U.S. Court Rules That Using Westlaw Summaries to Train a Competing AI Is Not “Fair Use”

U.S. Court Rules That Using Westlaw Summaries to Train a Competing AI Is Not “Fair Use”

On September 29, 2026, the United States Court of Appeals for the Third Circuit, a federal appellate court, held that summaries of judicial decisions prepared by a legal research platform are protected by copyright. The court also concluded that their use by a competing company to train its own artificial intelligence system does not qualify as fair use.

The decision was issued in the lawsuit brought by Thomson Reuters against ROSS Intelligence. Thomson Reuters operates Westlaw, a legal research platform that compiles judicial decisions accompanied by headnotes, which are short summaries, written by editors, of the points of law in each decision. ROSS, a competitor, built an AI-powered search tool that answered legal questions by pointing to passages from judicial decisions. To train it, a company hired by ROSS drafted approximately 25,000 memos containing legal questions formulated from Westlaw headnotes. ROSS presented itself as a substitute for Westlaw, at comparable prices. The District Court ruled in favor of Thomson Reuters with respect to 2,243 headnotes, and it fell to the court of appeals, in an opinion by Judge Montgomery-Reeves, to answer two questions – whether the headnotes are original enough to be protected, and whether ROSS’s use qualifies as fair use.

Under U.S. law, copyright protects original works, and to meet this requirement it is enough for a work to be independently created with a minimal degree of creativity. For the court, the headnotes meet this requirement because editors choose which points of law to highlight and how to word them so that they can be understood without reading the decision. In addition, the court held that the headnotes are not the law itself, which lies in the judicial decisions, and these remain free for all.

As to the cases in which the law allows certain uses of a protected work without authorization, known as fair use, the court considered four factors:

The first is the purpose of the use. Instead of displaying the headnotes, ROSS used them to train an AI. For the court, this was only a slight difference because, like Westlaw, ROSS sought, for commercial purposes, to help users find relevant decisions. Nor was the copying necessary, since ROSS had access to the decisions themselves, which are public, and chose the headnotes because it was the easier path. For the court, ease does not justify copying.

The second factor, the nature of the work, was the only one favoring fair use: the headnotes had already been published and are more factual than creative, as they describe judicial decisions. The third factor assesses how much of the work was used. ROSS claimed to have used a minimal fraction of Westlaw’s roughly 28 million headnotes. The court, however, considered each headnote an independent work, copied in its entirety, and concluded that ROSS copied more than it needed, since it could have worked from the decisions themselves.

The fourth factor examines the effect of the use on the market. ROSS argued that there is no market for headnotes sold on their own. The court responded that the law also protects the value of the work: the headnotes attract subscribers to Westlaw, and ROSS used them to build a competing platform. The court also recognized a rapidly developing market for licensing text for AI training. Thomson Reuters has never licensed its headnotes to third parties, but it uses them to train its own products, and ROSS appropriated that opportunity. The alleged public benefits, such as expanding access to the law, were not proven, since the decisions are public and ROSS charged prices comparable to Westlaw’s.

With three factors against and one in favor, the court concluded that the use was not fair use and affirmed the District Court’s decision. It emphasized that the case concerned an ordinary copyright dispute, rather than a general ruling on AI. The decision indicates that the use of copyrighted works to train AI systems does not automatically qualify as fair use: the analysis depends on the circumstances of each case, including the purpose of the use, the extent of the copying, and its effects on the market.

The decision can be accessed via the following link: United States Court of Appeals for the Third Circuit

Note: For quick release, this English version is provided by automated translation without human review.