France. The National Edition Union (SNE), the Société des Gens de Lettres (SGDL) and the National Syndicate of Authors and Composers (SNAC) introduced, on March 6, 2025, before the judicial court, a legal action against Meta Platforms Inc. They support that Meta would have proceeded to use, without their authorization, of works protected by copyright for the training of its artificial intelligence model (IA) “Llama” a chatgpt competitor or Mistral AI cat.
More specifically, authors and publishers notably argue that the exploitation of the works of the mind that they have created or published to fuel a generative AI would constitute an act of reproduction subject to the exclusive monopoly of the author and that, for lack of authorization, Meta would have engaged in acts of counterfeiting (and parasitic actions). According to them, the AI ”Llama” was indeed drawn to digital books from a “books3” database containing nearly 200,000 pirated pounds, including many French titles.
Without prejudging Meta’s position, the multinational could oppose several arguments on the basis of copyright. First, Meta could contest the very application of copyright on the ground that reproduction consists, according to article L. 122-3 of the Intellectual Property Code (ICC), in a material fixing of the work by all processes “Which allow it to be communicated to the public in an indirect manner”. However, it is by no means obvious that there is, downstream, in terms of the results generated, a communication to the public in an indirect manner, especially when the generated content does not allow to recognize the creations used. Especially since the latter are rarely restored in a recognizable form. Unrecognizable, the work could therefore not be considered as “Communicated to the public”. Reproduction would therefore not be an act of exploitation, but a simple technical act, escaping the author’s monopoly.
As attractive, this interpretation seems difficult to remember. First, national law must be interpreted in the light of the text and the purpose of European Union law; However, in this case, European law is less restrictive than French law, article 2 of Directive 2001/29 not posing this condition of “Communication to the public in an indirect manner”. Then, the aforementioned analysis only retains the act of reproduction if, downstream, at the results stage, it is possible to identify the content that made it possible to generate a result. But to be further upstream, at the training stage of generative AI models, it seems to be a little doubt that the use of works of the mind to cause a model implies their reproduction in the aforementioned sense, at least temporary. This is the reason why EU regulation on artificial intelligence requires providers of generative AI models to make available to the public “A sufficiently detailed summary of the content used to train models” And this, in particular in order to “ Respect Union’s law on copyright ”.
In the event that copyright finds it to be applied, META could still try to take shelter behind certain exceptions, including in particular that of excavation of texts and data provided for in article L. 122-5-3, III of the ICC which allows free and free copies on the Internet. Its application is however subject, except for purposes of scientific research in this case inapplicable, to the absence of opposition (or “Opt out”) rights holders. The text adds the condition “A lawful access” to digital copies or reproductions of works and specifies that these copies “ are stored with an appropriate level of security and then destroyed after the excavation of texts and data ”. Assuming that Meta had lawfully access to the works and that the authors did not exercise their right of opposition before their reproduction, the second condition relating to the necessary destruction of copies could constitute an insurmountable obstacle. In any event and supposing the applicable exception, it will still have to pass the three-step test which subjects all the exceptions to the obligation to be limited to special cases, not to infringe the normal exploitation of the work and not to cause unjustified damage to the author’s legitimate interests. Which seems difficult.
From there to say that Meta will be condemned, there is a step that cannot be taken. First for reasons of proof: AI are often qualified as “black box” to signify that it is difficult to grasp their operation, but also to identify the works of the mind in question, used as a training, among the billions of data. Although real, these probationary difficulties can however be resolved both technically, but also procedurally. Then for applicable jurisdiction and law questions; But here again nothing directing, the supranational texts fixing well -known principles of lawyers.
