Thomson Reuters v. Ross: what the AI fair use ruling decided
Thomson Reuters v. Ross is the first appeals ruling on AI training and fair use. What the court decided, what it left open for generative AI, and what's next.
Source-based. Written from the documents, reporting and reviews linked in the text. Nothing here was tested hands-on by The Ruling Desk. How we work
A federal appeals court has ruled for the first time on whether training an AI system on copyrighted material is fair use, and in Thomson Reuters v. Ross the answer was no. The Third Circuit held that ROSS Intelligence infringed Thomson Reuters' copyright when it used 2,243 Westlaw headnotes to train a rival legal search tool. The ruling is narrower than many headlines suggest: ROSS's AI didn't generate text, and it was built to replace Westlaw. Here is what the court decided, what it said it was not deciding, and where you'll see it cited next.
Key takeaways
- The court ruled: a three-judge Third Circuit panel affirmed, in an opinion filed September 29, 2026, that the headnotes are copyrightable and that ROSS's copying was not fair use.
- The opinion is public: it was sealed for redactions on September 29, and the full text was on the court's website by September 30.
- A competitor case, by the court's own framing: the judges called it "no more than an ordinary copyright case" because ROSS used the headnotes to build a substitute for Westlaw.
- Generative AI is set aside: a footnote says ROSS's AI "cannot generate original expression," unlike the models in Bartz v. Anthropic and the copyright cases against OpenAI.
- Licensing markets count: the court found a "rapidly developing" market for licensing headnotes as AI training data, a point plaintiffs in other cases are likely to press. That is our reading, not a holding about their cases.
What the Thomson Reuters v. Ross ruling decided
Thomson Reuters sued ROSS in 2020. ROSS, founded by three University of Toronto students after an IBM Watson competition, built a search engine that answered plain-language legal questions with passages from about 10 million public judicial opinions. To train it, ROSS hired a contractor, LegalEase, which wrote roughly 25,000 training memos. The writers used Westlaw headnotes, the short summaries of legal points that Thomson Reuters' editors write above each opinion, to frame the memos' questions, because they offered "an easy way" to do it, according to the record quoted in the Third Circuit's opinion.
In February 2025, Judge Stephanos Bibas, a Third Circuit judge sitting as a trial judge in Delaware, granted Thomson Reuters partial summary judgment on 2,243 headnotes and rejected ROSS's fair use defense. He then let ROSS appeal two questions early: whether the headnotes are original enough to be copyrighted, and whether ROSS's use was fair.
The appeals panel, Judges L. Felipe Restrepo, Tamika Montgomery-Reeves and Emil Bove, heard argument on June 11, 2026. Montgomery-Reeves wrote the opinion. It answers both questions for Thomson Reuters:
- Originality: each of the 2,243 headnotes has the "creative spark" copyright requires, because editors choose which points of law matter and how to word them so each stands on its own.
- Fair use: three of the four statutory factors weigh against ROSS. Only the nature of the work, factual summaries of public law, tilts slightly its way.
The court also noted evidence that ROSS "at times acted in bad faith," including attempts to reach Westlaw through a law-firm investor's credentials and a student account, in breach of Westlaw's terms.
Why the court rejected fair use
The core of the ruling is purpose. Fair use under the Supreme Court's Warhol decision asks whether a new use has a different purpose from the original, weighed against how commercial it is. The panel found that Westlaw uses headnotes to help researchers find relevant opinions, and ROSS used them to build a platform that does the same thing, priced "in line with" Westlaw. Training an AI model was an intermediate step, the court said, which offers "a slight degree of difference" and no more.
ROSS leaned on two lines of precedent. The court distinguished both:
| ROSS's argument | Why the panel rejected it |
|---|---|
| Google Books (2015): scanning whole books to make them searchable was fair | Google's snippets served a new purpose and could lead readers to buy the book. ROSS aimed to replace Westlaw, not send users to it. |
| Google v. Oracle, Sega, Connectix: copying code as an intermediate step was fair | Those copies were necessary to reach unprotected functional elements. ROSS could have used the free judicial opinions instead. "Unlike necessity, ease is not a justification for copying." |
On the third factor, the panel went further than Bibas. He had found the amount copied favored ROSS; the appeals court found it weighed against, because each headnote is its own work and ROSS took them whole. ROSS's point that it used only 0.08% of Westlaw's 28 million headnotes did not help.
On the fourth factor, market harm, the court found damage twice over: to Westlaw's place in the legal research market, and to a potential market for licensing headnotes as AI training data. Thomson Reuters never licensed its headnotes to others, but the court said that doesn't prove the market doesn't exist.
What the ruling did not decide
The opinion is explicit about its limits, and they matter more than the headline.
