In a technical and legal study published a few days ago, two researchers conclude that the use of copyright protected works to train generative AI models must be authorized by the rights holders, and that without such authorization, generative AI developers are acting illegally. What are the obligations of generative AI developers towards the copyright holders of works they use to train their models ? The question is simple, but the answer is not obvious. The European Commission considers that the « text and data mining » exception (TDM exception) introduced by article 4 of the 2019 Copyright Directive does benefit generative AI developers, and that consequently, to ensure that developers do not use their works, right holders must, in accordance with the Directive, make a « reservation of rights ». This interpretation, if…
