Third Circuit upholds Thomson Reuters win in Ross AI copyright case

The U.S. Court of Appeals for the Third Circuit has upheld Thomson Reuters’ copyright victory over Ross Intelligence, rejecting Ross’s fair-use defense over Westlaw material used to train a competing legal-search tool. The Sept. 29 decision is the first U.S. federal appellate ruling to reject a fair-use argument over publisher content used to train a commercial AI product. Its reasoning was sealed, however, and the public judgment was one page long, stating: “AFFIRMED.” The appeal affirmed a February 2025 Delaware federal court ruling by Judge Stephanos Bibas. He found that Ross infringed 2,243 Westlaw headnotes and granted Thomson Reuters summary judgment on fair use. Westlaw’s headnotes are editorial summaries of legal points in court opinions. The district court held that they contain sufficient creativity to be copyrightable, including in instances where a headnote quotes an opinion verbatim. It also found the Westlaw Key Number System met the minimum originality threshold. Ross had asked Thomson Reuters to license Westlaw content for training data, but Thomson Reuters declined because Ross was a competitor. Ross instead obtained about 25,000 legal-question-and-answer documents, known as Bulk Memos, from LegalEase to train its search tool. LegalEase instructed lawyers creating the documents to use Westlaw headnotes as a guide while not copying and pasting them directly. Thomson Reuters alleged that Ross obtained the underlying headnote material through a Westlaw licensee rather than creating it independently. Ross’s product was designed to return relevant court opinions in response to legal questions. It was not a large-language model or a generative-AI system: it retrieved existing passages from judicial opinions rather than producing original text. Ross shut down the platform in 2021, citing the cost of Thomson Reuters’ litigation. The district court concluded that Ross took the headnotes to ease development of a competing legal-research service, making the use non-transformative. It also found that a competing substitute could harm a potential market for AI-training data, even if Thomson Reuters had not itself used its material to train legal tools. Nieman Lab reported that Judge Tamika Montgomery-Reeves wrote that Thomson Reuters’ materials had a “creative spark” and that Ross sought to compete directly by using them for a highly similar purpose. That assessment is significant because the appellate court’s full reasoning was not publicly available at the time of the ruling. The case began with Thomson Reuters’ 2020 lawsuit accusing Ross of copying thousands of Westlaw headnotes. The appeal, Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., was docketed as No. 25-2153 after proceedings in Delaware federal court under No. 1:20-cv-00613. The outcome arrives amid dozens of U.S. copyright cases over AI training, many involving generative systems. The U.S. Copyright Office’s May 2025 analysis did not take a position on individual disputes, but said some training uses may qualify as fair use and others may not, with harm to markets for original works carrying substantial weight in the analysis. Other rulings have reached different results in different circumstances. In Bartz v. Anthropic, a judge found the use of published books to train Claude language models was fair use, while also finding liability for pirating and storing copies of books; that litigation later produced a $1.5 billion settlement for authors and publishers.