A Study on the Judicial Application of Fair Use for Artificial Intelligence Training Data in the United States

Abstract The rise of generative artificial intelligence has made the fair use determination of training data a cutting-edge controversy in copyright law. This article systematically examines the judicial application experience of U.S. federal courts in the cases of Thomson Reuters v. ROSS Intelligence, Bartz v. Anthropic, and Kadrey v. Meta. The study finds that the divergences among the three cases in the weighting of the four factors and the conclusions of their rulings are not logical contradictions, but rather reveal the flexible application mechanism of the fair use doctrine under different technological scenarios: the functional relationship between the AI product and the original work constitutes the core variable determining the finding of transformative use; the legality of the data source constitutes an independent factual basis for negating fair use; and the market impact analysis presents a typological structure ranging from direct substitution, licensing market harm, to market dilution. Keywords: Artificial intelligence training data; Fair use; Transformative use; Market dilution; Four-factor analysis framework