headnotes by using them to train its AI-powered legal research tool (opinion unsealed September 30, 2026). The court upheld the Delaware district court's revised summary judgment (Feb 2025), finding ROSS's fair-use defense failed three of the four statutory factors: the use was not transformative because ROSS's system performed the same commercial function as Westlaw (helping legal researchers find authority); each headnote counted as its own complete work, so copying any one weighed against fair use on the amount factor; and ROSS directly competed with Westlaw while using content from a market — licensing copyrighted material as AI training data — the court found cognizable even in its nascent state.

The court drew an explicit boundary: the ruling applies to ROSS's non-generative, retrieve-only system, and the panel declined to hold that AI training is categorically infringing or that fair use can never apply to machine-learning use cases. The opinion notes the DOJ's September 1, 2026 statement of interest in the OpenAI MDL cases, which calls LLM training "extraordinarily transformative," and treats that reasoning as not applying to ROSS's architecture. ROSS, a San Francisco startup that raised about $13 million, wound down in January 2021 under the weight of this litigation; Thomson Reuters welcomed the ruling.

Why it matters to creators

This is the first federal appellate decision to answer whether AI training can constitute fair use — and on its facts the answer is no. Two holdings reach beyond legal databases: (1) copying a competitor's human-curated editorial layer to build a tool that serves the same commercial function is not transformative; (2) once a rightsholder is licensing its content for AI training, courts can treat that developing licensing market as cognizable harm under fair-use factor four. That second point applies directly to content types where AI-training licensing markets now exist — news articles, academic journals, and music — and strengthens the bargaining position of creators whose work sits behind catalogs already being licensed to AI companies.

What is not yet known

The ruling does not decide generative AI training: courts have separately found fair use for Claude (Bartz v. Anthropic, June 2025) and Llama (Kadrey v. Meta) training on lawfully obtained books, and the Third Circuit distinguished those cases on the facts. No court is bound by this decision outside the Third Circuit, and a circuit split on AI-training fair use remains possible; the Supreme Court has not spoken.