Thomson Reuters v Ross: Third Circuit rejects fair use for AI training

On 29 September 2026 the US Court of Appeals for the Third Circuit affirmed that ROSS Intelligence infringed Thomson Reuters’ copyright by using Westlaw headnotes to train an AI legal-research tool, and that the use was not fair use. Thomson Reuters v Ross in the Third Circuit is an appellate ruling on copying for AI training, and it protects editorial content against a direct competitor. It matters to publishers, database owners and anyone building or buying AI tools trained on third-party material.

Key takeaways

  • The court held that the 2,243 Westlaw headnotes at issue are original and protected by copyright, even though they summarise public judicial opinions.
  • Three of the four fair use factors weighed against ROSS, which used the headnotes to build a competing substitute.
  • The court separated this case from generative AI disputes such as Bartz v Anthropic and Kadrey v Meta.
  • The emerging market for licensing content as AI training data counted as a market the copying harmed.

What did the Third Circuit decide in Thomson Reuters v Ross?

ROSS built a non-generative AI search engine that answered legal questions with passages from judicial opinions. To train it, ROSS commissioned around 25,000 memos from LegalEase Solutions, whose writers used Westlaw headnotes to frame the questions.

In a memorandum opinion of 11 February 2025, Judge Stephanos Bibas, a circuit judge sitting by designation in the District of Delaware, granted Thomson Reuters partial summary judgment on 2,243 headnotes and certified two questions for interlocutory appeal: originality and fair use.

The Third Circuit’s opinion in No. 25-2153, written by Judge Montgomery-Reeves for a panel with Judges Restrepo and Bove, answered both against ROSS. The headnotes show the “creative spark” required by Feist: editors choose which points of law matter and how to word them. Protecting them gives no monopoly over the law, because the opinions remain free, and the merger doctrine does not apply, since a point of law can be expressed in many ways.

Why did ROSS lose on fair use?

US fair use (17 U.S.C. § 107) weighs four factors; ROSS bore the burden of proof.

Factor Court’s finding Weighs
1. Purpose and character Commercial use with the same ultimate purpose as Westlaw: helping users find relevant case law. Training was an intermediate step, not a new purpose. Against
2. Nature of the work Headnotes are published and more factual than fictional. For ROSS
3. Amount used Each headnote is a work, copied in full; copying was not necessary because the opinions were freely available. Against
4. Market effect Harm to Westlaw’s value and market, and to the developing market for licensing headnotes as AI training data. Against

Authors Guild v Google did not help ROSS: Google Books sent readers to the original works, while ROSS aimed to replace Westlaw. Nor did the intermediate-copying cases (Google v Oracle, Sega, Connectix), where copying was needed to reach unprotected software interfaces; ROSS simply found headnotes convenient. In the court’s words, “Unlike necessity, ease is not a justification for copying.” The court also noted evidence that ROSS sought Westlaw access through other people’s accounts despite the terms of service.

How did the court distinguish generative AI cases?

In a footnote, the court addressed the US Department of Justice’s statement of interest filed on 1 September 2026 in the OpenAI copyright litigation in New York, which relied on Bartz v Anthropic to argue that training a large language model is transformative. The court said those concerns do not apply: ROSS’s tool cannot generate original expression, and it was trained to create a commercial substitute for Westlaw. The ruling is not a verdict on generative AI training, but it signals that copying a competitor’s editorial layer to build a rival product will be hard to defend.

Would the result be the same in the EU and Spain?

Europe has no open-ended fair use, and its tools are in some respects stronger for content owners:

  • Original texts, which headnotes or summaries can be, are protected as works (Article 10 of Spain’s consolidated Intellectual Property Act, TRLPI).
  • Databases enjoy a separate sui generis right that protects substantial investment in obtaining, verifying or presenting content and lets the maker prohibit extraction or re-use of a substantial part (Article 133 TRLPI, implementing Directive 96/9/EC).
  • The text and data mining (TDM) exception in Article 4 of Directive (EU) 2019/790 requires lawful access and does not apply if the rightsholder has expressly reserved its use, by machine-readable means for online content. Spain transposed it in Article 67 of Royal Decree-Law 24/2021.
  • Unlike the research exception, Article 4 is not shielded from contractual override, so licence terms matter. Access through borrowed credentials would hardly count as lawful.

What this means for your business

  1. If you publish editorial content or databases, document the creative and investment work behind them and record your rights reservations for TDM in machine-readable form.
  2. Review your licence terms: state whether AI training is permitted, by whom and for what purpose.
  3. If you build AI tools, trace every training source and avoid using a competitor’s annotations when the underlying public material is available.
  4. Treat AI training licences as a revenue line: the court recognised that market.

Our team for AI training data, licensing and digital asset protection can audit datasets and licences across Europe and Latin America.

Where companies get this wrong

  • Assuming summaries of public material are free to copy. The court protected headnotes precisely for the editorial choices they reflect.
  • Outsourcing data preparation without controls. ROSS did not contest that its contractors’ copying was attributable to it.
  • Ignoring terms of service. Access obtained in breach of contract undermines both US fair use arguments and EU lawful access.
  • Relying on US case law in Europe or Latin America, where the legal tests differ.

Reviewing data sources and contracts with one team handling IP licensing and disputes avoids fixing them after a claim arrives.

Frequently asked questions

Does Thomson Reuters v Ross mean AI training is never fair use in the US?

No. The Third Circuit stressed that ROSS’s tool was not generative and was built as a direct substitute for Westlaw. It expressly distinguished generative AI cases such as Bartz v Anthropic and Kadrey v Meta. Those disputes turn on different facts, and courts may weigh transformation and market harm differently there.

Are case law headnotes protected by copyright?

According to the Third Circuit, yes, when editors make creative choices about which points of law to include and how to word them. The judicial opinions themselves remain free to use. The court left open whether headnotes that copy opinion text verbatim would also be protected, and the case now returns to the Delaware court for the remaining issues.

Can IP Global Guard review our AI training data and licences?

Yes. We map training sources, check licences and rights reservations, and advise on EU, Spanish and Latin American exposure. Where US law is relevant for clients of the corridor, we coordinate with qualified US counsel, keeping a single point of contact for the whole review.

How IP Global Guard can help protect your content and your AI projects

IP Global Guard, the IP services line of META Channel Corporation Limited, advises on copyright, databases and AI licensing with one strategy and one billing relationship across more than 25 jurisdictions in Europe, Latin America and Africa. Within the same group we also cover the AI Act, so copyright and regulatory questions are answered together.

Tell us what content you publish or what data you train on, and in which markets you operate. We will assess your exposure and propose licence and opt-out wording. Contact our AI and copyright team.

This article is general information, not legal advice, and reflects the position on the date of publication.

Sources

Share