Generative AI, Reproductions Inside the Model, and the Making Available to the Public

Abstract The training of generative artificial intelligence (AI) models requires the collection and analysis of a staggering amount of data, most of which consist of copyright-protected works. To date, the question whether reproductions of these works are created inside the models during their training has seldom been discussed. This is a serious blind spot in the debate given that such reproductions – e.g., inside ChatGPT’s or Stable Diffusion’s models – could be made available to end users and, therefore, to the public when AI services are offered online. Under the InfoSoc Directive, this might be copyright infringement. EU Member States’ national copyright laws would then apply and their national courts would have international jurisdiction. Seen in this light, the widely propagated narrative that non-EU AI developers are not subject to EU copyright law is an illusion.