Table of contents
The conference of the eleven group presidents of the French National Assembly did not retain, on the morning of May 12, 2026, the Darcos bill on copyright and artificial intelligence for the bipartisan week in June. The central mechanism of the text, a future article L. 331-4-1 of the Intellectual Property Code establishing a rebuttable presumption of use of protected content by AI system providers, remains in suspense. It is now up to rights holders to rely on Article 53 of Regulation (EU) 2024/1689, applicable since August 2, 2025, which imposes transparency on training datasets without reversing the burden of proof.
The Essentials
- Fact: The bill submitted on December 12, 2025, by Senator Laure Darcos (Les Indépendants - République et Territoires group in the Senate, affiliated with the Horizons party, elected from Essonne) was unanimously adopted in the Senate on April 8, 2026 (text no. 85), transmitted to the National Assembly on April 9 under number 2634, then removed from the bipartisan calendar for June by the conference of group presidents.
- Implication: Article 53(1)(c) of Regulation (EU) 2024/1689 requires providers of general-purpose AI models to "implement a policy to comply with Union law on copyright and related rights, including by state-of-the-art technologies, a reservation of rights expressed in accordance with Article 4(3) of Directive (EU) 2019/790" - thus, it is a preemptive transparency obligation, without a presumption of use benefiting rights holders, who remain, as it stands, solely responsible for bearing the burden of proof.
- Nuance: This is a political rejection, not a legal one. In its opinion of March 19, 2026, the Council of State judged the mechanism compliant with the Constitution and European law, confirming that the national legislator retains the freedom to institute a specific evidentiary mechanism under the procedural autonomy of member states.
A Reversed Evidentiary Mechanism, Blocked on the Eve of Examination
The core of the text adopted in the Senate consists of inserting into the Intellectual Property Code a future article L. 331-4-1, analyzed in detail by the Schmitt law firm, which establishes a rebuttable presumption of use: once an indication related to the development, deployment, or outcome of an AI system makes the use of a protected work plausible, it is up to the provider of the model or system to provide contrary evidence (text adopted in the first reading, text no. 85, Senate, April 8, 2026). This construction reverses the ordinary evidentiary regime: the rights holder who detects an indication of use, an entire passage of a novel reproduced by a model, an editorial layout copied, no longer has to positively demonstrate the capture.
The Council of State's opinion of March 19, 2026 validates the mechanism after adjustments to the material scope - distinction between providers and deployers integrated into the text transmitted to the Assembly. The opinion concludes compliance with the Constitution and European law. "The Council of State," notes the analysis of the Schmitt firm dedicated to the opinion of March 19, 2026, "recognizes the legitimacy of the approach: correcting the structural information asymmetry that places rights holders in the practical impossibility of proving the use of their works in AI model training processes - whose data remains opaque, with companies opposing business secrecy protection." The rejection on May 12, 2026, therefore, does not stem from a jurisdictional lock; it relates to the parliamentary calendar and the political arbitration conducted by the conference of presidents.
The weeks preceding the vote saw structured lobbying, documented by Télérama, which notably describes an argument distributed to parliamentary groups and the suggestion to amend the text to limit its effect to quality cultural content. According to Pascal Rogard, CEO of SACD - a collective management society directly involved in the case - cited by Le Figaro on May 11, 2026, Arthur Mensch (Mistral AI) and Yann LeCun, Executive Chairman of AMI Labs (Advanced Machine Intelligence Labs) since his departure from Meta, were among the tech personalities who met with group presidents to oppose the text - an allegation emanating from an organization party to the dispute, not publicly confirmed by the interested parties as of May 12, 2026.
The rejection on May 12 is part of a longer sequence than the senatorial chronology suggests. The PPL was submitted following a consultation conducted in 2025 between AI stakeholders and cultural industries, whose rights holders have publicly noted the lack of outcome. On the rights holders' side, Cécile Rap-Veber, managing director of Sacem, warns in her public statements about the urgency for the creative sector and calls for the conclusion of fair exploitation agreements with AI laboratories. It is from this exhaustion of the amicable route that the legislative strategy was born.
The mobilization around the text reflects this shift. As reported by Le Monde on May 7, 2026, a coalition of 81 cultural and media organizations was formed to defend the passage in the Assembly, accompanied by a petition of 25,000 signatures from creative professionals. Bercy did not support an accelerated procedure, a position that stakeholders interpret as a desire not to send a negative signal to the French AI ecosystem.
The parliamentary arithmetic of the bipartisan week did the rest: according to the elements gathered by Décideurs Juridiques, the conference of the eleven group presidents prioritized other texts scheduled for the June slot. The details of the internal arbitrations have not been the subject of official communication. The tight schedule before the presidential election adds a constraint: according to Télérama, the legislative bottleneck in the fall - with the budget vote at the forefront - leaves few windows for a text placed at the bottom of the pile.
