On September 29, 2026, the Third Circuit affirmed Judge Stephanos Bibas's judgment for Thomson Reuters in its case against ROSS Intelligence. According to the Chat GPT Is Eating the World tracker, the panel rejected ROSS's fair use defense for its use of Westlaw headnotes in memos that trained its AI model. Judge Montgomery-Reeves is listed as the opinion's author. It is the first US appeals-court ruling on fair use in AI training.
One caveat matters for everything below. The full opinion is under a ten-day sealing order, and only the bare judgment was public when this was written. What follows rests on the district-court record, the parties' briefing, and the reported outcome. It does not rest on the panel's reasoning, which we have not read.
The strongest case for the ruling
The case for Thomson Reuters is serious. Westlaw's headnotes are editorial work: short summaries of points of law, each tied to a key number. Judge Bibas's 2023 memorandum opinion records that Westlaw holds a registered copyright on its text and compilation of legal material, including headnotes and the Key Number System. It also records that ROSS first tried to license Westlaw, was refused because Thomson Reuters does not let users build competing platforms, and then commissioned about 25,000 question-and-answer memos through a third party, LegalEase. Thomson Reuters said those questions were essentially headnotes with question marks at the end. If a startup can't license a product and then rebuilds it from the product's own editorial layer, that is a fair-sounding account of free-riding. Copyright exists to protect the incentive to produce such work.
What the case is, and isn't
The facts here are narrow. ROSS was a legal-research startup building a "natural language search engine" that returned quotations from judicial opinions in response to lawyers' questions. Thomson Reuters sued in 2020, and the AP's February 2025 report noted that ROSS is now defunct. The tool was a competitor, built to answer the same questions Westlaw answers.
That is why the reported framing, that the tool was a market substitute, is the part that matters. The fourth fair use factor, harm to the market for the original, is where ROSS fought hardest. In its May 2026 supplemental briefing, ROSS argued the court should look at harm to the headnotes themselves rather than the whole legal-research market, and said West showed no lost subscribers. Thomson Reuters answered that ROSS copied to build a direct commercial competitor. It also cited Bartz v. Anthropic, where, per the same account, a California court said using a proprietary research system to build a competing AI tool was not transformative.
The district court's ruling was also confined to non-generative AI. The case concerned a search tool that surfaced existing judicial text, not a model that produces new text, images or code. The Copyright Office's Part 3 report on generative AI training, released in pre-publication form on May 9, 2025, addresses the generative question separately. Whether the Third Circuit's reasoning reaches that question depends on language we cannot yet see.
Where proportionate policy should land
Our view is that the best reading of this outcome is also the least alarming. Courts that hold a direct competitor liable for building a substitute from a rival's protected editorial product are applying fair use as it has long worked. Substitution, not the act of learning from data, is the harm copyright law is built to address. A startup that copies a rival's curated product to undercut it is a different case from a lab that trains on a broad corpus to build something with different outputs and a different market.
The risk is overreading. If lower courts take "AI training is not fair use" as the holding, developers face uncertainty that falls hardest on small firms. Large incumbents can pay for licenses or settle. A startup, like ROSS, often cannot. A ruling that turns on market substitution is workable. A ruling read as a general rule against training would raise entry costs without protecting any author's actual market.
Three things should happen next:
- Read the opinion before drawing conclusions. The sealing order lifts within days. Commentators on both sides will claim it, as they did the Third Circuit's earlier UpCodes decision.
- Keep the factor-four inquiry evidence-based. Courts should ask who actually lost revenue and whether the output substitutes for the original. Harm should not be inferred from the fact of copying alone.
- Let licensing markets develop without mandates. Voluntary licensing for training data is growing. Statutory compulsory licenses would freeze terms before anyone knows what the market is worth.
What to watch
The ruling binds only the Third Circuit, and ROSS could seek rehearing or Supreme Court review. Generative-AI cases pending in other circuits will test whether this reasoning carries over to models whose outputs do not directly compete with the training data. Until then, the safe reading is the narrow one: a court has held that copying a competitor's editorial product to build a rival is not fair use. It has not held that all AI training needs a license.