A federal appeals court has determined that using copyrighted editorial summaries to train a competing artificial intelligence product constitutes infringement, rejecting the argument that machine learning processes automatically qualify as fair use. The U.S. Court of Appeals for the Third Circuit ruled that Ross Intelligence violated Thomson Reuters’ copyrights by utilizing Westlaw headnotes to develop a legal research platform designed to replicate Westlaw’s core functionality.
The decision clarifies a critical distinction for the generative AI sector: while underlying judicial opinions are public domain, the editorial infrastructure built around them is not. Westlaw headnotes are concise summaries drafted by attorney-editors to identify specific legal issues within court rulings. These annotations require independent judgment regarding which legal points to extract, how to phrase them accurately, and what factual context is necessary for the summary to stand alone. The court found this human curation meets the originality threshold for copyright protection, distinguishing it from the raw government text of the opinions themselves.
Ross Intelligence had argued its use was transformative because the headnotes served as training data rather than content displayed to end users. The panel disagreed, characterizing the ingestion of copyrighted material as merely an intermediate step toward creating a direct market substitute. Because Ross aimed to answer the same legal research questions as Westlaw using Westlaw’s own editorial labor, the court deemed the use “minimally transformative at best.” The ruling emphasizes that technical differences in how content is processed do not override commercial realities when the end product targets the same customer base with the same value proposition.
This finding significantly impacts the fair use analysis regarding market harm. Beyond the immediate competition between legal research platforms, the court recognized a potential licensing market for AI training data. By bypassing licensing negotiations to build a rival service, Ross undermined both the existing market for Westlaw subscriptions and the emerging market for authorized dataset licensing. This aspect of the ruling signals that courts may increasingly view unlicensed training on proprietary datasets as economic displacement rather than technological innovation.
Legal experts note that this precedent is fact-specific but highly instructive. It does not establish a blanket prohibition on AI training with copyrighted works, nor does it resolve disputes involving non-commercial research or fundamentally different use cases. However, it establishes that direct competitors cannot rely on fair use to appropriate curated, expressive content simply because they employ novel technology. Developers working with public records must now carefully separate unprotected primary sources from protected secondary analysis, metadata, and organizational structures created by publishers.
Thomson Reuters filed suit against Ross Intelligence in 2020, alleging the startup had induced a third-party contractor to create training materials derived from Westlaw. The Third Circuit’s ruling affirms liability based on the undisputed evidence that Ross used these specific headnotes to benefit its commercial platform.





