Thomson Reuters v. Ross: What the First Appellate AI Fair Use Ruling Means for Legal Tech Startups

On 29 September 2026, the US Court of Appeals for the Third Circuit affirmed that ROSS Intelligence infringed Thomson Reuters’s copyright when it used Westlaw headnotes to train an AI legal research tool, and that its use was not fair use. It is the first federal appellate decision on fair use in the context of AI training, and it concerns legal research, which is exactly the field where many legal tech startups, including teams working from India, are building.

The ruling is narrower than the headlines suggest, and it does not decide how fair use applies to generative AI. But it contains reasoning that every legal tech founder should understand, especially about training data and licensing markets. This guide explains what the court decided, what it left open, and what legal tech startups should do about their training data and contracts. It is written from the Indian-law perspective; US copyright points should be confirmed with US counsel.

What the Third Circuit decided

The case is Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153. Thomson Reuters, which owns Westlaw, sued ROSS after ROSS built a legal search tool using training material derived from Westlaw’s headnotes, the short summaries of legal points that Westlaw editors write for judicial opinions. In February 2025, the District of Delaware held that the headnotes were sufficiently original to be protected and that ROSS’s use was not fair use, and the appeal was taken on that ruling (Ballard Spahr; Klehr).

On 29 September 2026, the Third Circuit affirmed. Reports of the decision say the court found that 2,243 Westlaw headnotes were protected by copyright and that ROSS’s copying to train its tool was not fair use (Jenner & Block; Akin Gump; Haynes Boone). The court weighed the four statutory fair use factors and held that the first, third and fourth factors weighed against ROSS, with the second factor weighing slightly in its favour.

A note of caution on sources. The public initially saw a one-page judgment reading “affirmed”, and reports differ on whether the panel’s full opinion was sealed or has since been unsealed (Let’s Data Science). Read the opinion itself, on the court’s docket, before relying on any detail in this article or in law-firm summaries.

Why the court found no fair use

On the first factor, the purpose and character of the use, the court treated ROSS’s use as highly commercial and called it “minimally transformative, at best”, because ROSS used the headnotes for the same ultimate purpose as Thomson Reuters: helping lawyers research the law. Training an AI system on copyrighted material is not automatically transformative, and where the new product competes with the original, the argument is weaker (Akin Gump; Haynes Boone).

The fourth factor, the effect on the market for the original, carried the most weight for businesses. The court found harm in two markets: the existing market for legal research tools like Westlaw, and a developing market for licensing headnotes as AI training data. According to reports, it credited evidence that Thomson Reuters already uses the headnotes to train its own AI search products, and held that the fact that Thomson Reuters had not yet licensed them to anyone did not make the market illusory (Ballard Spahr; Mondaq).

That second point is why the ruling matters beyond legal research. A court at the appellate level has accepted that licensing material as AI training data can be a market that a copyright owner is entitled to protect. For AI companies, it means that arguing that no licensing market exists is harder to win, and for rights holders, it is a reason to build one.

What the ruling does not decide

The court was careful to limit its holding. ROSS’s tool was not generative: it returned existing judicial opinions in response to a query and did not produce new text. The court distinguished it from generative AI systems, and commentators emphasise that the decision leaves open how fair use applies to large language models that generate original expression (Ballard Spahr; Mondaq). Other US district courts that have looked at generative training have reached mixed results, so there is still no binding appellate answer on that question.

The facts were also particular. ROSS copied headnotes that were proprietary editorial material, ROSS was a direct competitor, and the copying took place before the current wave of generative AI. A court looking at training on the open web, on public-domain judgments or on licensed material would start from different facts. And the ruling concerns US law, so it does not bind Indian courts, which apply the Copyright Act, 1957. In ANI v. OpenAI, a Delhi High Court judge refused an interim injunction in July 2026 and viewed training as arguably fair dealing, but that ruling is interim and under appeal.

