Will AI have to prove it doesn't copy? France wants to change the rules

For several months, the training of AI models has been a subject fueling a long legal and political debate in Europe, pitting rights holders against technology companies.

In France, several senators have decided to take up the issue with a proposed law aimed at rebalancing this power dynamic, and a first step has just been taken. Indeed, as revealed by L’informé, the Council of State has just affirmed that this text is indeed compatible with the Constitution and European law.

This validation opens the way to a parliamentary debate and could profoundly change how AI companies will have to justify the use of cultural content…

A reversal of the burden of proof for AI providers

The bill, tabled in December 2025, aims to establish a presumption of use of protected works by AI systems.

Specifically, if an AI system generates a response that cites If a text is original or imitates an author's style, as was recently the case with Mistral AI, the use of a copyrighted work could be considered probable. In this case, it would no longer be up to the author to demonstrate that their work was used to train the model, but rather up to the AI provider to prove otherwise.

Today, the situation is reversed, as creators must demonstrate that their works were used to train a model. This is a particularly complex task, as technology companies generally do not publish a complete list of the data used during training. Thus, the senators speak of an information asymmetry between rights holders and AI providers.

The future system would therefore introduce a specific evidentiary regime into the Intellectual Property Code, with a so-called “rebuttable” presumption where companies could always contest the accusation by providing evidence.

The Council of State validates the legal principle

At the request of the President of the Senate, Gérard Larcher, the Council of State has therefore issued its opinion, and the highest administrative court considers that the French legislature is indeed competent to establish this type of mechanism.

According to its analysis, the proposed law does not modify European rules relating to copyright, but simply creates a specific procedural regime that allows for the establishment of proof of potential infringement, which falls under the jurisdiction of the Member States. members.

However, the court clarifies that this is a presumption of use, and not an automatic presumption of infringement, which means that the final legal classification would remain the prerogative of the judges.

A text contested by the technological ecosystem

Despite this legal green light, the proposal continues to generate numerous criticisms. Some companies, starting with Mistral AI, believe that this presumption could, in particular, weaken the competitiveness of European AI players by introducing an additional legal risk for companies developing language models.

For his part, Mistral AI CEO Arthur Mensch has advocated a different approach in the Financial Times, proposing the creation of a European levy paid by AI providers to finance a fund supporting cultural creation.

The French bill is scheduled to be debated in a public session of the Senate on April 8. The debates should clarify the scope of the system and its conditions of application, while the question of remuneration for content used by AI models remains one of the major legal challenges in the sector…