On 19 November 2026, France’s National Assembly is due to examine the so-called “Darcos” bill, which establishes a presumption that protected works have been used by artificial intelligence providers. Passed unanimously by the Senate in the spring and scheduled on the day reserved for the Socialist group, it pits the cultural sector squarely against Mistral AI and the digital industry. For companies that develop or use AI, what is at stake is the burden of proof in the event of a dispute.
A single article that reverses the burden of proof
The text was tabled in December 2025 by senators Laure Darcos, Agnès Evren and Pierre Ouzoulias, then adopted by the Senate on 8 April 2026 almost unchanged, Silicon.fr recalls. Its principle fits in one sentence: in a civil case, a protected work would be presumed to have been used by the provider of an AI model or system as soon as there is evidence making such use plausible.
In practical terms, as Next explains, the rights holder must still provide initial evidence, but it is then up to the developer to show that it did not use the work. Under current law, an author, publisher or media outlet that suspects a model has been trained on its content runs up against the opacity of datasets. The presumption is designed precisely to get around this wall.
Why the debate slipped from June to November
The text was due to be debated in the chamber in June. According to Next, around a hundred amendments tabled by members of the Ensemble group prevented the debate from concluding; Cécile Rap-Veber, managing director of Sacem, counts 110 of them and speaks of obstruction. At the end of October, the Socialists announced that they would place the bill on their parliamentary opposition day.
Among the proposed changes, MP Éric Bothorel wanted to replace the word “system” with “model”, which would have cleared downstream players, such as companies that integrate a model into their tools. According to Silicon.fr, the committee rejected this amendment, as well as the one calling for the article to be deleted outright.
Mistral and AFNUM warn of a litigation risk
On 8 November, Arthur Mensch, head of Mistral AI, came out publicly against the text. He considers the objective “legitimate” but believes, according to Next, that it is “impossible” to prove non-use without revealing one’s methods, and therefore one’s trade secrets. Instead, he proposes a levy calculated on the revenue of all AI providers active in Europe, foreign ones included, which would feed a European fund for creative work.
AFNUM, which represents Google, Nvidia and Microsoft among others, published a statement on 9 November entitled “Negotiate rather than litigate”. The organisation argues that:
- lawsuits could target the entire chain, from developers to professional users, whatever their actual role;
- start-ups would have to bear high legal costs, at the risk of jeopardising their fundraising;
- some models intended for the French market might no longer be offered there;
- a national law would fragment the single market, at a time when the Commission is consulting on the copyright directive.
It puts a figure on the risk: if a quarter of the investments pledged at the AI summit were postponed or relocated, €27 billion in direct investment and more than 40,000 jobs over ten years would be lost. This figure is AFNUM’s own estimate.
The cultural sector and the government
Laure Darcos retorted that trade secrets cannot take precedence over property rights and that judges know how to protect confidentiality. According to Next, she points out that more than 160 lawsuits over training on protected content are under way in the United States. Sacem, SACD and the Syndicat national de l’édition (the French publishers’ association) co-signed an op-ed asserting that France does not have to choose between technological power and cultural power. During the Senate debate, Next recalls, Culture Minister Catherine Pégard and Anne Le Hénanff, Minister Delegate for AI and Digital Affairs, had argued for negotiation rather than litigation and pointed to changes in the European framework.
Finally, Next, citing Le Monde, mentions the possibility of a compromise amendment that would allow adoption while sparing Mistral. Nothing had been settled as of 12 November, a week before the session.
Frequently asked questions
Does the Darcos law already apply?
No. The Senate adopted it in April 2026, but the National Assembly has yet to examine it, in public session on 19 November. According to Next, this day reserved for the Socialists may well be the last opportunity to put the text before MPs during the current parliamentary term.
Could an SME using an AI tool be affected?
The text targets the provider of the “AI model or system”. AFNUM believes that professional users could be sued; it was precisely to exclude downstream players that MP Éric Bothorel wanted to restrict the text to models alone, without success in committee.
What does Mistral AI propose instead?
Arthur Mensch advocates a levy proportional to the revenue of all AI providers present in Europe, paid into a European fund for creative work. He presents it as a complement to licensing agreements, not a replacement for them.
Sources
- AI and copyright: Mistral and the digital industry step up their lobbying — Next, 2026-11-09
- AI and copyright: why tensions are rising in France — Silicon.fr, 2026-11-10
- AI and copyright: Negotiate rather than litigate — AFNUM, 2026-11-09