A courtroom in the James A. Byrne US Courthouse in Philadelphia, where the 3rd Circuit sits (2007). Image: Carol M. Highsmith / Wikimedia Commons, Public domain, cropped
For the first time, a US appeals court has ruled on whether training AI on someone else’s copyrighted work counts as fair use, and the answer was no. The 3rd US Circuit Court of Appeals on Tuesday upheld Thomson Reuters’ win against Ross Intelligence, a former legal research rival accused of using material from Westlaw to train an AI search engine, Reuters reported.
Update: what the unsealed opinion says
The court’s full opinion, written by Circuit Judge Arianne Montgomery-Reeves for a unanimous three-judge panel, is now public, and it goes out of its way to keep the ruling narrow. ROSS had framed the case as a fight over the future of AI legal technology. The judges disagreed:
Under ROSS’s framing, this case appears to concern the future of AI legal technology. But appearances can be deceiving. In truth, this is no more than an ordinary copyright case.
Judge Arianne Montgomery-Reeves, US Court of Appeals for the Third Circuit
The key points from the opinion:
- Headnotes are protected. The court found the 2,243 Westlaw headnotes at issue “possess a creative spark” and are original enough for copyright, even though editors often follow the court’s own wording.
- Training was only a small change. Training an AI on the headnotes was “an intermediate step” that “arguably presents a slight degree of difference in use”, but ROSS used them to build a legal research tool, the same purpose as Westlaw. That made the use “minimally transformative, at best”.
- It isn’t about chatbots. In a footnote, the judges noted that ROSS’s AI “cannot generate original expression”, unlike the models in copyright cases against Anthropic and OpenAI. The Justice Department has argued in the OpenAI case that training a large language model is transformative, but the court said those concerns “do not apply here”.
- AI training data is a market. The court said the market for licensing headnotes as AI training data “is rapidly developing”, and that ROSS “usurped Thomson Reuters’s opportunity” to sell into it. Rights holders will lean on that line in other cases.
- “AI” is no excuse. The judges rejected ROSS’s argument that AI matters for national security, saying it “does not give ROSS carte blanche to violate copyright law merely because it incorporates AI.”
Of the four fair use factors, only the nature of the work leaned slightly towards ROSS. The purpose of the use, the amount taken and the harm to the market all counted against it.
What the case was about
Thomson Reuters sued Ross in 2020, accusing it of copying thousands of Westlaw “headnotes”, the short summaries of legal points that sit at the top of court opinions. Ross used them to train its own AI tool, which found relevant court decisions in answer to users’ questions. Ross shut its platform down in 2021, blaming the cost of the lawsuit.
Ross argued that training its AI on the headnotes was fair use, the part of US copyright law that allows some copying without permission, for example when the new use transforms the original. A federal judge in Delaware rejected that argument last year:
Ross took the headnotes to make it easier to develop a competing legal research tool. So Ross’s use is not transformative.
US Circuit Judge Stephanos Bibas, sitting as the trial judge in Delaware
The Philadelphia-based appeals court has now backed that decision.
Why AI companies are watching
Dozens of lawsuits from authors, news publishers, music labels and others accuse AI companies of training on their work without permission. The biggest names, including OpenAI, argue that training is fair use because it turns the material into something new. OpenAI made exactly that case in its fight with authors over books taken from the LibGen piracy site.
This is the first of those disputes to reach a US appeals court, so other judges will look closely at it. But there’s an important difference: Ross’s tool wasn’t generative AI. It didn’t write new text, it pointed users to existing court opinions, which made it a more direct competitor to Westlaw. AI companies are likely to argue that chatbots and image generators are far more “transformative” than a search engine built to replace the product it copied from.
“Respecting copyright is essential”
A Thomson Reuters spokesperson said the company was pleased with the ruling and “firmly believes that respecting copyright is essential for fostering innovation while protecting the intellectual property that fuels fiduciary-grade AI solutions.” Ross and its lawyers hadn’t commented. Thomson Reuters is the parent company of Reuters News.
The ruling also lands as more publishers look for ways to get paid for their work, such as the libraries and archives that have quietly struck deals with AI labs.
Why it matters
Until now, AI companies could point out that no appeals court had ruled against training on copyrighted work. That argument is gone. The ruling doesn’t decide the big generative AI cases, and the judges said so in writing. But its finding that AI training data is a real licensing market gives rights holders their strongest precedent yet.
Sources: Third Circuit opinion, Thomson Reuters v. ROSS Intelligence, No. 25-2153 (Sept. 29, 2026), Reuters (September 29, 2026).
Latest Policy & Safety news
