France may flip the burden of proof on AI training: SACEM, its opt-out and the Darcos Bill

In a release reported on 8 June 2026, SACEM, France's songwriters' society, renewed its call for MPs to pass the Darcos Bill. It's a one-article bill that the Senate passed unanimously in April. It wouldn't create any new right. It would change who has to prove what when a rights holder takes an AI company to court, and for musicians that's the part that has been missing.

What would the Darcos Bill actually change?

Senator Laure Darcos and colleagues filed the bill on 12 December 2025, and the Senate adopted it on 8 April 2026. The adopted text adds a new article, L. 331-4-1, to France's Intellectual Property Code. In any civil dispute, a work protected by copyright or a related right is presumed to have been used by the provider of an AI model or system, unless the provider proves otherwise. The presumption applies when a sign linked to the system's development, deployment or output makes that use plausible. It would also apply to cases already in progress, except those that have reached a final decision.

The Senate's own summary lists what could count as that sign: outputs "in the manner of" or "in the style of" an author, an AI regurgitating part or all of a work, resemblance between a work and a generated result, or evidence from technical analysis of the model or its data supply chain. The aim, in the summary's words (our translation), is to "restore equality of arms" in litigation.

The Conseil d'État reviewed the bill at the Senate's request. In its 19 March 2026 opinion it found no constitutional or treaty obstacle, subject to some adjustments to the drafting. The Senate's culture committee then amended the text to take account of that opinion. One amendment swapped "exploited" for "used", because the reproduction right is a matter for EU law, not French law. So this is a rule about evidence, not a new copyright.

Where does SACEM's opt-out fit in?

French law lets AI developers mine works they accessed lawfully, unless the rights holder has opted out (article L. 122-5-3 III of the code, which implements the EU's text-and-data-mining exception). SACEM says it opted out in October 2023 for its members' whole repertoire, so members don't need to do it themselves. Any AI provider that wants to train on those works needs SACEM's prior authorisation and has to negotiate the terms, including payment.

The problem, as the Senate summary puts it, is that the opt-out has had very limited effect in practice. Rights holders don't know whether their works were used, and AI providers don't always honour opt-outs. In our reading, the two measures depend on each other. The opt-out decides when mining is unlawful. The presumption decides who has to prove that mining happened at all.

What changes for an artist in a dispute?

Right now the person suing has to prove that their work went into a model, and only the AI company holds the training records. The Munich case GEMA largely won against Suno shows what that takes. The judgment records 3,779 prompts run by GEMA, 338 of them about the six songs at issue, including 176 for "Atemlos durch die Nacht" alone. We covered the ruling in Munich court: a model that memorises songs is copying them.

Under the Darcos Bill, a songwriter or their society would show a plausible sign, such as an output that closely tracks their song or imitates their style. The AI provider would then have to prove it didn't use the work. Note what the bill doesn't do: the text presumes use, not infringement. We expect a provider could still argue that the use was licensed or otherwise lawful. That's why SACEM's opt-out matters so much here: it closes off the "lawful mining" defence for SACEM members' works, at least for use after October 2023.

Why is it stuck?

The Senate sent the bill to the National Assembly on 9 April as bill no. 2634, and the culture committee filed its report and text on 2 June (Assemblée nationale). NosParlementaires reports that the committee approved it 29 votes to 8. Between 29 May and 8 June, MPs filed 110 amendments to its single article, 100 of them from the presidential camp's Ensemble pour la République group. On 11 June the bill sat last on a day reserved for the GDR group's business and was never reached.

Livres Hebdo reported on 3 July that Renaissance MP Paul Midy objected to what he called a "presumption of guilt" for AI providers. Digital minister Anne Le Hénanff warned of more litigation and pointed to future revisions of the EU framework as the appropriate level. Several MPs described intense lobbying by tech companies, notably Mistral AI. NosParlementaires reported on 6 August that the bill had never been put to a vote in the chamber, and that the amendments were still waiting for it to be scheduled again. The Assembly's file lists no plenary sitting.

The bill isn't only a French cause, though. CISAC, which has 227 member societies in 111 countries, held its centenary general assembly in Paris on 4 June. Society representatives held up signs saying that creators around the world support the Darcos Law, according to Digital Music News. SACEM called unlicensed AI training "the greatest plundering of creative and artistic works ever perpetrated."

What we think

The Darcos Bill goes after the right problem. Europe already gave rights holders an opt-out. What it didn't give them was a way to find out whether anyone honoured it. An opt-out you can't check protects you only as far as AI companies choose to respect it. Making the provider prove non-use puts the burden on the side that holds the evidence.

The less obvious effect is on AI companies' own record-keeping. Under the bill, the only way to rebut the presumption is to show what you trained on. Training-data records stop being a compliance chore and become a company's legal defence. The Senate summary makes this argument directly: providers that document their sources and license their data get an advantage over those that scrape. If the bill passes, licensed developers win and undocumented scrapers lose. Each new pending case raises the cost of having no records.

The critics have a point, though not the one they lead with. "Presumption of guilt" overstates it: the presumption can be rebutted and needs a plausible trigger first. The real risks are elsewhere. A loose reading of "plausible" could invite speculative claims, and a French-only rule raises the question the minister asked about whether this belongs at EU level. We'd note that the Senate summary itself points to a European Parliament resolution of 10 March 2026 that mentions the same idea.

For artists, the honest caveat is that few individual musicians will ever sue an AI company. GEMA's Munich case needed a collecting society's staff and lawyers. The bill's value will mostly reach creators through SACEM and other societies, so the practical steps sit with you:

• Check your mandate. Confirm your works are registered with a society that holds and exercises an AI opt-out.
• Keep dated evidence. Save prompts, outputs and dates whenever a tool produces something that tracks your music. Under this bill, that's exactly the kind of sign that shifts the burden.
• Ask where the money goes. If the presumption pushes AI firms into licensing, ask your society how that licence income will be shared with members and reported.