France faces aggressive AI liability push

France is testing a much harder line on AI training data. The core idea is simple — if an AI model looks plausibly tied to protected works, the developer may have to prove those works were not used. That is a big shift from the usual copyright fight, where the claimant has to do the proving. And this week the push got louder, with 81 French organizations pressing the National Assembly to take up the bill quickly. (assemblee-nationale.fr) ### What changed this week? On May 5, a coalition of 81 French organizations from music, film, publishing, and news put out a joint call backing the proposal and urging deputies to pass it before summer. They framed it as a way to stop AI companies from building businesses on unlicensed cultural works and to force real licensing talks. (snepmusique.com)Where is the bill now? This is not just an advocacy memo. The French Senate adopted the text unanimously on April 8, 2026, and it was formally transmitted to the National Assembly on April 9. So the live question is no longer whether someone floated the idea — it is whether the Assembly schedules and advances it. (senat.fr), a copyrighted work or neighboring-rights object would be presumed to have been used by the provider of an AI model or system if an indicator tied to the system’s development, deployment, or generated output makes that use plausible. “Sauf preuve contraire” — unless the provider proves otherwise. Basically, burden first lands on the AI company. (senat.fr) ### Why is that such a big deal? Because training-data cases are usually hard to prove from the outside. Rightsholders do not see the corpus, the filtering pipeline, or the fine-tuning layers. Model developers do. The bill is trying to solve that information asymmetry by changing procedure, not by waiting for perfect transparency. That sounds narrow, but procedure is often where lawsuits are won. (laweconc([senat.fr)resume-frances-ai-copyright-shortcut/)) ### What counts as a clue? That is the sharp edge. The trigger is not direct proof of copying. It is any indicator linked to development, deployment, or outputs that makes use “vraisemblable” — plausible. So a claimant may not need a leaked dataset list. Patterns in outputs, product behavior, or other evidence around the model could be enough to shift the burden. (senat.fr) ### Does it only hit future cases? No — and that is one of the most aggressive parts. The Senate text says the rule would apply to ongoing civil proceedings once the law takes effect, except where a judgment is already final. In plain English, live cases could suddenly get easier for rightsholders to press. (senat.fr) ### How does this connect to the EU AI Act? Fr(senat.fr)at makes the EU AI Act’s copyright and transparency duties enforceable in practice. Their argument is that transparency rules mean less if rightsholders still cannot turn suspicious evidence into a workable claim. So this is France trying to add litigation teeth on top of the EU framework. (snepmusique.com)tions-se-mobilisent-pour-la-ppl-presomption/)) ### Who is pushing it? The bill was introduced in the Senate by Laure Darcos and a cross-party group of senators, then adopted unanimously. The public campaign around it is being driven by a broad rights-holder coalition — groups including SNEP and SACEM, plus organizations across cultural and information sectors. That breadth matters because it makes this look less like a music-industry niche fight and more like a national creative-industries push. (senat.fr) ### Bottom line? France is not banning AI training. It is trying to make “prove you had the rights” the default courtroom posture when the evidence points that way. If the National Assembly goes along, model developers — especially anyone relying on murky training histories — face a much tougher legal environment in France. (assemblee-nationale.fr)