Abstract:
This article explores whether generative AI image models, which employ latent-space diffusion techniques, infringe upon the copyright of works used as training data. A doctrinal analysis based on US principles of transformative use and substantial similarity observed that the core technical process of latent-space generation does not typically replicate copyrighted expression. A comparative legislative analysis of text and data mining (“TDM”) exceptions in Japan, Singapore, and the European Union highlights their growing importance in facilitating data-driven innovation. The findings support adopting a dedicated TDM exception in South Korea to provide legal certainty and promote AI innovation.