AI and copyright: France flips the burden of proof
France's Senate unanimously passed a bill forcing AI companies to prove they did not use protected works in training. A legal reversal that classic copyright law never managed to achieve.
40 meetings. Zero result.
Between June and November 2025, rights holders and AI companies met five times in plenary sessions and held 40 bilateral meetings. The goal: agree on value sharing when a generative model is trained on protected works.
On April 8, 2026, the Senate passed a single-article bill unanimously. What a year of negotiation had failed to produce, a parliamentary vote settled in a few hours.
What classic copyright could not do
The problem with classic copyright is that it assumes you can identify the theft. You see your book reproduced without authorization, you sue. But the training data inside an AI model is buried in corpora of hundreds of terabytes. Nobody publishes the complete list.
Proving that an AI ingested your work is like convincing a judge that a burglar entered your home without seeing their face, without fingerprints, without cameras. You know something was taken. You cannot prove it under existing rules.
The result: since 2022, lawsuits against AI developers for unauthorized data use have been piling up, but they rarely succeed for lack of evidence. The legal terrain was asymmetric by design.
The presumption of exploitation: what it actually means
The law creates a "rebuttable presumption of exploitation." The mechanism comes down to one idea: when an AI produces output that resembles a protected work (same style, similar excerpts, characteristic similarities), courts presume that content was used during training.
The burden of proof reverses. The artist no longer has to prove the pillage. It is the AI company that must demonstrate to a judge it did not use a specific work. This is comparable to requiring a food manufacturer to prove the absence of an allergen, rather than waiting for a consumer to fall ill before the proof is their problem.
The Council of State, consulted by the Senate, validated the constitutional approach of this mechanism with minor drafting adjustments. The law applies to model developers and to systems that deploy them. It also covers proceedings already underway at the time of the vote.
The proposal is cross-party: co-signed by senators from Horizons (Laure Darcos) to the PCF (Pierre Ouzoulias), through LR, the Centrists, the PS, and the UDI. The unanimous vote is not a footnote.
What this changes for a musician, an author, a photographer
Consider a composer who discovers that a generative model reproduces their style closely. Before this law, the process was: prove your catalog appeared in the training data, obtain the model's logs (often nonexistent or confidential), convince a judge without documentation.
After the law, the frame reverses. The company developing or deploying the model must be able to document where its training data came from. If it cannot prove the non-use of a specific work, the presumption runs against it.
SACEM welcomed the vote as "an important step for creators, against the profound imbalance between AI providers and rights holders." The SNE (national publishers' association) called it "a necessary first step."
The technical argument from the opposition
Mistral AI and France Digitale (France's tech sector lobby) oppose it. Their argument: proving non-use of a particular piece of content among dozens of terabytes is technically impossible. The argument is serious, not just posturing.
But it points precisely to the structural problem. If you cannot trace what was ingested, it means the current business model rests on an acknowledged gray area. "Impossible to prove non-use" amounts to admitting the total absence of training data traceability. That is not an argument against the law: it is a description of the problem the law tries to solve.
Digital Minister Anne Le Hénanff called it a "devastating legal risk for innovation." Bercy and Matignon are holding back: protecting French champion Mistral weighs in the balance. The real fault line is not between pro-artists and pro-tech: it runs through the state itself, torn between its cultural industry and its national startup.
The real decision is still ahead
The text now goes to the National Assembly. That is where the political balance of power will become visible. The Senate, more insulated from immediate economic interests, could vote unanimously. The Assembly, under more direct influence from the executive and from Bercy, will be different terrain.
If the law passes, France would be the first country in the world to legislate in this way. Laure Darcos put it plainly: "We simply want to put a foot in the door and send a signal to operators and to Brussels." The European Union is watching: the Commission is still debating its approach, and a French precedent would create legislative pressure across the continent.
The question before the Assembly: when a player cannot document what it used to build a commercial product, should the legal void work in its favor or against it?
The Senate has answered. The Assembly must decide.
Further reading: if you want to understand how ongoing legal proceedings (including SACEM's case against several AI developers) could evolve under this new legal framework, follow the legislative process in the National Assembly over the coming weeks.
