Heppe-Smith Publishing

← Back to blog

Thomson Reuters v. ROSS: what the first appeal ruling on AI training means for authors

James Heppe-Smith

A stack of leather-bound law reports with a paperback novel leaning against it, and a brass balance scale behind.

On 29 September a US appeal court ruled for the first time on whether training an AI system on copyrighted material can be fair use. In Thomson Reuters v. ROSS Intelligence, three judges of the Third Circuit said that on the facts in front of them it was not.

You’ll see it reported as a win for authors. It is a win for a copyright owner, but the case had nothing to do with books and the AI in it couldn’t write a sentence. Even so, one passage does bear on the cases authors are fighting.

What Thomson Reuters v. ROSS was about

Thomson Reuters owns Westlaw, the legal research service. Its editors write headnotes, short summaries of the points of law in each court opinion. ROSS, a start-up building a rival legal search tool, hired a contractor to write about 25,000 training memos, each pairing a legal question with passages from court opinions. The contractor’s staff built those questions from Westlaw headnotes.

The trial judge found that 2,243 headnotes had been copied and that the copying was not fair use. ROSS appealed. It had already closed its platform in 2021, Reuters reports, so the appeal was about liability and the rule itself.

Why ROSS lost

American courts weigh four factors to decide fair use. The panel’s summary: “the second factor weighs slightly in favor of fair use, but the first, third, and fourth factors weigh against it.”

  • Purpose. ROSS used the headnotes to build a paid product for the same customers. The court called that “minimally transformative, at best.”
  • Necessity. The court opinions themselves are free for anyone to use. ROSS copied the headnotes because it was quicker. “Unlike necessity, ease is not a justification for copying.”
  • The market. ROSS “copied the headnotes to create a competing legal-research platform in the same market as Westlaw.”

ROSS also argued that a ruling against it would hold back AI development and touch on national security. It had brought no evidence for either, the court said, and on security it added: “this does not give ROSS carte blanche to violate copyright law merely because it incorporates AI.” This was, in the judges’ words, “no more than an ordinary copyright case.”

What the ruling does not decide

The court was careful about the limits of its ruling, and they matter more to authors than the result.

It is not about generative AI. The opinion says: “ROSS’s AI was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions.” The tools in the authors’ cases against OpenAI, Anthropic and Meta do generate new text, and that is the argument those companies rely on to call training transformative.

The court set the authors’ cases to one side by name. In a footnote it noted the Justice Department’s filing in the OpenAI case, which we covered in our piece on the OpenAI authors case, and said: “The concerns raised in that separate case do not apply here.” It went on: “Unlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression.” The footnote does not say those earlier rulings were right or wrong.

It says nothing about books, pirated libraries or AI output crowding out human writing. None of those appears in the opinion.

It binds only part of the country. The Third Circuit covers Delaware, New Jersey, Pennsylvania and the US Virgin Islands. The OpenAI case is in New York and the Anthropic and Meta cases are in California. Judges there can read this ruling and be persuaded by it, but they are not bound to follow it.

The passage authors should read

One part of the ruling does travel. Fair use turns partly on harm to the market for the original work, including markets the owner could reasonably license. Thomson Reuters argued that ROSS had damaged its chance to license headnotes as AI training data. The court agreed:

Here, the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing.

It also said: “That Thomson Reuters did not license its headnotes to others does not disprove that a market exists to do so.”

So an appeal court has now treated licensing text for AI training as a real market that copying can harm, which is what authors’ and publishers’ groups have been arguing for. The finding rests on the evidence in this case, though, including the fact that Thomson Reuters trains its own AI search tools on the same headnotes, and the court found harm in the ordinary market as well, because ROSS was building a direct substitute for Westlaw.

AI companies often cite Authors Guild v. Google, the Google Books case, in which scanning books to make them searchable was held to be fair use. The panel drew the contrast itself. Google’s search “may have driven users toward a book the users would then purchase from the original author.” ROSS’s tool “does not lead a user to Westlaw; instead, as ROSS admits, it aims to replace Westlaw.”

How it is being read

The opinion was filed under seal on 29 September and made public the next day, so the first reports were written without the reasoning. Reuters called it “the first copyright dispute over AI training to be heard by a US appeals court,” and noted that the case “does not involve generative AI.”

Commentators differ on how far it reaches. The legal technology writer Bob Ambrogi put the narrow view before the text was out: “An affirmance that closely follows the facts could be consequential but fairly confined.” Publishers take the wider one. When the Association of American Publishers filed a brief in the case in November 2025, its chief executive Maria Pallante said: “AI is a market to which copyright owners are fundamentally entitled.” That was the publishers’ argument before the ruling. Since it, both the publishers and the Authors Guild have welcomed the decision, Publishers Weekly reported on 2 October. Pallante said the court “correctly noted that the licensing market for AI training not only exists but is ‘rapidly developing.’” Mary Rasenberger, chief executive of the Authors Guild, said the ruling “vindicates copyright law.”

ROSS does not intend to stop here. On 2 October its lawyer said: “We intend to seek review by the Supreme Court to obtain much-needed clarity on these issues.” That is a statement of intent, not a filing, and the Supreme Court chooses which cases it hears.

What it means for you

Nothing changes today. No money flows to authors from this case and no book was involved.

The arguments have shifted a little your way. Lawyers for authors now have an appeal ruling that rejects “AI is different” as a reason to set copyright aside, and that counts a licensing market for training text as something copying can damage. Lawyers for the AI companies have the court’s own statement that generative AI is another question.

The ruling to watch is still Judge Stein’s in the OpenAI authors case, where both sides have asked for summary judgment. That one is about books, and about AI that writes.

In short

Thomson Reuters v. ROSS is the first appeal ruling on AI training and fair use, and the copyright owner won. It concerned a legal search tool that could not generate text, built to replace the product it copied from, and the court said in terms that the authors’ cases raise different concerns. Read the headlines with that in mind, and keep watching New York.

Keep reading: Your First 100 Readers: What Actually Works for Indie Authors, Is Kindle Unlimited Worth It for Indie Authors in 2026?.

Sources

← More posts