Thomson Reuters v. ROSS: The First US Appeals Ruling on AI Training and Fair Use

By Dafa Attas

Photo: kewl / Pixabay

In what commentators describe as the first ruling of its kind, a US federal court of appeals has decided whether copying protected material to train an AI system can be fair use. In Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153, the Court of Appeals for the Third Circuit said it was not. The opinion, written by Judge Montgomery-Reeves for a panel with Judges Restrepo and Bove, was filed on 29 September 2026 and affirms partial summary judgment for Thomson Reuters.

The court called it "no more than an ordinary copyright case", and its reasoning is narrower than the headlines suggest.

The facts

Thomson Reuters owns Westlaw. Its editors write headnotes, short summaries that identify the points of law in a judicial opinion. ROSS set out to build a competing legal search engine that would answer plain-language questions with passages from court opinions. The opinion stresses that this tool was not generative AI and produced no new expression.

To train the system, ROSS hired LegalEase Solutions, which wrote about 25,000 training memos. Each memo posed a legal question and paired it with four to six opinion passages graded by relevance. The memo writers used Westlaw headnotes to frame the questions. The district court, with Circuit Judge Stephanos Bibas sitting by designation, held that ROSS's use of 2,243 headnotes infringed copyright and was not fair use. It then certified two questions for interlocutory appeal: whether the headnotes are original, and whether ROSS's use was fair.

Headnotes are protected

The Third Circuit found that the 2,243 headnotes have the creative spark that copyright requires. Choosing which points of a judgment to summarise, and how to word them, involves editorial judgment.

The four fair use factors

The first factor looks at the purpose and character of the use. ROSS's use was commercial. On transformation, the court found the purposes close: Thomson Reuters uses headnotes to help researchers find relevant law, and ROSS used them to build a tool that finds relevant opinion passages. The intermediate step of training an AI made at most a slight difference, so the use was "minimally transformative, at best" and the first factor weighed against fair use. The court distinguished Authors Guild v. Google, where scanning books to power a snippet search tool served a function different from reading the books. According to Ballard Spahr, the court also rejected ROSS's analogy to software intermediate-copying cases, because copying the headnotes was a convenient route to the memos, not a necessary one.

On the second factor, the nature of the work, headnotes sit near the factual end of the spectrum, so this factor slightly favoured ROSS.

The third factor, the amount used, went the other way: the memos copied the headnotes in full, so it favoured Thomson Reuters.

On the fourth factor, market effect, the court found harm in two markets. ROSS designed its product to compete with Westlaw in the existing legal research market. The court also recognised a potential derivative market for licensing headnotes as AI training data, which it described as rapidly developing, relying on evidence that Thomson Reuters uses its own headnotes to train its AI search products. That Thomson Reuters had not yet licensed headnotes to others did not mean no such market exists. The court rejected ROSS's public-interest arguments, saying AI does not give ROSS "carte blanche to violate copyright law".

With three factors against fair use and one slightly in favour, the court held that ROSS's use was not fair.

Footnote 7 and the generative AI cases

The US Department of Justice had filed a statement of interest in the consolidated OpenAI copyright litigation in New York. Relying on Bartz v. Anthropic, it argued that training a large language model that can generate original responses is a transformative use. In footnote 7, the Third Circuit said those concerns do not apply here, because ROSS's platform cannot generate original expression and was trained to create a commercial substitute for Westlaw.

The ruling therefore leaves the central generative AI question open. Ballard Spahr notes that two federal district courts in California have already reasoned differently about market harm, in Bartz v. Anthropic and Kadrey v. Meta. The Third Circuit's opinion is precedential only in its own circuit, which covers Delaware, New Jersey, Pennsylvania and the US Virgin Islands.

What developers and rights holders can take from it

For AI developers, the case shows that labelled training data can be the weak point. The court opinions themselves were free to use. The problem was Thomson Reuters' editorial summaries. Teams building retrieval or classification tools on third-party annotations, tags or summaries should check where that material came from and on what terms.

For rights holders, the market analysis matters most. A federal appeals court has accepted, in a precedential opinion, that licensing content as AI training data is a market copyright law can protect, even where the owner has so far used the content only for its own models.

A note on Indonesian law

Indonesia does not have an open-ended fair use test. Law No. 28 of 2014 on Copyright lists specific exceptions instead. Under Article 44(1)(a), using, reproducing or altering a work, in whole or in substantial part, is not infringement if the source is fully cited and the use is for education, research, scientific writing, preparing reports, or writing criticism or a review, without harming the reasonable interests of the author or copyright holder. The Elucidation explains those reasonable interests as a balance in enjoying the economic benefit of a work.

We are not aware of an Indonesian court applying these exceptions to AI training. A use like ROSS's, commercial and aimed at building a competing product, would be difficult to fit into a list of exceptions organised around research, criticism and review.

Sources

• US Court of Appeals for the Third Circuit, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153, opinion filed 29 September 2026 (copy of the court filing)
• Ballard Spahr, "Third Circuit Addresses Fair Use in AI Training, But Leaves Generative AI Questions Unresolved" (2 October 2026)
• Womble Bond Dickinson via The National Law Review, article on the Third Circuit's fair use ruling in Thomson Reuters v. ROSS (7 October 2026)
• Law No. 28 of 2014 on Copyright (Indonesia), official text (BPK)