When Humans Train Machines and Machines Create Art: Rethinking Copyright in the Age of Generative AI

The rapid rise of generative artificial intelligence (GenAI) has transformed creative production, giving rise to a new form of human–AI co-creation in which humancreators rely on AI tools to enhance creativity, while AI models depend on vast repositories of human-generated content for training. This reciprocal relationship challenges the foundations of copyright law, raising two interdependent policy questions:whether AI-generated content should qualify for copyright protection (output regulation) and whether copyrighted human works may be lawfully used to train AI models (input regulation). Although these two debates have been widely discussed in legal and policy circles, existing research largely examines them in isolation, overlooking their dynamic interaction and the resulting welfare implications for creators, AI developers, and consumers.This project aim to fill this critical gap by developing a dynamic game-theoretical framework that integrates the input and output dimensions of AI copyright regulation into a unified analytical model. The model captures the strategic interactions among three key stakeholders—a GenAI platform, human creators, and consumers—within an evolving creative ecosystem. Specifically, the study investigates how granting copyright to AI-generated content influences the optimal level of restriction or compensation for using copyrighted human materials as training data, and vice versa. The framework aims to further identify welfare-maximizing policy combinations that balance human creative effort, AI technological progress, and social welfare.Methodologically, the project models the creative economy as an N-period dynamic game. In each period, the GenAI platform sets a training data price, creators decide their creative effort and data-sharing proportion, and consumers choose between human-and AI-generated content based on perceived quality and authenticity. The equilibrium outcomes under four policy regimes—ranging from full openness to full restriction—are compared in terms of human effort, model quality, and total welfare.The findings will offer both theoretical and policy contributions. Theoretically, the project extends classic copyright economics by incorporating the co-creative nature of GenAI and the feedback loop between input and output regulation. Practically, it provides timely policy guidance for jurisdictions such as Hong Kong, which currently adopts an AI-friendly regime allowing unrestricted training data use while granting copyright to AI outputs. By identifying the optimal balance between innovation and fairness, this research informs sustainable policy design for the global AI-driven creative economy.