On March 26, 2026, by an overwhelming majority of 569 votes to 45, the European Parliament adopted the postponement of the application of rules on high-risk artificial intelligence systems provided for by the AI Act. The initial deadline of August 2, 2026, has been pushed back to December 2, 2027, for systems in Annex III, and even to August 2, 2028, for those in Annex II. Presented as indispensable support for European industry, the decision deserves sober reflection: is it an act of pragmatism or a strategic abdication?
The Context: A Regulation Born in a Different World
The AI Act was adopted in June 2024, in a climate very different from today's. At the time, Europe believed it could simultaneously become a global regulatory leader – as it had with the GDPR – and build a sovereign AI industry. Two years later, the assessment is harsh: OpenAI, Anthropic, Google DeepMind, and now Mistral AI continue to operate primarily from the United States or with non-European capital, while China has launched its own large-scale models. The window of opportunity has largely closed, and the regulation, designed to frame a nascent industry, risked freezing what little European innovation remained.
Arguments for Postponement
The proponents of the postponement — with Commissioner Thierry Breton no longer in office, MEP Eva Maydell has become its main advocate — put forward three solid arguments. Firstly, the technical standards from the European Committee for Standardization (CEN-CENELEC) are not yet ready: applying the rules without verifiable standards would expose companies to massive legal uncertainty. Secondly, the penalties foreseen — up to 35 million euros or 7% of global turnover — are deemed dissuasive to the point of discouraging European startups from deploying any risk system. Finally, the United States and China have no equivalent constraints: penalizing the European industry would accentuate its relative decline.
Arguments Against
To these arguments, opponents — rights advocacy NGOs, several Green MEPs, some researchers like the French Mireille Hildebrandt — offer an inverse interpretation. Postponing the application of the rules by fourteen months means offering a clear path for foreign operators to deploy un-audited AI systems in Europe, in areas as sensitive as predictive justice, social scoring, or biometric recognition. It also weakens the European regulatory model, which drew its strength precisely from its speed of application. In the longer term, it risks the European market becoming a free experimentation ground for models designed elsewhere, without reciprocation.
A Possible Third Way
Between brutal application and unconditional postponement, a third way exists: apply the rules only to foreign operators deploying models in Europe, while temporarily exempting European companies until the CEN-CENELEC standards are finalized. This reversed asymmetry — using the European market as a lever for sovereignty rather than a uniform constraint — was not seriously debated in Strasbourg. However, it would have the merit of aligning regulation with the continent's industrial interests, instead of indiscriminately penalizing everyone in the name of a legal universalism that primarily serves competitors.
Editorial Opinion
The vote on March 26 undoubtedly marks the end of the European regulatory exception. Brussels wanted to be the global sheriff of AI; for now, it is content with being its lazy referee. The postponement is not in itself a mistake — it was technically necessary — but it becomes a political mistake if it is not accompanied by an aggressive industrial strategy: massive funding for Mistral AI and other European champions, preferential public procurement, sovereign purchase doctrines. Without these, Europe will have simply chosen not to regulate what it can no longer produce. That would be dramatic.
Key Takeaways
- On March 26, 2026, the European Parliament votes 569 to 45 to postpone the AI Act.
- Deadline for Annex III systems pushed to December 2, 2027, Annex II to August 2, 2028.
- CEN-CENELEC technical standards not yet finalized, partly justifying the postponement.
- Penalties foreseen: up to 35 million euros or 7% of global turnover.
- A third way could have targeted only foreign operators, without penalizing European industry.
