AI Training Fair Use: Thomson Reuters v Ross

Published 13 October 20267 minute read

Is copying protected material to train an AI model fair use in the United States? On 29 September 2026 the Third Circuit gave an answer, on narrow facts. In Thomson Reuters v Ross Intelligence, the Third Circuit rejected AI training fair use. The copied works had been used to build a competing product.

The ruling matters to anyone who builds, licenses or buys AI tools trained on third party content. This article explains what the court decided and what it did not. It then compares the position in India, the UK, the EU and Japan.

Key points

Thomson Reuters Enterprise Centre GmbH and West Publishing Corp v Ross Intelligence Inc, No. 25-2153, was argued on 11 June 2026 and decided on 29 September 2026. Judges Restrepo, Montgomery-Reeves and Bove sat on the panel. Judge Montgomery-Reeves wrote the opinion.
The court held that the 2,243 Westlaw headnotes at issue were original enough to be protected by copyright. It rejected the merger argument.
Ross's use was commercial and "minimally transformative, at best". Factors one, three and four weighed against fair use. Factor two favoured Ross only slightly.
The court recognised a potential market for licensing headnotes as AI training data. It held that Ross usurped that market.
The court distinguished generative AI models. It said Ross's tool "cannot generate original expression".

The background

Thomson Reuters owns Westlaw. Westlaw editors write headnotes that summarise points of law in judicial opinions. Ross Intelligence wanted to build an AI legal research tool. It obtained training material based on Westlaw headnotes through a third party, LegalEase. For this appeal, Ross did not dispute that LegalEase and Morae Global copied headnotes. It also did not dispute that their actions were attributable to it.

Judge Stephanos Bibas, sitting by designation in the District of Delaware, granted Thomson Reuters partial summary judgment. He then certified two questions for interlocutory appeal. The first was whether the headnotes and the West Key Number System are original as a matter of law. The second was whether Ross's use of the headnotes was fair use.

Copyright in the headnotes

The court applied the low originality threshold of Feist v Rural Telephone. It found that each headnote at issue showed some creative spark. Editors choose which points of law to state and how to word them.

Ross argued that the idea and its expression had merged. The court disagreed because "there are many ways to express points of law from judicial opinions". It answered the policy concern briefly. Headnotes are not law. Judicial opinions are, and they stay free to use. The court distinguished the Second Circuit's Matthew Bender decision. Industry conventions did not dictate how these headnotes were written.

The four fair use factors

Fair use in the US is governed by 17 U.S.C. 107. Ross, as the party relying on the defence, carried the burden.

Purpose and character. The court asked whether Ross had a distinctive purpose. It found that Ross shared the same ultimate purpose as Westlaw, which is legal research. Training an AI model was an intermediate step that did not change that purpose. The court relied on Warhol v Goldsmith. It distinguished Authors Guild v Google, where search served a different function. It also distinguished Sega, Connectix and Google v Oracle, where copying was needed to reach unprotected functional elements. Ross could have used the judicial opinions themselves. In the court's words, "Unlike necessity, ease is not a justification for copying."

Nature of the work. The headnotes were published and more factual than fictional. This factor favoured Ross slightly. The court noted that it rarely plays a significant role.

Amount used. Ross copied the full text of each headnote, and each headnote was treated as a whole work. The copying was not necessary because the underlying opinions were freely available. This factor weighed against Ross.

Market effect. Ross's tool was meant to compete with Westlaw. The court also found a potential derivative market for licensing headnotes as AI training data. It held that "ROSS usurped Thomson Reuters's opportunity to enter that derivative market". This factor weighed against Ross. The court found no evidence of public benefit that could outweigh the harm.

Weighing the factors together, the court held that the use was not fair. It affirmed the partial summary judgment. The interlocutory appeal ends, and the remaining issues stay before the District Court.

What the ruling does not decide

It does not decide whether training a generative AI model is fair use. The court said Ross's tool, unlike the models in Bartz and In re OpenAI, cannot generate original expression.
It does not decide whether the West Key Number System is original. Ross forfeited that challenge.
It leaves open whether headnotes that quote an opinion word for word are original.
It is a Third Circuit ruling. Courts in other circuits may weigh the factors differently, especially for highly transformative models.

India

India has no open ended fair use doctrine. Section 52(1)(a) of the Copyright Act, 1957 allows fair dealing only for listed purposes. These are private or personal use including research, criticism or review, and reporting current events. Commercial training of a competing product fits none of them easily. The Act has no express text and data mining exception.

Indian law also treats editorial content in law reports as protectable. In Eastern Book Company v D.B. Modak (2008) 1 SCC 1, the defendants conceded before the Delhi High Court that the publisher had copyright in its headnotes. They undertook not to copy them. At the same time, section 52(1)(q)(iv) allows anyone to reproduce a judgment or order of a court unless the court prohibits it. The same line as in Ross therefore applies. Judgments are free to use, but the editor's own expression is not.

For Indian AI developers, the safer course is to train on primary legal materials and licensed data. Our guide to the types of intellectual property protection explains how copyright fits alongside other rights.

Europe and the United Kingdom

EU law does not use fair use. It uses specific exceptions. Article 3 of the Digital Single Market Directive (EU) 2019/790 covers text and data mining for scientific research. It benefits research organisations and cultural heritage institutions. Article 4 requires Member States to allow anyone to mine lawfully accessible works, through national law. Under Article 4(3), it applies only if the rightholder has not expressly reserved that use. For online content, the reservation may be made by machine readable means. Article 4 also covers extractions from databases protected by the sui generis database right. For a legal publisher, a clear reservation blocks commercial mining under Article 4. It does not block research mining under Article 3.

The UK is narrower. Section 29A of the Copyright, Designs and Patents Act 1988 allows copies for computational analysis. It applies only to non commercial research by a person with lawful access. A commercial tool like Ross's would fall outside it. Article 30-4 of Japan's Copyright Act is wider and allows data analysis. It does not apply where the use would unreasonably prejudice the copyright owner's interests.

Practical steps

AI developers should map the source of every training dataset and record its licence. In Ross, the developer did not contest that its contractors' copying was attributable to it.
Never access a rival's database through borrowed or disguised accounts. The court noted that Ross had at times acted in bad faith in this way.
Prefer primary legal materials such as judgments and statutes, which are free to use in the US and India.
Publishers should assert copyright in editorial content and, for Europe, reserve text and data mining rights in machine readable form.
Buyers of AI tools should seek warranties and indemnities on training data in their contracts.

Our earlier note on copyright ownership of a logo in India explains how Indian courts approach proof of authorship. For advice on licensing, data audits or disputes, see our copyright services.

Frequently asked questions

Is training AI on copyrighted material fair use?

It depends on the facts. In Thomson Reuters v Ross, the Third Circuit held it was not fair use on the facts. Copied headnotes trained a tool that could not generate new expression and competed with the original. The court did not decide the question for generative AI models.

Are case law headnotes protected by copyright?

In the US, the Third Circuit held that the Westlaw headnotes at issue were original enough for protection. In India, the defendants in Eastern Book Company v D.B. Modak conceded that the publisher's headnotes were protected. The judgments themselves remain free to reproduce.

Does India have a text and data mining exception?

No. The Copyright Act, 1957 has no express text and data mining exception. Fair dealing under section 52(1)(a) is limited to listed purposes, including private or personal use and research.

AI training fair use: Thomson Reuters v Ross and what the Third Circuit decided