A federal appeals court says training an AI on Westlaw headnotes wasn't fair use. Its footnotes say the case stops there.

The first federal appeals court to rule on whether training an AI on copyrighted material is fair use did it in a case about a legal search tool, not a chatbot. On September 29 a three-judge panel of the Third Circuit ruled that ROSS Intelligence’s use of Thomson Reuters’s Westlaw headnotes was not fair use. The opinion says more than the headline, and it pulls back in one place.

What the court actually held

Westlaw’s headnotes are the short summaries that sit above a court opinion and state a point of law. Thomson Reuters’s editors write them, and the panel said each one has to stand on its own without the opinion or the other headnotes. ROSS wanted a search engine that answers plain-language legal questions with passages from real judicial opinions. To train it, a contractor called LegalEase wrote about 25,000 training memos. Each memo posed a legal question and listed the opinion passages that answered it, and the questions were built from Westlaw headnotes.

The panel made two findings. The 2,243 headnotes at issue were original enough for copyright because they show a “creative spark.” And ROSS’s use was not fair. Circuit Judge Tamika Montgomery-Reeves wrote that the case was “no more than an ordinary copyright case.” That framing matters later, because it’s the court’s own reason for not going further.

The footnote that matters more

Footnote 7 is where the panel discusses the Justice Department. In early September the DOJ filed a statement of interest in the OpenAI copyright litigation in Manhattan, arguing that training a large language model qualifies as fair use, because the model can “generate original responses.” It relied on a 2025 ruling in Bartz v. Anthropic. The Third Circuit’s answer is two sentences long: those concerns “do not apply here,” because ROSS’s system “cannot generate original expression.” It only returns passages that already exist in court opinions.

That’s the distinction the whole ruling hangs on. A tool that hands back existing text is a different thing from a model that writes new text, and the panel didn’t decide the second question. The court says itself that this isn’t a broad ruling on AI training. Read the footnote and the narrower version is the one that holds: training a non-generative legal search tool on a competitor’s headnotes was not fair use.

Why the 0.08 percent argument lost

ROSS argued the amount it took was tiny, about 0.08 percent of Thomson Reuters’s roughly 28 million headnotes. The panel said that’s the wrong way to count. Small is beside the point. Taking a small slice isn’t fair if the slice is the important part, and the opinion says ROSS copied the entire text of the 25,000 Westlaw headnotes into its memos. The court also said ROSS didn’t need the headnotes at all, because the underlying judicial opinions were freely available to it.

On market harm, the panel wrote that the market for licensing headnotes as AI training data is “rapidly developing.” Thomson Reuters, it noted, uses its own headnotes as training data for its AI search products. ROSS offered no evidence to rebut that.

What happens next

ROSS is now defunct, but its lawyers at White & Case told Courthouse News they’ll ask for Supreme Court review, so this isn’t final. The court also left one question open. In footnote 5 it says the lower court had mused that headnotes quoting opinions word for word might still be original, and it did not answer that, because the 2,243 headnotes at issue don’t copy opinion text verbatim.

My view is that the ruling is strong on its facts. A direct competitor, near-verbatim copying of a paid editorial product, and a licensing market that already exists: that’s a lot of facts pointing one way. As a template for a chatbot trained on the open web, it tells you almost nothing. The court says as much. The next test is a case about a model that writes new text, where the DOJ’s argument that training changes what the work is for has to be answered directly. Until then, I’d treat this as a useful precedent for editorial databases and not as a verdict on AI training.

For publishers, the practical lesson is narrow. The easiest copyright case to win is a competitor copying your structured editorial work to build a rival product. The harder case, publishers letting search engines use their pages, got its own ruling the same week. Judge Amit Mehta dismissed two publisher antitrust suits against Google, which is a different theory and a different fight. Related: the Google AI Overviews suits and why they were dismissed.

Sources

1. Third Circuit opinion, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153, filed September 29, 2026: full opinion (PDF)
2. Courthouse News Service: AI training of copyrighted material not fair use: Third Circuit, September 30, 2026
3. Nieman Journalism Lab: Federal appeals court upholds Thomson Reuters’ landmark AI copyright win, September 30, 2026
4. Stanford Fair Use Project: case summary, Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc.