Friday

World news with the sources attached

US appeals court rejects fair use defence for AI trained on a rival’s legal summaries

The Third Circuit sided with Thomson Reuters against the defunct start-up ROSS Intelligence, but said its ruling does not reach AI systems that generate new text.

Event date
Published
Reading time
3 min

Key points

  • The Third Circuit affirmed on 29 September 2026 that ROSS Intelligence infringed copyright in 2,243 Westlaw headnotes.
  • It is the first US federal appeals ruling on fair use and AI training.
  • The court recognised a developing market for licensing content as AI training data.
  • The opinion expressly distinguishes generative AI, leaving that question undecided.

For the first time, a US federal appeals court has ruled on whether copying copyrighted material to train an artificial intelligence system can be excused as “fair use”, and in the case before it the answer was no.

The Court of Appeals for the Third Circuit on 29 September 2026 upheld a lower-court judgment for the information company Thomson Reuters against ROSS Intelligence, a legal research start-up that has since ceased operating. The opinion was filed under seal and made public the following day.

The dispute

Thomson Reuters owns Westlaw, a legal research service that includes “headnotes”, short summaries of the points of law in court decisions. ROSS set out to build a competing search tool that used AI to return relevant passages from judicial opinions in response to a lawyer’s question.

To train it, ROSS commissioned an outside firm, LegalEase Solutions, to prepare roughly 25,000 training memos. Those memos were built on Westlaw headnotes. Thomson Reuters sued in 2020 in federal court in Delaware. In February 2025 Judge Stephanos Bibas ruled that 2,243 headnotes had been infringed and that ROSS could not rely on fair use, reversing a position he had taken in 2023, when he said the question should go to a jury. ROSS appealed before trial.

What the appeals court decided

The three-judge panel, Judges Restrepo, Montgomery-Reeves and Bove, affirmed on both of the questions before it. It held first that the headnotes were original enough to be protected by copyright.

On fair use, US law asks courts to weigh four factors. The panel found that three of them favoured Thomson Reuters:

  • Purpose of the use. ROSS’s use was commercial and only minimally transformative, because it served the same end as the original: helping lawyers do research. ROSS “aspired to be a direct competitor”, the court said.
  • Amount taken. Each headnote is a separate work, and ROSS copied each one whole.
  • Effect on the market. ROSS’s product competed directly with Westlaw. The court also recognised a second market: “the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing,” it wrote. That Thomson Reuters had not licensed its headnotes to others did not mean no such market existed.

Only the second factor, the nature of the work, leaned slightly towards ROSS, because headnotes are largely factual.

The court also dismissed the argument that copying as an intermediate step towards a non-infringing product should be excused. The underlying court opinions were freely available, so ROSS did not need the headnotes. “Unlike necessity, ease is not a justification for copying,” the opinion said.

What it did not decide

The panel described the dispute as “no more than an ordinary copyright case” and drew a line around it. “Unlike the AI models in Bartz and In re: OpenAI, ROSS’s AI platform cannot generate original expression,” it wrote, referring to pending cases about chatbots and other generative systems.

That limit matters. ROSS’s tool retrieved existing passages; it did not write anything. The far larger legal fight, over whether developers of generative models such as Anthropic, Meta and OpenAI may train them on books and other creative works without permission, remains open. Two federal trial judges have approached that question with different reasoning: in Bartz v. Anthropic, Judge Alsup likened training to human learning, while in Kadrey v. Meta Judge Chhabria rejected the comparison because such models can produce “countless competing works”, the law firm Ballard Spahr noted in an analysis of the ruling. No appeals court has yet settled the matter.

Rights holders nonetheless saw the decision as support for their position. The US recording and music publishing trade bodies, the RIAA and NMPA, had filed a brief backing Thomson Reuters, and the court’s recognition of a licensing market for training data is the argument their members are making in their own lawsuits against AI music generators, the trade publication Music Business Worldwide reported.

Next steps

A Thomson Reuters spokesperson said the company was pleased with the outcome, the legal news site LawSites reported. ROSS’s lawyer said the company would ask the Supreme Court to review the decision, according to the same reports. Because the appeal was taken before trial, the case otherwise returns to the district court in Delaware.

Sources

  1. Thomson Reuters v. ROSS Intelligence, No. 25-2153, opinion (3d Cir., public version filed 30 September 2026) US Court of Appeals for the Third Circuit (copy hosted by Plagiarism Today), 30 Sep 2026 · primary source
  2. Third Circuit Addresses Fair Use in AI Training, but Leaves Generative AI Questions Unresolved Ballard Spahr, 2 Oct 2026 · independent report
  3. A US appeals court has rejected a 'fair use' defense over AI training. What does it mean for the labels' fight with Suno? Music Business Worldwide, 5 Oct 2026 · independent report
  4. 3rd Circuit Rules for Thomson Reuters in Its Copyright Fight Against Legal Research Startup ROSS LawSites (LawNext), 30 Sep 2026 · independent report
  5. Third Circuit Affirms Revised Fair Use Ruling Against ROSS' AI Legal Research Platform in Sealed Opinion IPWatchdog, 30 Sep 2026 · independent report

Spotted an error? See how to request a correction.

From other sections