Executive Summary

Is training new AI models on vast libraries of unlicensed works an act of copyright infringement? This essay examines the emerging rules governing the training of AI systems across four jurisdictions—the E.U., the U.S., Japan, and China. Many AI companies have already drawn heavily on shadow, unlicensed libraries, raising copyright risk in the absence of legal protections. This comparative study uncovers convergence on the general principle of carving an exception from copyright liability for AI training, but divergence on the mechanisms and the conditions for the exception. While all four jurisdictions seek to position themselves as innovation friendly, the precise rules present a complex terrain and reveal surprisingly high risk in the U.S..