Copyright Exceptions and Limitations for AI Innovation: A Need for Tailored Reform?

The intersection of copyright law and artificial intelligence training presents a defining challenge for global innovation policy.Training large-scale generative AI models requires ingesting thousands of copyrighted works, rendering traditional licensing and rights-clearance frameworks economically infeasible.This chapter examines how major jurisdictions have responded to the question of whether and how copyright law should accommodate computational uses of protected works for AI training purposes.Through comparative analysis, it identifies three divergent regulatory approaches: the flexible but unpredictable fair use doctrine employed in the United States, statutory text and data mining (TDM) exceptions with opt-out mechanisms adopted by the European Union, and Japan's permissive framework allowing non-expressive uses without prior authorization.These jurisdictional variations function not merely as technical differences but as competitive tools that shape AI investment flows and innovation trajectories.Existing frameworks, conceptualized for different technological contexts, fit AI training poorly.No international consensus has emerged regarding fair compensation mechanisms, transparency requirements, or whether current exceptions require adaptation or wholesale replacement.This chapter proposes a tailored reform framework designed specifically for AI training contexts, addressing the idea-expression distinction, global AI competition dynamics, and the emerging "race to the middle" among copyright reformers.