On 29 September 2026, the United States Court of Appeals for the Third Circuit affirmed a partial summary judgment in favour of Thomson Reuters against ROSS Intelligence. ROSS had used legal headnotes drafted by Westlaw editors to prepare data with which it trained its own artificial intelligence-based legal search engine.
The court concluded that the 2,243 summaries examined in the partial summary judgment possessed the minimum originality required for protection and that the use was not covered by fair use, the United States doctrine that permits certain uses of protected works without authorisation.
The opinion was initially filed under seal to allow the parties to propose potential redactions. The public and precedential version, signed by Judge Tamika Montgomery-Reeves, is available in the official court archive.
Westlaw publishes judicial decisions, which do not belong to Thomson Reuters, alongside its own editorial materials. Among these are headnotes: brief summaries that identify and formulate relevant legal points from each decision and link to the corresponding passage.
ROSS intended to build a search engine capable of answering legal questions phrased in ordinary language. To train it, the company engaged LegalEase Solutions, which produced approximately 25,000 memoranda containing a question and several passages from decisions classified by their relevance. Those who drafted these memoranda utilised thousands of Westlaw headnotes to formulate the questions. ROSS did not dispute on appeal that this conduct could be attributed to it, nor that the headnotes had been copied.
The court of first instance compared the texts and granted partial summary judgment regarding 2,243 headnotes: the questions in the memoranda closely resembled those summaries, but not the text of the underlying judicial decisions. The appeal affirmed that decision.
The court analysed the four legal factors of fair use. It did not base the outcome on a rule stating that training artificial intelligence with protected works is always unlawful.
The first factor, the purpose and character of the use, weighed against ROSS. The use was commercial and, according to the court, minimally transformative at best. Westlaw employed its summaries to assist in locating and understanding decisions; ROSS utilised them to train a platform also intended to find relevant legal passages. The intermediate training phase introduced a degree of difference, but both uses shared substantially the same purpose and ROSS aspired to replace Westlaw in the same market.
The second factor, the nature of the work, favoured ROSS slightly. The summaries were published and, although they contained sufficient editorial decisions to be original, they possessed a more factual than fictional character because they were required to accurately describe judicial decisions.
The third factor, the amount and substantiality of the portion used, again weighed against ROSS. The company argued that the materials represented only 0.08% of Thomson Reuters' approximately twenty-eight million summaries. The court responded that each headnote was a protected work and that it had been copied in its entirety. Furthermore, such copying was not necessary: ROSS had access to approximately ten million unprotected judicial decisions and could have created its training materials from them. Convenience did not equate to necessity.
The fourth factor, the effect upon the market, also favoured Thomson Reuters. ROSS utilised the headnotes to construct a substitute tool in the legal research market. The court further considered a developing derivative market: the licensing of such materials as training data. Thomson Reuters already utilised its headnotes to train its own products, and the unauthorised copying deprived the company of the opportunity to license that use.
ROSS invoked the public interest in expanding access to the law. The court noted that the underlying judicial decisions were freely accessible and that ROSS offered its service at prices comparable to those of Westlaw. It also found no evidence connecting this specific tool to the general national security arguments that ROSS raised regarding artificial intelligence.
The balance was, therefore, uneven: the second factor slightly favoured ROSS; the first, the third and the fourth weighed against it. Taken together, the use was not a fair use.
ROSS was not developing generative artificial intelligence. Its system did not create new expression: it returned existing passages from judicial decisions. The Third Circuit itself distinguished this case from litigation concerning large models capable of generating original responses.
The opinion expressly mentions first-instance decisions regarding generative models and a brief filed by the Department of Justice in other litigation. According to the court, the reasons invoked therein did not resolve this case: ROSS trained its system to create a commercial substitute for Westlaw and the available evidence led to a different conclusion regarding the transformative nature of the use.
The judgment does not establish that all competition renders a use unlawful. Substitutive competition weighed within a broader analysis: similar purpose, commercial character, complete copying of each headnote, existence of non-protected alternatives and harm to the original market and the potential licensing market.
The appeal was interlocutory and affirms a partial summary judgment. It resolves the originality of the 2,243 headnotes examined and rejects the fair use defence. It does not yet determine damages or bring the district court proceedings to a close. Matters that may remain open there include, among others, whether some rights had expired and what follows from the infringement.
In my interpretation, the most transferable element is not that artificial intelligence was involved, but the relationship between the source and the product constructed with it. ROSS did not utilise the headnotes for a purpose removed from legal research: it employed them to train a tool that aspired to replace the service from which they originated.
That logic may appear in lawsuits concerning generative AI when plaintiffs allege that the results substitute for the works or services from which the data were extracted. However, the opinion also warns of the limit of the comparison. A model that generates new expression may present a different degree of transformation and different market effects. These will have to be proven in each case.
In Europe, the question is framed through other rules: the exceptions for text and data mining, the reservations of rights and the obligations applicable to model providers. The guide 'Artificial Intelligence and Copyright in Europe' develops that framework.
About the open conversation
This article continues the Notebook’s examination of artificial intelligence, copyright and training data. It is based on the judgment of the United States Court of Appeals for the Third Circuit and the first-instance ruling. It distinguishes what was decided in this litigation from the questions that remain open in relation to generative systems. If you work in intellectual property, legal research or the development of artificial intelligence systems and can correct or refine this reading, write to me.