What Article 53 AI Act Does and Does Not Do
Regulation (EU) 2024/1689 locked in a transparency regime for general-purpose model providers, without affecting the evidentiary regime. Article 53(1)(c) requires to "put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790", and Article 53(1)(d) requires the publication of a sufficiently detailed summary of the content used for training, according to a template provided by the AI Office. The logic is therefore preemptive: document, identify opt-outs, publish - not reverse the burden of proof of use to the benefit of the holder.
The application schedule is now fixed. As reminded by the AI Act Service Desk, the obligations of Article 53 have been applicable since August 2, 2025, for new models, while the AI Office's sanctioning powers - up to 15 million euros or 3% of global turnover according to Article 101 - only come into effect on August 2, 2026. Models already on the market before August 2025 have an additional deadline until August 2, 2027, to comply. Thus, in practice, a window of twelve to twenty-four months during which the AI Office can oversee and request corrections, but cannot yet impose fines.
The GPAI Code of Practice published on July 10, 2025 extends the system on a voluntary basis. Its Copyright chapter provides a complaint mechanism and a point of contact for rights holders, as well as a real-time transparency obligation on indexing robots. Doctrinal analyses published in 2025 in the IIC (International Review of Intellectual Property and Competition Law) journal note, however, that these obligations remain voluntary for signatories of the GPAI Code and do not themselves create a probative right of action enforceable against non-signatories.
Internationally, two opposing trajectories highlight the solitude of the French choice: the United Kingdom abandoned in March 2026 its broad text and data mining exception project, without considering a reversal of the burden of proof, while in Japan Article 30-4 of the Copyright Act has since 2018 enshrined a default permission to use works for training - the opposite logic of the Darcos PPL. At the European level, Article 53 AI Act sets a documentary obligation upstream of litigation; the future article L. 331-4-1 CPI would have redefined the evidentiary mechanics downstream, in the civil procedure itself. Two distinct regimes, which could have coexisted.
Three Operational Readings of the Evidentiary Gap
For rights holders, the immediate effect of the non-enrollment is mechanical: without a rebuttable presumption, they retain the full burden of positively demonstrating use. However, this demonstration requires access to training datasets that neither the summary provided by Article 53(1)(d) AI Act, nor the technical documentation of Annex XI guarantee in a granularity useful to the judge. Indirect clues - almost exact textual reproduction of passages of works, layout reproduced by an image generation model - can form the basis of a submission, but their probative value remains arbitrated on a case-by-case basis. The gap between the presumption that the future article L. 331-4-1 CPI would have instituted and the common law regime explains why SACD, Sacem, and the coalition of 81 cultural organizations supported this text with unusual intensity.
For GPAI providers, the May 12 arbitration leaves intact a schedule they know. The compliance policy provided for by Article 53(1)(c) must be in place for new models, the publication of a detailed summary according to the AI Office template as well. The GPAI Code of Practice offers a presumption of compliance to signatories; it does not dispense with an effective copyright policy. The 2026-2027 period constitutes mainly an observation horizon: the AI Office can request information, demand corrective measures, but the triggering of fines on August 2, 2026, only concerns new models, previous models have until August 2, 2027, to comply. The imposed transparency creates an auditable administrative trace, without automatically leading to a probative right of action in national jurisdiction.
For deployers in France - publishers integrating a model into a public service, press companies feeding their editorial tools, cultural platforms - the boundary remains blurred. The Council of State had precisely requested an adjustment on the scope between provider and deployer, integrated into the text transmitted to the Assembly. Without this adopted text, the qualification remains that of the European regulation, which distinguishes the model provider and the system deployer, but without explicit articulation with national civil liability regimes. A transitional asymmetry opens between August 2026 and August 2027: the AI Office's sanctions are activatable on new GPAI models, not yet on previous models in the process of compliance. Rights holders could, in this interval, attempt actions before French civil courts based solely on Article 53 AI Act and Directive 2019/790, without a presumption in their favor. The litigation will create doctrine, in the absence of a text.
One question remains open after May 12: will France seize the procedural autonomy that the Council of State recognized on March 19, 2026, before the full sanctioning competence of the AI Office renders the debate moot? The window is narrow. August 2, 2027, marks the compliance deadline for models prior to August 2, 2025, under the full sanctioning competence of the AI Office. If a new parliamentary initiative - resumption of the text by government means, niche of another group, complementary senatorial submission - is not undertaken by the fall of 2026, the national route will have been deactivated by default, and the evidentiary trajectory of French rights holders will be arbitrated by civil jurisprudence based solely on Article 53.
The lock holds in a positive law distinction: Article 53 of Regulation (EU) 2024/1689 organizes a preemptive transparency obligation, not an ex-post evidentiary regime. It is the same topology that separates, for five years, the information obligations of the data controller (articles 13 and 14 GDPR) from the effectively enforceable access right of the data subject (article 15 GDPR) - a gap that five years of administrative and judicial litigation have not bridged.