The Rise of AI Training Data Licensing: Reassessing Fair Use and MarketHarm

Abstract

The unlicensed use of copyrighted works to train generative artificial intelligence (AI) systems is among the most contested questions in contemporary copyright law. Courts and scholars remain divided on how the fourth fair use factor should apply to a process spanning data ingestion, model training, and commercial deployment. This article asks how the Factor Four market harm inquiry should be structured as training data licensing markets emerge, and what this means for Kazakhstan, which has adopted an AI training opt-out rule without a developed doctrinal framework. Using doctrinal analysis of United States and European Union copyright law, a comparative method, and a critical synthesis of three scholarly positions, the article reexamines six recent decisions including Andy Warhol Foundation v. Goldsmith, Kadrey v. Meta, and Thomson Reuters v. ROSS Intelligence for what each actually decided. It finds that existing approaches apply one standard of scrutiny to a use occurring in two doctrinally distinct phases, and proposes a phase-sensitive framework: a permissive test for upstream training paired with heightened scrutiny for downstream deployment, supplemented by collective licensing and transparency obligations. Applying this framework to Kazakhstan, the article evaluates Article 23(5) of the Law of the Republic of Kazakhstan “On Artificial Intelligence” and proposes an institutional design built on existing collective-management infrastructure.