合理使用视角下生成式人工智能训练数据著作权风险与制度完善

Copyright Risks and Institutional Perfection of Generative AI Training Data from the Perspective of Fair Use

摘要: 生成式人工智能技术快速发展，训练数据著作权合规风险日益凸显，其数据驱动运行模式与现行著作权法律规范产生结构性冲突。针对人工智能数据输入、数据分析、内容生成全流程存在的著作权侵权争议，结合我国封闭式立法模式下合理使用规则难以适配大规模模型训练的现实困境，区分表达性行为与非表达性行为展开规范分析，尝试搭建“非享受性使用”与“转换性使用”联动的动态评估机制。研究发现，现有合理使用条款受制于适当引用、少量复制等适用条件，无法覆盖人工智能深度学习场景。依托现有立法修订路径，可在《著作权法实施条例》中增设人工智能深度学习对应的合理使用情形，并配套建立以使用性质与转换程度为核心的反向校验规则，为平衡技术发展与著作权人权益提供规范思路。

Abstract: With the rapid advancement of generative artificial intelligence (GenAI), copyright compliance risks concerning training data have become increasingly prominent. Its data-driven operation model creates structural conflicts with the existing copyright legal norms. In view of copyright infringement disputes arising throughout the whole process of artificial intelligence data input, data analysis and content generation, this paper conducts normative analysis by distinguishing expressive acts from non-expressive acts, considering the practical dilemma that fair use rules under China’s closed legislative model can hardly adapt to large-scale model training. It attempts to construct a dynamic evaluation mechanism combining “non-consumptive use” and “transformative use”. The study finds that the existing fair use provisions, constrained by application prerequisites such as appropriate quotation and partial reproduction, fail to cover the deep learning scenarios of artificial intelligence. Relying on the path of revising current legislation, a fair use circumstance corresponding to AI deep learning may be added to the Regulations for the Implementation of the Copyright Law of the People’s Republic of China, with supporting reverse verification rules centered on the nature of use and the degree of transformation. This provides normative ideas for balancing technological development and the rights and interests of copyright holders.