Examining artificial intelligence training as fair use: a comparative law reflection on Bartz v. Anthropic PBC

Traditional legal doctrinal analysis easily regards the use of copyrighted works in artificial intelligence (AI) training as infringement; however, in substance this conclusion goes against the purpose of copyright law to balance interests. In Bartz v. Anthropic PBC , a U.S. federal district court held for the first time that the use of copyrighted books to train large language models is fair use, but did not accept fair use for the pirated copies acquired to build a permanent central library. Transformative use forms the core of this holding; training aims to build a technical tool, its results are transformative and differ in value from the original works, and no market substitution follows. The class action settlement reached later only concerned the acquisition of pirated copies; thus, the reasoning still stands. From the perspective of comparative law and legal interpretation, the decision offers a reference for China; Article 24 of the Copyright Law of the People's Republic of China lists fair use in a closed enumeration. It is better to give this provision a limited expansive interpretation. By means of teleological expansion, AI training can be included in the scope of "illustrating a point" under Article 24(1)(2) and treated by analogy as a special kind of illustrative use; its object is the linguistic patterns, statistical regularities, and knowledge structures in data, and three restrictive requirements apply. Then the degree of transformation serves as an auxiliary criterion and works inside the three-step test as a residual check.