Appeals court rules training AI on legal summaries was not lawful

The US Court of Appeals for the Third Circuit has upheld an earlier ruling that it was not fair use when Ross Intelligence used Thomson Reuters' Westlaw headnotes to train an AI system. Headnotes are short summaries of legal points in court decisions. Fair use is a US doctrine that, under certain conditions, allows the use of protected material without permission.

Thomson Reuters filed the lawsuit in 2020. Ross is a legal research company that has since shut down. The case centered on Ross using material from Westlaw to train an AI without permission.

Why did Ross lose?

According to Plagiarism Today's review, the court weighed four factors that go into a fair use analysis. Three of them favored Thomson Reuters. The use was only minimally transformative, meaning it did not create anything substantially new, and it was commercial. Ross copied entire headnotes and built a competitor to Westlaw.

One factor weighed slightly in Ross's favor, because headnotes are published and have factual content. The court also acknowledged that a market is emerging for licensing content used for AI training.

According to AI Weekly, it is said to be the first federal appeals ruling to directly address fair use in AI training. It sets precedent in the Third Circuit.

Not a ruling on ChatGPT and Claude

Ross's AI was not generative, meaning it was not a model that writes text or creates images on its own. According to the law firm BakerHostetler, it answered legal questions by returning existing court decisions, and users were never shown the headnotes.

The court itself distinguished between Ross and cases involving generative AI, because Ross aimed to replace a competitor rather than create original expression. The question of whether it is fair use to train generative models from companies such as OpenAI or Anthropic is therefore still unresolved. It is pending in, among others, the New York Times case.

BakerHostetler assesses that the ruling applies a strict necessity standard: The fact that copying is easy does not make it justified. According to the law firm, this could limit the development of fine-tuned, specialized models.

The date of the ruling varies across sources between September 29, September 30 and October 1, 2026.