The U.S. Court of Appeals for the Third Circuit ruled on September 29 that ROSS Intelligence’s use of copyrighted legal headnotes to train a non-generative legal-research tool was not fair use. The decision in Thomson Reuters v. ROSS Intelligence affirmed copyright protection for the headnotes and recognized a potential market for licensing them as AI training material.

The ruling gives weight to training-data licensing in the assessment of copyright market harm, even when the copyright owner has not licensed the material to outside developers. Its scope is nevertheless tied to the use before the court: a legal-research system that did not generate original expression, rather than a generative AI model.

Headnotes are editorial summaries of legal points in judicial decisions. Although copyright does not protect legal principles themselves, original expression used to summarize or explain them can qualify for protection. The dispute concerns that editorial material, rather than ownership of the underlying law.

Fair use requires courts to consider four factors: the purpose and character of the use, the nature of the copyrighted work, the amount taken, and the effect on the work’s potential market or value. In evaluating market effects, the Third Circuit found harm both to the existing market and to the value of the headnotes. It also treated licensing those materials for AI training as a potential derivative market.

Evidence that the copyright owner used its headnotes to train its own AI search products supported that assessment. The court did not regard the absence of third-party licenses as sufficient to rule out a licensing market. The burden of establishing fair use remained with the party invoking the defense.

The decision takes a different approach from two June 2025 rulings in the Northern District of California concerning books used for AI training. Judge William Alsup accepted that a training-data licensing market might emerge but concluded that copyright law did not necessarily reserve that market to authors. Judge Vince Chhabria likewise rejected reliance on lost licensing revenue for a transformative use, reasoning that doing otherwise would make the analysis circular: an owner could defeat fair use simply by demanding payment for the disputed activity.

The Third Circuit did not resolve that circularity question. It had already found that ROSS’s use involved little, if any, transformation. That distinction limits direct comparisons with cases involving uses that courts consider substantially transformative. The rulings do not establish a uniform answer for all AI training disputes.

A separate issue is how courts should evaluate licensing evidence when they consider it relevant. In its analysis of the decision, Baker Botts identified possible parallels with patent royalty litigation, where courts scrutinize whether existing agreements genuinely match the technology and rights at issue. A bundled agreement may cover services or access beyond training rights, while a settlement license may reflect litigation pressures as well as commercial value.

Those comparisons are not a legal test adopted in the ROSS ruling. They point instead to unresolved evidentiary questions, including how much weight a court should give particular agreements and whether expert testimony accurately describes their terms. The appellate decision recognizes a potential training-data market; how parties prove that market’s boundaries and value remains a separate question.

Sources: Baker Botts