The United States Court of Appeals for the Third Circuit filed its opinion in Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. ROSS Intelligence Inc., No. 25-2153, on 29 September 2026. A panel of Judges Restrepo, Montgomery-Reeves and Bove affirmed the District of Delaware's grant of partial summary judgment to Thomson Reuters in No. 1:20-cv-00613, Judge Montgomery-Reeves writing. The appeal resolved the fair use defence and the originality of the copied headnotes.
The governing provision is 17 U.S.C. § 107. On the first factor the court found the use commercial and only minimally transformative, because the product served the same end as the Westlaw headnotes it was built from. The second factor leaned slightly toward ROSS, the headnotes being already published and closer to factual than creative. The third factor ran against ROSS, which copied entire headnotes when the underlying judicial opinions were freely available, and the court rejected the argument that taking 0.08 percent of the headnote corpus was immaterial. The fourth factor also ran against ROSS, on harm both in the legal research platform market and in a potential market for licensing headnotes as AI training data. The court held the 2,243 headnotes at issue original and copyrightable.
Model developers training on editorial content now face an appellate holding that a potential licensing market for training data counts in the fourth factor analysis. Legal research vendors, publishers and database owners gain a citable authority against treating bulk ingestion of their editorial layer as fair use. Developers whose training sets include headnotes, abstracts, case summaries or other editorial apparatus built on public domain text cannot answer a claim by pointing to the free availability of the underlying text.
The court did not decide whether training a generative model is fair use. It distinguished the Bartz and In re OpenAI rulings on the ground that the ROSS system cannot generate original expression, and said the concerns raised in the OpenAI matter do not apply here. It left open whether headnotes that reproduce opinion text verbatim can be original, treating the district court's remark on that point as dictum, and did not reach the originality of the Key Number System because ROSS forfeited that challenge.
Licentium advises AI developers, data licensors and publishers on training data rights, licensing terms and copyright exposure under United States, United Kingdom and European Union law. Work we undertake includes training data provenance reviews, drafting and negotiating content licensing agreements for model training, fair use and text and data mining risk assessments, and dispute positioning where an editorial corpus has been ingested.