So the right reading is neither “training is always infringement” nor “training is safe”. The decision says that where an AI company copies a competitor’s protected editorial content to build a competing tool, fair use is hard to establish, and that the existence of a training-data licensing market counts against the copier.

Practical lessons for legal tech startups

Start with provenance. For every dataset in your pipeline, record where it came from, who owns it and on what terms you have it. Primary legal materials, such as statutes and judicial opinions, are generally in the public domain or free to use in many jurisdictions, but editorial layers added by commercial publishers, including headnotes, summaries, annotations, citators, classification schemes and curated databases, are exactly the kind of material the Ross courts protected. If your tool was trained on a vendor’s editorial content, you need a licence or a plan to replace it.

Review your sources of derived data. Some startups obtain training data from third parties that scraped it, or use labels produced by humans or models that were themselves trained on someone’s protected content. Ask suppliers for warranties about the origin and rights in what they deliver, and for indemnities. Check the terms of any legal database you subscribe to, because many prohibit using content to train models, and a contract breach can exist even where copyright does not.

Think about how your product competes. The court’s reasoning was strongest where the AI tool substituted for the original product. A tool that analyses documents a user uploads, drafts from open sources or cites to public law presents different risks from one that was built to replace a publisher’s search product. Also watch your marketing: statements that your system was trained on “all of Westlaw” or “every case summary” are evidence for the other side. Keep a record of how you decided to use each source.

Risk management, remedies and quick answers

If your product has already been built on questionable data, don’t wait for a demand letter. Audit the training sets, quarantine content whose rights you cannot show, consider retraining on licensed or public-domain material, and approach likely rights holders about licences while the commercial relationship is still friendly. Keep communications factual. If you receive a claim, preserve records, limit internal commentary and get US counsel involved immediately, because remedies in US copyright cases include injunctions, damages and orders affecting the product. Founders based in India should also remember that an Indian entity can be sued in the US where it targets US customers, and that Indian rights holders can assert claims in India. Build licensing costs into your pricing from the start, because after Ross, a licence for training data is more likely to be seen as a normal cost of doing business than as a favour.

Related reading: copyright in legal headnotes and AI training, how US AI companies should structure data deals after Ross, the October 2026 AI music copyright suits, AI copyright in India vs the US, ANI v. OpenAI and AI training as fair dealing, AI training data licensing with Indian publishers and open-source licence audits for US startups; also AI training data provenance audits after Bartz.

Quick answers

Is training an AI model on copyrighted material always fair use?

No. The Third Circuit held in Thomson Reuters v. Ross that copying Westlaw headnotes to train a competing, non-generative legal research tool was not fair use, though the ruling does not decide how fair use applies to generative AI.

Does the Ross ruling apply in India?

No. It is a US decision. India’s Copyright Act, 1957 has no text-and-data-mining exception, and an Indian court decided ANI v. OpenAI differently at the interim stage in July 2026, with an appeal pending.

What should a legal tech startup do about its training data?

Document where each dataset came from, license or replace any proprietary editorial content such as headnotes, review database terms, and get warranties and indemnities from data suppliers.

Does Ross decide whether training a generative AI model is fair use?

No. The tool in Ross was not generative, and the court distinguished generative systems. It is the first federal appellate ruling on fair use in AI training, but other courts have reached different results on different facts.

Is the Ross opinion available?

Reports conflict on whether the full opinion was sealed at first and later unsealed. Read it directly from the court before relying on any summary, including this one.

Final word

Halverton & Co. is an Indian law firm advising US companies, founders, investors and law firms on Indian law, including AI and copyright strategy, training data licensing and technology contracts for legal tech startups. We practise in Jharkhand, Maharashtra and before the Supreme Court of India, and work alongside US counsel, who advise on US law. Halverton & Co.: Where tech needs law! If you have a question about Indian law, write to us at office@halvertonandco.com, or get in touch.

This article reflects developments reported up to late October 2026. It is for general information only, is not legal advice, and does not create an attorney-client relationship. Halverton & Co. is an Indian law firm and does not advise on US law; US-law points should be confirmed with US counsel.