A federal appeals court ruled that ROSS Intelligence broke copyright law by training its legal search tool on 2,243 Westlaw summaries, and said the court’s reasoning does not reach AI that writes new text.

The U.S. Court of Appeals for the Third Circuit ruled that legal research startup ROSS Intelligence broke Thomson Reuters’ copyright by using summaries written by Westlaw editors to train an AI search tool built to compete with Westlaw. The court held that this copying was not “fair use,” the part of copyright law that allows copying protected work without permission in some cases.

The court called the dispute “no more than an ordinary copyright case.” ROSS’s tool did not write new text. The tool returned passages from existing court opinions. The court said arguments the Justice Department has made in a separate case for AI systems that generate new text, such as OpenAI’s and Anthropic’s models, do not apply to ROSS.

ROSS trained its tool on copies of Westlaw’s summaries

Westlaw is Thomson Reuters’ online legal research service. Its editors write short summaries, called headnotes, that appear above court opinions on Westlaw. Each headnote states one legal point from the opinion in a way that makes sense without reading the opinion.

ROSS built a tool that answered legal questions typed in plain English with matching passages from about 10 million court opinions. To teach the tool which passages answer which questions, ROSS hired a company called LegalEase to write about 25,000 training memos. Each memo paired a legal question with four to six passages, graded from “great” to “irrelevant.” The memo writers built the questions from Westlaw headnotes, which the case evidence described as “an easy way” to frame them. ROSS then trained its AI on the memos.

A federal trial court in Delaware compared the memo questions with the headnotes they resembled and found that 2,243 headnotes had been copied: the questions matched the headnotes’ wording, not the opinions’. ROSS did not dispute on appeal that the memo writers copied them.

Court opinions carry no copyright, and anyone can copy them. The appeals court held that headnotes are different, because an editor decides which legal points to include and how to word them.

The court found ROSS copied Westlaw’s work to do Westlaw’s job

Courts decide fair use by weighing four questions: why the copier used the work, what kind of work it is, how much was copied, and whether the copying hurts the market for the original. The court found that purpose, amount, and market harm weighed against ROSS. Only the kind of work favored ROSS, because headnotes mostly state legal facts, and copying factual work is more often allowed than copying creative work.

On purpose, the court said Westlaw uses its headnotes to help researchers find court opinions that answer their questions, and ROSS used the headnotes to build a tool that does the same. Training an AI on the headnotes arguably made ROSS’s use slightly different from Westlaw’s, the court said, but not enough, because ROSS used that AI for a research service that competed with Westlaw. ROSS ran ads comparing itself to Westlaw, set prices “in line with” Westlaw’s, and some law firms switched from Westlaw to ROSS.

The court said ROSS did not need the headnotes, because it could have written its own questions from the freely available court opinions. “Unlike necessity, ease is not a justification for copying,” the court wrote.

The court also cited evidence that ROSS employees tried to access Westlaw using login details from a law firm that had invested in ROSS, and that one employee used a student account while hiding that he worked for a competitor. The court said those actions count against ROSS, to the extent that a copier’s honesty still matters in deciding fair use.

The court said ROSS cost Thomson Reuters the chance to sell its summaries as AI training data

The court found that ROSS harmed Thomson Reuters in two markets. The first is legal research itself: ROSS built a rival service from Westlaw’s work, which made the headnotes less of a reason to pay for Westlaw.

The second is the sale of licenses to use headnotes as AI training data. The court called that market “rapidly developing” and noted that Thomson Reuters trains its own AI search product on its headnotes. Thomson Reuters had never licensed its headnotes to anyone else for that purpose, but the court said that did not mean the market did not exist. By using the headnotes as AI training material without permission, ROSS got for free what Thomson Reuters could have charged other companies to license.

The court said its ruling does not decide generative AI cases

On Sept. 1, the Justice Department filed a statement in a separate copyright case against OpenAI in federal court in New York. The department argued that training a large language model, the technology behind chatbots such as ChatGPT, gives copied material a new purpose because the model can write original responses. The department relied on a 2025 ruling in a copyright case against Anthropic. The department also argued that the training in the OpenAI case did not create a product competing with the copied works.

The Third Circuit said the department’s arguments do not apply to ROSS, whose tool cannot write anything new and was built to replace Westlaw. The court also noted that the Justice Department did not file anything in the ROSS case.

The ruling came partway through the case

Early in the case, the Delaware trial court ruled that Westlaw’s headnotes are protected by copyright and that fair use did not cover ROSS’s copying of 2,243 of them. Both sides asked to appeal those two questions before the rest of the case went ahead. The appeals court agreed with the trial court on both.