The rapid deployment of artificial intelligence systems capable of generating text, images, music, and code has intensified legal and policy debates surrounding the use of copyrighted works as training data. At a broad level, existing copyright frameworks were not designed to address large-scale, automated ingestion of protected content for machine learning purposes, resulting in regulatory uncertainty, fragmented litigation, and tension between rights holders and AI developers. Current approaches ranging from unlicensed use claims to narrow exceptions struggle to balance innovation incentives with the protection of creative labor. This article proposes collective licensing as a regulatory design solution for governing the use of copyrighted works in AI training. Drawing on established models from music, broadcasting, and reprographic rights, the analysis situates collective licensing as a middle-ground mechanism capable of reducing transaction costs, ensuring remuneration, and providing legal certainty at scale. The article examines the structural limitations of individual licensing and fair use–based approaches when applied to high-volume, non-expressive uses inherent in machine learning workflows. Narrowing its focus, the article outlines a conceptual framework for collective licensing tailored to AI training contexts. It explores institutional design considerations, including rights aggregation, scope of licensed uses, opt-out mechanisms, remuneration models, transparency obligations, and governance safeguards. Comparative insights from existing collective management organizations inform the proposal, highlighting how regulatory oversight and standardized tariffs could mitigate market power imbalances while supporting technological development. The article argues that a well-designed collective licensing regime can align copyright law with data-driven innovation by preserving incentives for creators, enabling lawful AI development, and reducing adversarial enforcement dynamics. By reframing AI training as a compensable, regulated use rather than an ungoverned exception, the proposal offers a pragmatic pathway for modernizing copyright law in the era of large-scale artificial intelligence.