# Third Circuit Rejects Fair Use in First Appeal Over AI Training Page: https://hirantha.me/blog/third-circuit-rejects-fair-use-in-first-appeal-over-ai-training Published: 2026-09-30 Written by: hirantha.me Newsroom (AI-generated). An AI model wrote this from the sources listed below; how it is made: https://hirantha.me/about/newsroom Topics: ai copyright, fair use, thomson reuters, ross intelligence, ai training data > The Third Circuit upheld Thomson Reuters' win over ROSS Intelligence on 29 September, but its opinion is sealed and the facts make it a narrow precedent. An appeals court has sided with a copyright holder over an AI developer for the first time. The catch is that nobody outside the courtroom can yet read why. ## What happened On 29 September 2026 the US Court of Appeals for the Third Circuit affirmed a lower court's ruling in Thomson Reuters v. ROSS Intelligence. The public record so far is a one-page judgment reading "AFFIRMED". The written opinion, authored by Judge Tamika Montgomery-Reeves, was filed under seal while the parties propose redactions. Chat GPT Is Eating the World, a blog that tracks AI litigation, reported the seal would last about ten days. Law360 and MLex both describe it as the first US appellate ruling on whether training an AI system on copyrighted material can be fair use. ## The case behind it ROSS built a legal research tool. Thomson Reuters, which owns Westlaw, sued because ROSS used material derived from Westlaw's headnotes, the short summaries of legal points attached to court opinions, to build its product. In February 2025 Judge Stephanos Bibas, sitting in the Delaware federal court, ruled for Thomson Reuters on summary judgment. He found that ROSS infringed 2,243 headnotes and that its use was not transformative because it was building a competitor to Westlaw. ROSS, now defunct, appealed before a full trial. ## Why this is narrower than it sounds Headlines will say AI training is not fair use. That overstates it, and a September 2026 analysis in the American Bar Association's Business Law Today spells out why. Three facts drove the original ruling: - ROSS sold a direct competitor to the product it copied from. - Westlaw's headnotes sit behind a paywall for subscribers. - ROSS was not a generative system. It used the headnotes to power a search tool rather than to produce new text. Chatbots and image generators differ on each point. The same ABA piece notes that in Bartz v. Anthropic, a federal court in Northern California treated training on lawfully obtained books as fair use, while rejecting training on pirated copies. Together, the two rulings suggest that how data was obtained and whether the product competes with the source both carry weight. ## What the sealed opinion could settle Until the opinion is public, several questions stay open, according to Copyright Lately: - Whether the panel accepted Judge Bibas's view of how original a headnote is. - Whether the mere existence of a potential licensing market is enough to weigh against fair use. - Whether the doctrine of intermediate copying, developed mostly for computer code, applies to this kind of material. Copyright Lately lists Judge Emil Bove on the panel alongside Judge Montgomery-Reeves. The affirmance carries no reasoning today, so its value to other courts depends entirely on what the opinion says once unsealed. ## What the judgment does not tell us A one-line affirmance says the result stands. It does not reveal whether the panel agreed with every step of Judge Bibas's reasoning, or whether it rested on narrower grounds. That is why lawyers are waiting for the unsealed opinion before drawing conclusions, and why early commentary is mostly about scope rather than substance. Copyright Lately names Judge Bove on the panel besides the author. ## Why other cases will notice Dozens of generative AI copyright suits are pending across the US. An appellate decision, even a narrow one, gives trial judges something they had lacked: a higher court's view on the fair use factors applied to AI training. Trial courts elsewhere are not bound by the Third Circuit, but they read it. ## What this means in practice - Teams building AI features on third-party text should note the factors that counted against ROSS: a competing product, paywalled inputs and no licence. - Data provenance now looks like a legal question as well as an engineering one. Keep records of where training and retrieval data came from. - Do not treat this as settling generative AI. Wait for the opinion, then read what it says about market harm and licensing. - Publishers and data owners may find the licensing point the most useful part, and it is the one still unknown. ## Questions **Did the Third Circuit rule that AI training is not fair use?** It affirmed that ROSS could not claim fair use for its non-generative legal search tool. The reasoning is sealed, and analysts note the facts were narrow. **When will the Third Circuit opinion be public?** It was filed under seal pending redaction proposals; Chat GPT Is Eating the World reported about ten days. **Does this affect generative AI lawsuits?** It gives trial courts an appellate view to consider, but ROSS was not generative AI and other courts are not bound by it. ## Sources - [3rd Circ. Affirms Thomson Reuters' Westlaw AI Copyright Win](https://www.law360.com/pulse/courts/articles/2531563), Law360 - [Third Circuit affirms summary judgment rejection of fair use defense by ROSS Intelligence. Opinion under seal for now.](https://chatgptiseatingtheworld.com/2026/09/29/third-circuit-affirms-summary-judgment-rejection-of-fair-use-defense-by-ross-intelligence-opinion-under-seal-for-now/), Chat GPT Is Eating the World - [Signed, Sealed, Affirmed: Third Circuit Backs Thomson Reuters over ROSS](https://copyrightlately.com/thomson-reuters-v-ross-third-circuit/), Copyright Lately - [Thomson Reuters v. Ross, One Year Later: A Narrower Precedent Than the Headlines Suggested](https://businesslawtoday.org/2026/09/thomson-reuters-v-ross-one-year-later-a-narrower-precedent-than-the-headlines-suggested/), Business Law Today (ABA)