First US appeals ruling on AI training rejects fair use in Thomson Reuters v. Ross
The Third Circuit affirmed that Ross Intelligence infringed Westlaw headnotes to train its AI legal search tool. The opinion is sealed, so the only public reasoning is the 2025 ruling on a non-generative competitor.
A one-page judgment, a sealed opinion, 2,243 Westlaw headnotes. The court's reasons are not public yet.
The US Court of Appeals for the Third Circuit on September 29 affirmed that Ross Intelligence infringed Thomson Reuters' copyrights by training its AI legal search tool on material built from Westlaw headnotes, and that the copying was not fair use. The written opinion is sealed.
The judgment is public, the reasoning is not
What the public can read so far is a one-page judgment that says "AFFIRMED," according to Copyright Lately. LawNext reported that the court gave the parties 10 days to propose redactions, with reasons, before it decides what to unseal.
The appeal, docket 25-2153, was interlocutory. Judge Stephanos Bibas, a Third Circuit judge who heard the case in the District of Delaware by designation, certified two questions in April 2025: whether Westlaw's headnotes are original, and whether Ross's use was fair use. Law360 calls the decision the first US appellate ruling on whether AI training can be fair use. LawNext describes Ross as now shuttered.
What Bibas decided in February 2025
Everything in this section comes from Bibas's February 11, 2025 opinion, not from the sealed appellate one.
Ross asked to license Westlaw content, and Thomson Reuters refused because Ross was a competitor. Ross then bought roughly 25,000 "Bulk Memos" from LegalEase: legal questions that LegalEase's lawyers wrote with Westlaw headnotes as their guide. Bibas found actual copying of 2,243 headnotes.
On factor one, he held the use commercial and not transformative, because Ross used the headnotes to build a research tool that competed with Westlaw. He rejected Ross's intermediate-copying argument. The cases Ross relied on, including Google v. Oracle and Sega v. Accolade, were about computer code, where copying was necessary to reach unprotected ideas. Neither was true of headnotes.
On factor four, he named two markets: legal research platforms and "data to train legal AIs." It did not matter whether Thomson Reuters had trained its own tools on the headnotes. The effect on a potential training-data market was enough. Factors two and three went to Ross, partly because Ross never showed a user a headnote.
Why the ruling may stay narrow
Bibas fenced in his own opinion. Ross's tool was not generative: a user typed a legal question and got back existing judicial opinions. He wrote that "only non-generative AI is before me today."
The facts were also unusually direct. Both sides agreed Ross and Westlaw were competitors, and the copied material was editorial work built for the same research job. The generative cases, such as the Authors Guild suit against OpenAI and Microsoft, involve models that write new text from books that are not search products. Whether the Third Circuit kept Bibas's limits is not public.
What the unsealed opinion has to answer
- Is a potential market for AI training data enough, on its own, to tip factor four?
- Does intermediate copying for training get fair-use room only when the material is computer code?
- Are Westlaw headnotes original enough to protect, the first certified question?
- Does the panel say anything about generative models, or leave them out as Bibas did?
Copyright Lately lists a similar set of open issues, including whether potential licensing markets are enough for factor four.
Why a build studio cares
Ross's product never showed anyone a Westlaw headnote, and it still lost on its training data. The copying sat two steps upstream, inside memos a contractor wrote for Ross from a database Ross was refused a license to. That upstream step is where we look first. When we draw a data-flow map in an audit, every dataset that trains, fine-tunes, or evaluates a model gets a line: where it came from, who produced it, and under what terms. A vendor that bought training data should be able to say what the seller built it from. This is not legal advice, and the Ross facts are narrow. But three questions cost a build team nothing this week: which datasets trained this model, did a contractor build any of them from a licensed source, and does the product compete with that source.
Next step: read Bibas's 2025 opinion, then watch docket 25-2153 on CourtListener for the unsealed opinion. If your model was trained on data a contractor assembled and nobody has traced its source, write to us at hello@gattyworks.com.