French Report on the Applicable Law to Generative AI Models

On 15 December 2025, a French Report on the Applicable Law to Generative AI Models Available in the European Union has been published under the auspices of the advisory body to the French Minister of Culture on intellectual property (CSLPA).

Tristan Azzi (University of Paris 1 Panthéon-Sorbonne, Sorbonne Law School) and Yves El Hage (University of Jean Moulin Lyon 3), the authors of the report, were tasked with working on the interplay between EU conflict-of-laws in the field of copyright and the recently adopted AI Act. The objective was to respond to a highly topical question: Do EU copyright law apply to non-European GenAI models in the EU jurisdiction?

Background

GenAI systems such as ChatGPT, Claude or Le Chat are based on AI models that are trained using vast amounts of data (i.e. training data) which may include copyrighted works. Users of these systems could also incorporate copyrighted works in their prompts. Finally, the systems’ outputs may as well reproduce parts of copyrighted works. All of these circumstances arise during the AI input and output phases within an international context involving several legal systems. This raises a crucial PIL question: which law applies to any potential copyright infringement? (see already here on the blog, and here for an ongoing case report).

This is exactly the legal research that the authors of the report have carried out, and the findings are extremely interesting, detailed and well-documented.

Main Findings

The authors argue that, at the AI input stage, the lex loci protectionis should be identified not by reference to the technical location of AI providers’ servers, but by treating AI training and output generation as a single complex process whose downstream effects determine the location of the infringing act. Therefore, this process could point to the law of the country in which the AI model is offered and used as a service. While not enacting formal conflict-of-laws rules, the AI Act reinforces this analysis, based on its Recital 106 which requires providers placing general-purpose AI models on the EU market to comply with EU copyright law.

As regards copyrights’ infringements at the output stage, the authors explain that the situation is comparable to that of a classic infringement disseminated via a website. Thus, the lex loci protectionis should be determined according to the location of a website’s targeted audience.

Finally, regardless of the lex loci protectionis and where the training operations took place, the AI Act could allow French courts to apply EU copyright law, as soon as the AI model is placed on the market on its territory, based on PIL methodologies of overriding mandatory provisions and the public policy exception.

The English Executive Summary of the Report as well as the full Report (in French) are available here.