by Dennis Crouch
ROSS Intelligence has filed its opening brief in the Third Circuit, challenging Judge Stephanos Bibas's decision that the AI company's use of Westlaw headnotes to train its legal search engine constituted copyright infringement. The case presents fundamental questions about the intersection of copyright law and artificial intelligence training that could significantly impact the development of AI systems across industries. Thomson Reuters Enter. Ctr. GmbH v. Ross Intel. Inc., No. 25-2153 (3d Cir. appeal filed Sept. 2025).
ROSS Intelligence built an AI legal search engine designed to let users ask legal questions in plain English and receive ranked, cited passages from judicial opinions as answers. To train the system, ROSS commissioned 25,000 legal memoranda from LegalEase Solutions, each posing a question and providing multiple ranked answers. Many of the questions were adapted directly from Westlaw headnotes summarize points of law from judicial opinions. These headnote-derived questions served as labeled examples that helped the AI learn how to identify, rank, and distinguish between relevant and irrelevant case passages. Although the headnotes themselves were not displayed to ROSS users, they were incorporated into the training process as inputs, making them central to the copyright infringement dispute between Thomson Reuters (owner of Westlaw) and ROSS.
Judge Bibas' February 2025 decision granted summary judgment in favor of the copyright holder -- finding that Westlaw's headnotes were sufficiently original for copyright protection and that ROSS's use for training did not qualify qualified as fair use. The district court certified the copyright questions for immediate interlocutory appeal under 28 U.S.C. § 1292(b), recognizing the novel and difficult nature of the issues presented.
In its newly filed appeal, ROSS presents two issues:
- Is a short quote or paraphrase of a judicial holding copyrightable?
- Does the fair use doctrine protect ROSS’s internal use of Westlaw’s headnotes in memos that served as training data for an AI legal search engine that produced only non-infringing outputs?
To continue reading, become a Patently-O member. Already a member? Simply log in to access the full post.