Thomson Reuters v. ROSS: what the first AI-training appellate ruling actually decided

On September 29, 2026, the U.S. Court of Appeals for the Third Circuit affirmed the district court's partial summary judgment for Thomson Reuters in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. It is the first federal appellate decision on whether using copyrighted material to train an AI system is fair use. The opinion was initially filed under seal, and a public version followed on September 30.

The short version. Copying a competitor's copyrighted editorial content to train a non-generative tool built to compete with that competitor was not fair use. The court did not hold that "AI training is not fair use," and it did not decide how fair use applies to generative AI.

The facts that drove the result

Westlaw headnotes are short summaries that Thomson Reuters editors write to identify points of law in judicial opinions. ROSS Intelligence wanted to build a competing AI legal-research tool. Thomson Reuters declined to license Westlaw to it, so ROSS obtained training material from a third party: memos of legal questions and answers that were written using Westlaw headnotes. ROSS used that material to train a system that took a natural-language question and returned passages from existing judicial opinions.

Two features of that system shaped everything that followed:

  • It was not generative. It retrieved and ranked existing judicial text. It did not write new prose.
  • It was built to replace the product it learned from. ROSS was marketed as an alternative to Westlaw for legal research.

Judge Stephanos Bibas, a Third Circuit judge sitting by designation in the District of Delaware, granted partial summary judgment to Thomson Reuters in February 2025, revising his own earlier ruling. He held that more than 2,000 headnotes were infringed and rejected ROSS's fair-use defense. He then certified those questions for interlocutory appeal, and the Third Circuit heard argument in June 2026.

What the Third Circuit held

  1. The headnotes are copyrightable. Choosing which points of law to summarize, and how to word them, clears the "extremely low" originality bar of Feist.
  2. Merger does not apply. ROSS argued that a headnote merges with the legal point it states. The court treated merger as limited to the rare case where an idea can be expressed in only one or a few ways, and found that headnotes can be written in many ways.
  3. The use was not fair. The court found the use highly commercial and only minimally transformative. ROSS used the headnotes for essentially the same purpose, legal research, in a product meant to compete with Westlaw. On market harm, it found harm to Westlaw's existing legal-research market and to a developing market for licensing headnotes as AI training data.

What it did not decide

  • Generative AI. ROSS's tool returned existing passages. Whether training a generative model, which produces new expression, is transformative remains open. Commentators across the spectrum have stressed this limit.
  • A rule for every training use. The result turned on direct competition and on copying a competitor's editorial work product. Training a model for a different purpose, on different material, in a different market, raises different factor-one and factor-four questions.
  • Law outside the Third Circuit. The decision binds federal courts in Delaware, New Jersey, Pennsylvania and the Virgin Islands. Elsewhere it is persuasive authority only. The generative-AI rulings in Bartz v. Anthropic and Kadrey v. Meta remain district-court decisions in the Northern District of California.

Why it still matters

The training-data licensing market now has appellate recognition. Whether lost licensing revenue for training data counts as market harm under factor four was one of the central open questions in AI copyright law. An appellate court has now treated that market as cognizable, at least where a licensing market is developing and the user is a direct competitor. Expect publishers and data licensors to cite this holding in every training dispute.

Competitive substitution weighs heavily. The facts that sank ROSS were that it used a rival's content to build a rival's product. The closer an AI product sits to substituting for its source material, the weaker a fair-use defense becomes.

Licensing is now part of the risk analysis. If content you train on is available for license and you did not license it, this case gives the rights holder a factor-four argument. That changes how in-house teams should weigh "license it" against "rely on fair use."

What to do with it

  • If you build AI products: document where your training data came from, and look harder at any source owned by a company your product competes with.
  • If you publish or license content: a documented, priced training-data license strengthens the market-harm side of any future dispute.
  • If you advise either side: read the opinion itself before relying on summaries, including this one, and keep the generative and non-generative questions separate.

Go further: take the free 12–15 minute lesson on the case, read the founder briefing on training AI on licensed content, or work through fair use and AI training in depth in IP Law in the Age of AI.

Sources

For education and general information only. Not legal advice.