It was presented as a model, the first structured legislative framework on artificial intelligence in the world, an ambitious text, carried by the European Commission, applauded by Member States, NGOs, European technological companies, but as it approaches its entry into force, AI Act is at the heart of a reversal of the situation And now comes up against the harsh reality of their application, which everyone now judges too expensive, too vague, and incomplete.
We can spoil without regret the conclusion of the last episode, those who defended it in arms, ask today … that we press pause with the greatest diligence.
From consensus to the sling
If the criticisms accumulate for several weeks, the rocking point has been played last week, with an open letter signed by 33 entrepreneurs, founders of European startups, which calls for “freezing the regulatory clock”, followed on July 4 by Fifty leaders of large European groups (Airbus, Publicis, Totalenergies, BNP Paribas, Axa, Siemens, etc.) who joined the call. All now denounce an overly rapid implementation, without safeguards, and at high legal risk for their companies.
“Proceeding without clarity can leave companies blocked between contradictory interpretations, compromise innovation and discourage investments”, write the signatories.
The technical trap of AI Act
AI Act is based on a simple principle, any IA system classified as “at high risk” must demonstrate its compliance with a series of requirements (transparency, security, human supervision, robustness, etc.). To do this, he relies on technical standards supposed to offer a common reference framework.
But these standards, written by CEN and CENELEC, will not be ready before 2026. Companies will therefore have to prove their compliance … without a repository. And in the absence of a “presumption of conformity”, the weight of proof is based on the producer.
For a startup, this means mobilizing legal teams, auditing each line of code, and documenting evolutionary processes. Estimated cost: € 200,000 per year.
The Mistral paradox, defend regulation as long as it only applies to others
The case of Mistral aivalued at more than 2 billion euros and perceived as the French champion of Open Source, is emblematic. The startup, including the co -founder Arthur Mensch One of the signatories of the call for a break in the application of AI Act, had however been presented as one of the natural beneficiaries of the European framework. The text was to offer a protective regulatory space for emerging players against GAFAM.
Today, if Mistral does not reject regulationthe startup alerts on The effects of an ill -calibrated applicationespecially for open source models. In its current version, Ai Act does not clearly distinguish the obligations applicable to GPAI depending on whether they are owners or open. This ambiguity could expose community projects to legal constraints similar to those imposed on giants such as Openai or Google Deepmind, without having the means. The paradox is again there, a regulation designed to balance the game risks, for lack of adjustments, to penalize the actors it intended to protect.
Brussels surrounded by its own startups, industrialists, and member states
The sling does not only come from startups, Several Member States now express explicit reserves on AI Actin particular on its application calendar and its technical feasibility. Since July 1, Denmark, which provides the rotating presidency of the Council of the European Unionaligned himself with the break as a break. The minister in digital Caroline Stage was spokesperson for 44 Danish companies calling for a freeze on the implementation. Sweden supported this positionby the voice of his Prime Minister ULF KrisSSONwhich also pleads for a regulatory break. In Italycritical voices are raised within the tech and associative ecosystem, denouncing the legal vagueness and the dissuasive effects of the text on innovation. In Germanyif the government remains officially favorable to AI Act, industrial giants as Siemens, SAP, Mercedes-Benz or Lufthansa publicly joined calls to an adjustment of the text. This Nordic and industrial front is gradually wider isolating the commissiontaken between political urgency and economic pressure.
A regulation designed overhanging economic reality
The most severe criticism formulated by the signatories, testifies to the lack of European lucidity, AI Act was thought As if Europe was already the world leader in AI. We must remember our dependence on the United States for its chips, of China for certain components, that our startups are structurally smaller, less capitalized, less legally equipped
Regulating what is not controlled, without equipping those who must apply the rule, amounts to discouraging those that are pretended to protect.
The commission delayed, without reassuring
Faced with pressure, the European Commission is trying to reassure and Thomas Regnier, spokesperson of the CE indicated various measures:
- A “AI Act Service Desk” will be launched to support companies.
- A “Digital Simplification Omnibus Package” is under study.
- THE GPAI code of conductscheduled for August 2025, could be rejected at the end of 2025.
But it remains vague, and above all, the binding rules of AI Act remain valid legally. Companies will have to respect them even in the absence of official standards, Thomas Régnier was more than clear on this subject: ” A Legal Text is a legal text. Legal Deadlines Are LEGAL DEADLINES. Adopted by Our Co-Legislators. »»
It is not by drawing the land that you create a winning team.
If AI Act is not rejected on the bottom, Its implementation, precipitated, unbalanced and legally risky, makes the ecosystem switch to uncertainty. Brussels wanted to give Europe a role of moral leader but above all risks producing an opposite effect: the slowness of its industrial adoption, the exodus of its talents which will develop their projects on less constrained land, and ultimately the reluctance of its investors.
