关键词: 生成式人工智能, 模型训练, 著作权, 侵权认定

Abstract: With the rapid development of generative artificial intelligence technology, the infringement issues arising from the use of copyright-protected works in its data training process have become a focal point. The typical domestic and international cases are studied from three aspects: determination of fair use, principle of imputation, and assumption of responsibility. In terms of fair use, relevant cases in China have not involved the determination of the use of works for large model training, while US courts have shifted from supporting the “intermediate reproduction” as constituting fair use to determining infringement. In terms of principle of imputation, considering the universality of data, technical complexity, and the need to encourage innovation, the principle of fault liability should be adopted to avoid excessively burdening technology development. In terms of assumption of responsibility, platforms in data training scenarios can be broadly interpreted as new types of “network service providers”, and the “safe harbor principle” can be analogically applied. Infringement can be determined through the “substantive similarity + likelihood of contact” standard and algorithm comparison technology, while referencing current mainstream theories from various countries, aiming to provide reference for the improvement of relevant legal regulations and industrial development.

Key words: generative artificial intelligence, model training, copyright, determination of infringement