A staggered timeline
The European regulation on artificial intelligence, known as the AI Act (Regulation (EU) 2024/1689), was adopted by the European Parliament on 13 March 2024 by 523 votes in favor, 46 against and 49 abstentions. It entered into force on 1 August 2024, but its obligations apply in stages: prohibited practices since 2 February 2025, rules for general-purpose AI models since 2 August 2025. 2 August 2026 marks the general application date of the text and, above all, the start of enforcement.
What changes since 2 August
Since this date, the AI Office of the European Commission and national authorities have their enforcement powers. Concretely:
- the AI Office can enforce the obligations of providers of general-purpose models: technical documentation, copyright compliance policy, public summary of training content, and additional obligations for the most advanced models presenting systemic risks;
- prohibited practices, such as manipulation, exploitation of vulnerabilities or social scoring, can now be sanctioned;
- new transparency obligations (Article 50) apply: a chatbot must indicate that it is not human, deepfakes must be flagged, and AI-generated content must carry a machine-readable marking. Systems already on the market before 2 August 2026 have until 2 December 2026 for this marking.
The division of competences is precise: the AI Office controls systems offered by the provider of the underlying model and those integrated into very large platforms or search engines designated under the DSA; national authorities control the others; the European Data Protection Supervisor handles EU institutions. The Commission has launched a complaint tool, a channel for whistleblowers and published a first list of more than 180 signatories of its code of good practices on transparency of AI-generated content. A scientific panel of 60 independent experts supports the system.
The big postponement of 'high risk'
The most restrictive part of the text, the rules on high-risk AI systems, does not yet apply. The 'Digital Omnibus on AI' regulation (Regulation (EU) 2026/1744), approved by the European Parliament in June 2026 (423 votes in favor, 57 against, 174 abstentions), validated by the Council on 29 June and entering into force on 27 July 2026, postponed their application:
- until 2 December 2027 for standalone systems listed in Annex III (recruitment, education, credit, access to public services, biometrics…);
- until 2 August 2028 for systems integrated into already regulated products, such as medical devices or machinery.
The Council justifies this delay by the lag in technical standards and national structures. The Omnibus also postponed the creation of national regulatory 'sandboxes' to 2 August 2027 and added a new prohibition: systems generating non-consensual intimate images or child sexual abuse content will be banned from 2 December 2026. The risk-based framework, however, remains intact: prohibitions, rules on general-purpose models, and transparency obligations follow their original timeline.
Dissuasive penalties
The AI Act provides for fines of up to 35 million euros or 7 % of the global annual turnover, whichever is higher, for prohibited practices (Article 99). This is higher than the GDPR ceiling (4 %). For providers of general-purpose models, the Commission may impose up to 3 % of global turnover or 15 million euros (Article 101). It remains to be seen how these penalties apply to providers established outside the Union.
France lagging behind on its authorities
The effectiveness of the mechanism depends on the member states, and the Commission itself reminds us: national authorities still need to be designated and provided with resources. In France, the government presented a plan in September 2025 based on existing regulators, coordinated by the DGCCRF, the single point of contact, and by the DGE, which represents France on the European AI Committee. During the examination of the bill adapting to Union law (DDADUE), government amendments strengthened the role of the CNIL, particularly regarding prohibited practices. The Senate adopted the text on 18 February 2026, but it is still awaiting examination by the National Assembly: the designation is therefore not final.
What this deadline reveals
2 August 2026 is neither the great evening announced in 2024, nor a renunciation. Europe has chosen to start with what is most mature (general-purpose models, transparency, prohibitions) and to defer what requires technical standards still under development. This pragmatism has a political cost: it fuels the idea of regulation that retreats in the face of industrial pressures. It also has a logic of sovereignty: an inapplicable text would have mainly penalized European companies, less equipped than American giants to absorb legal uncertainty.
The real test will come from the first investigations by the AI Office into major models, and from the ability of states, including France, to establish credible authorities in time before December 2027.
Sources: European Commission, IP/26/1714, 31 Jul. 2026 ; AI Act Service Desk, application timeline ; Council of the EU, 29 Jun. 2026 ; European Parliament, omnibus vote ; Regulation (EU) 2026/1744, Official Journal ; Regulation (EU) 2024/1689 (AI Act) ; Ministry of Economy, 9 Sep. 2025 ; Banque des Territoires, 18 Feb. 2026.
