INTRODUCTION 263

I. FAIR USE IN THE DIGITAL AGE 269
A. Historical Development and Purpose 269
B. Adapting Fair Use to Emerging Technologies 273

II. AI TRAINING CASES: RECENT DECISIONS 281
A. Introduction to the Cases 281
B. Generative AI and the Structural Tensions of Fair Use 285
1. The Breakdown between Factors One and Four 286
2. The Limits of the Human-Learning Analogy 291
3. Public Benefit, Market Failure, and Institutional Limits 294

CONCLUSION 298

This Article examines the first federal decisions addressing whether the use of copyrighted books to train large language models constitutes fair use: Kadrey v. Meta and Bartz v. Anthropic. Both courts held that training on lawfully acquired works was fair use, while diverging in their treatment of pirated materials and in aspects of their market-harm analysis. Using these cases as doctrinal case studies, the Article argues that generative AI exposes structural tensions within fair use doctrine. In particular, AI training unsettles the traditional relationship between transformativeness under the first factor and market substitution under the fourth. It also places pressure on the idea/expression distinction, complicates the assessment of emerging licensing markets, and brings renewed attention to public-benefit considerations in technology cases. Beyond substitution, the Article identifies a distinct “value-extraction” concern: AI systems may appropriate expressive value from copyrighted works without replenishing the incentive structure copyright is designed to sustain. The Article’s primary contribution is to identify and analyze these tensions, while offering some suggestions for how courts might approach them in future cases.