Third Circuit Rules AI Training on 2,243 Westlaw Headnotes Was Not Fair Use

On 29 September 2026, the US Court of Appeals for the Third Circuit held that ROSS Intelligence infringed Thomson Reuters' copyright by using 2,243 Westlaw headnotes to train an AI legal search engine, and that the use was not fair use (opinion, No. 25-2153). Judge Tamika Montgomery-Reeves, writing for a panel with Judges Restrepo and Bove, affirmed the February 2025 ruling of Judge Stephanos Bibas. LawNext called it the first federal appellate decision to address fair use in the context of AI training. The court was careful to say what it did not decide: ROSS's system was not generative AI.

What happened

ROSS, founded by three University of Toronto students, built a search engine that answered plain-language legal questions with passages from judicial opinions. According to the opinion, it hired LegalEase Solutions to write about 25,000 training memos, each pairing a legal question with four to six passages, and the memo writers used Westlaw headnotes to draft them. Thomson Reuters sued in 2020.

The panel framed the appeal plainly. "In truth, this is no more than an ordinary copyright case," the opinion says. It decided two questions:

• Are headnotes copyrightable? Yes. Editors choose which points of law to extract and how to phrase them, and the court found the 2,243 headnotes at issue have the "creative spark" that originality requires.
• Was training on them fair use? No. The second factor (nature of the work) weighed slightly for ROSS; the first, third and fourth weighed against it.

On the first factor, the court found ROSS used the headnotes for the same ultimate purpose as Westlaw, "making ROSS's use minimally transformative, at best." It rejected the argument that copying was needed to train the model, because the judicial opinions themselves were available: "Unlike necessity, ease is not a justification for copying." On market harm, it found that "the market for licensing headnotes as text to train AI is rapidly developing", and that ROSS usurped Thomson Reuters' chance to enter it.

ROSS, which has shut down, says it will ask the Supreme Court to review the ruling, per LawNext. The appeal was interlocutory, so the case returns to the District of Delaware for the remaining issues.

The numbers

Why it matters

It is binding law in the Third Circuit (Delaware, New Jersey and Pennsylvania, where a lot of companies are incorporated or sued) and persuasive everywhere else. Three parts of the reasoning will travel.

First, the market-harm analysis treats a developing market for licensing content as AI training data as a relevant derivative market. Thomson Reuters showed it uses its own headnotes to train its own AI products. As more publishers sign training data deals, that argument gets easier for every plaintiff.

Second, "ease is not a justification" narrows the intermediate-copying defence that software companies have relied on since the reverse-engineering cases. If the facts or uncopyrighted source material were available, copying the annotated version because it was convenient does not count as necessity.

Third, the court noticed the Justice Department. A footnote records that on 1 September 2026 the DOJ filed a statement of interest in the consolidated OpenAI copyright litigation in New York, relying on Bartz v. Anthropic to argue that training a large language model that generates original responses is transformative. The panel said those concerns "do not apply here", and that the DOJ "notably did not" weigh in on ROSS.

What sceptics say

The ruling is narrower than the headlines. The opinion stresses that ROSS's AI "was not a generative AI" and would only return existing passages. Commentators at ArentFox Schiff note the court expressly distinguished generative AI and that no federal appellate court has yet resolved fair use for generative model training. AI developers will argue, with the DOJ apparently on their side, that a model producing new text is a different purpose from a search engine competing directly with Westlaw.

The facts were also unusually bad for the defendant. ROSS wanted to be a direct substitute for Westlaw, and the court noted evidence that ROSS accessed Westlaw in breach of its terms of service. A lab with clean data sourcing and a clearly different output will look very different on factors one and four.

And on the other side, the Anthropic settlement shows the costs can be large even when training itself is found fair: the Bartz court found training on books was fair use, but storing pirated copies was not, and the case still settled for $1.5 billion.

What it means for engineering and DevOps teams

My take: this is a data provenance ruling in legal clothes, and data provenance is an engineering problem.

• Know what is in your fine-tuning and retrieval sets. If a dataset includes annotations, summaries or editorial content scraped from a commercial product, you are closer to ROSS than to a general-purpose LLM. Keep a manifest of sources and licences per dataset version, the same way you would keep an SBOM. My notes on RAG versus fine-tuning cover where that data usually comes from.
• Prefer primary sources over someone's annotations of them. The court's necessity point is practical advice: train on the opinions, not the headnotes, if the opinions are what you need.
• Check terms of service on anything you crawl. The court counted a TOS breach against ROSS on good faith. Your crawler's access pattern can become evidence.
• Budget for licensing. A court has now said a training data licensing market exists. Procurement will start pricing that in.

FAQ

Did the Third Circuit decide that AI training is not fair use?

It decided that ROSS's use of 2,243 Westlaw headnotes to train a non-generative legal search tool was not fair use. The opinion expressly distinguishes generative AI, so it does not settle whether training large language models is fair use.

What happens next in Thomson Reuters v. ROSS?

ROSS says it will ask the Supreme Court to review the decision. Meanwhile the case returns to the District of Delaware for the claims this interlocutory appeal did not resolve, since the summary judgment covered only 2,243 headnotes.

How does this affect the OpenAI and Anthropic copyright cases?

It is not binding on courts in California or New York, where most generative AI cases sit, but its reasoning on licensing markets and necessity will be cited. The DOJ has separately told the New York court it views LLM training as transformative.

Related: Open Weights in 2026: Chinese Labs Ship the Frontier, the Licenses Grew Clauses, and Brussels Started Enforcing, Is Fine-Tuning LLMs Worth It in 2026? and RAG vs Long Context vs Fine-Tuning.