- Generative AI. Bibas wrote in 2025 that "only non-generative AI is before me today." The panel's footnote 7 goes further, naming the generative cases. It notes that the US Department of Justice argued in a September 1, 2026 filing in the consolidated OpenAI copyright cases in New York that training a large language model is transformative, relying on Bartz v. Anthropic. "The concerns raised in that separate case do not apply here," the court wrote, because ROSS's AI cannot generate original expression and was trained to be a commercial substitute.
- Verbatim headnotes. The court left open whether headnotes that copy an opinion's words are copyrightable, since the 2,243 at issue don't.
- The Key Number System. ROSS never argued that Westlaw's topic index lacks originality, so the court didn't rule on it.
- Damages and the rest of the case. This was an early appeal on two questions. Bibas's 2025 order left other headnotes and whether some copyrights expired for trial, so the case goes back to Delaware.
That narrowness is the theme of most expert reading. Writing in the ABA's Business Law Today just before the appeal was decided, lawyers Giovanna Fessenden-Fairbank and Ben Silvers argued the case turns on facts that many AI suits lack: a direct competitor, paywalled content and a tool that doesn't generate new output. They also argue that how training data was acquired may matter more than whether a model is generative.
Where the Ross precedent will be cited next
Aaron Moss of Copyright Lately calls it the first federal appellate ruling on fair use in AI training and expects its reasoning to shape dozens of generative AI cases pending around the country. Which ones, and how, is still an inference, so here is ours.
- The OpenAI cases in New York. The panel itself engaged with the Justice Department's position there. AI companies will point to footnote 7 to say Ross doesn't reach chatbots; plaintiffs will point to the market-harm analysis.
- Suits where the AI product competes with the source. The court's reasoning is strongest when the copier builds a substitute. Plaintiffs whose work an AI tool can stand in for, rather than transform, will quote it most.
- Arguments about training-data licensing. Plaintiffs often argue that unlicensed training harms a market for licensing their work to AI companies. In Bartz, Judge William Alsup rejected that idea for books, writing that such a market "is not one the Copyright Act entitles Authors to exploit." The Third Circuit is the first appeals court to accept a training-data licensing market, on its facts.
The two 2025 trial rulings that found training fair use on their records, Kadrey v. Meta and Bartz v. Anthropic, were in California, in the Ninth Circuit. A Third Circuit decision doesn't bind courts there or in New York. It can persuade, and as the only appellate word so far, it will be read closely. The fight over who may train on whose data is also playing out between AI companies, as in OpenAI's distillation claim against Moonshot.
Sealed or public: clearing up the confusion
Early reports, including IPWatchdog's and Blake Brittain's story carried by Insurance Journal, said the reasoning was sealed. That was true on September 29: the court issued its judgment and gave the parties 10 days to propose redactions. Copyright Lately later reported that neither side proposed any, and the full opinion was posted on September 30, as ChatGPT Is Eating the World noted. Courthouse News quoted from it the same day. So there is no unsealing date left to wait for.
What it means for you
If you build AI products, the court's message is about purpose, not technology. It said AI does not give anyone "carte blanche" to copy, and it rejected "ease" as a reason to use protected material when free sources exist. If your tool is designed to do the same job as the data you trained on, this opinion is the one opposing counsel will cite. This is general reporting, not legal advice.
If you use chatbots, nothing changes today. The generative AI cases are still at the trial level. Courts aren't the only ones scrutinizing the big labs, either, as the FTC's probe into OpenAI and Anthropic shows. More of our coverage is in the AI section.
Bottom line
Thomson Reuters v. Ross is a real precedent with a narrow base. The Third Circuit ruled that copying Westlaw's headnotes to train a competing, non-generative search tool was not fair use, and it went out of its way to say the generative AI fight is a different case. Watch for two things next: whether ROSS, which shut down its platform in 2021 according to Brittain's report, seeks further review, and how judges in the OpenAI cases treat footnote 7.
FAQ
Is AI training fair use after this ruling?
Not as a blanket rule either way. The Third Circuit held that ROSS's specific use, training a non-generative tool to compete with Westlaw, was not fair. It expressly distinguished generative AI models, and two 2025 trial rulings in California found training fair use on their records.
Does the ruling apply to ChatGPT, Claude or Llama?
Not directly. The court said the concerns in the consolidated OpenAI cases "do not apply here" because ROSS's AI could not generate original expression. Lawyers on both sides will still cite its reasoning on market harm and licensing.
What happens to the case now?
It returns to the federal court in Delaware, where questions left for trial include whether some copyrights expired and damages. ROSS could also ask the full Third Circuit or the Supreme Court to review the ruling; none of the coverage we read reports that it has, as of October 1, 2026.