Is the postponement of 2 August 2026 final?

Not yet. The "Digital Omnibus" that pushes back the "high-risk" obligations was the subject of a provisional agreement on 7 May 2026 between the Commission, the Parliament and the Council; it still has to be formally adopted and published to become final.

Which AI practices are already banned outright in the European Union?

Since February 2025: social scoring, the untargeted scraping of facial images to build facial recognition databases, emotion recognition at work and in schools, and systems that manipulate or exploit vulnerable people. Those bans are not postponed.

Does the AI Act apply to non-European companies?

Yes. The regulation is extraterritorial: it applies as soon as the output of an AI system is used in the Union, even if the provider is established in the United States or elsewhere. That is one of the sources of tension with Washington.

Are ChatGPT, Gemini or Mistral covered?

Yes, as general-purpose AI models (GPAI). Their obligations (transparency, a summary of training data, copyright) have applied since August 2025, framed by a Code of Practice signed by OpenAI, Google, Microsoft, Anthropic and Mistral. Meta refused to sign it.

Who enforces the AI Act in France?

Several authorities: the CNIL (the French data protection authority) as lead, for deployments and banned practices such as emotion recognition; the ANSM (the French medicines and health products agency) for AI in medical devices; plus the DGCCRF (consumer protection) and Arcom (media regulator) depending on the field.

EU AI Act 2026: what changes and what Europe delays

The 2 August 2026 turning point… already postponed

2 August 2026 was meant to be the great switch-over for the AI Act, the European regulation on artificial intelligence presented as the strictest in the world. On that date, the obligations on “high-risk” AI were due to apply, and the Commission was due to gain the power to sanction large models. Except that Europe has just pushed back its core — under pressure from the tech giants and the Trump administration. A breakdown, without hype: what really changes, what is postponed, and what it implies.

What was due to switch over on 2 August 2026

The AI Act has been in force since 1 August 2024, but it applies in stages. 2 August 2026 concentrated two major deadlines:

the obligations for high-risk AI systems (Annex III) — recruitment, credit, education, essential services — with conformity assessment, documentation and human oversight;
the Commission’s sanction powers over general-purpose models (GPAI, such as ChatGPT or Gemini): demanding documents, running evaluations, even imposing the withdrawal of a model, and issuing fines.

The U-turn: the “Digital Omnibus”

On 19 November 2025, the Commission presented a “Digital Omnibus” proposing to push those obligations back. On 7 May 2026, a provisional agreement between the Commission, the Parliament and the Council confirmed the delay — the first revision of the AI Act since its adoption in 2024.

The result: the standalone “high-risk” obligations (Annex III) move from 2 August 2026 to 2 December 2027; those built into products that are already regulated (Annex I) move to 2 August 2028. The agreement also tightens the definition of “high risk”: an AI that merely assists or optimises, without creating a risk to health or safety, will no longer fall into the category automatically.

Why the retreat

Officially, the technical standards and the designation of national authorities are not ready. Unofficially, the pressure runs two ways. On one side, the tech giants: Meta refused to sign the GPAI Code of Practice, judging it legally vague and going beyond the law (OpenAI, Google, Microsoft, Anthropic and Mistral signed it). On the other, Washington: the Trump administration threatened the EU with “immediate and substantial retaliation” and tariffs if it tightened its tech rules; according to the Financial Times, Brussels is “in discussion” with Washington about adjustments. A sovereignty standoff we have already explored in relation to AI in banking and healthcare.

What applies anyway

Not everything is postponed. Already in force:

the banned practices (since February 2025): social scoring, mass scraping of facial images, emotion recognition at work and in schools;
the obligations on GPAI models (since August 2025): transparency, a summary of training data, respect for copyright.

The regulation also provides for labelling AI-generated content (deepfakes) — the kind of abuse seen with Grok’s sexualised deepfakes. And the fines remain deterrent: up to 35 million euros or 7% of worldwide turnover for a banned practice, 15 million euros or 3% for other breaches (and for GPAI).

What this changes in France

In France, enforcement is entrusted to several authorities: the CNIL (the data protection authority, acting as lead), the ANSM (the medicines and health products agency, for AI in medical devices), plus the DGCCRF (consumer protection) and Arcom (media regulator) depending on the field. In practice, if an AI screens you for a job or a loan, the “high-risk” guardrails meant to protect you… will wait until late 2027. Meanwhile, financial regulators are already alarmed by what the models can actually do.

Le Recul’s reading

Presented as the world’s bulwark against AI, the AI Act is being weakened before it has even bitten. The timetable is slipping, the definition of “high risk” is narrowing, and the key commitments of the large models remain… voluntary. The real question is no longer “is Europe regulating AI?”, but “is it holding up against lobbying and American pressure?”. For now, the answer leans towards postponement.

What to take away

2 August 2026 was due to apply the “high-risk” AI obligations and open up sanctions on GPAI models.
The Digital Omnibus (provisional agreement of 7 May 2026) pushes high risk back to 2 December 2027 (and to 2 August 2028 for AI built into regulated products).
The drivers of the retreat: lobbying by the giants (Meta refused the Code) + pressure from Trump (threats of retaliation).
Still in force: banned practices (February 2025) and GPAI obligations (August 2025); fines of up to 35 million euros / 7% of worldwide turnover.

The figure to remember: 16 months. That is the delay to the core of the AI Act, pushed from August 2026 to December 2027.

Our verdict. AI promises, Europe was meant to check — but it is giving itself 16 more months, under pressure. On paper, the AI Act remains the most ambitious law in the world; in practice, its enforcement is retreating as the giants and Washington push. To watch: the formal adoption of the Omnibus, and whether the sanctions on GPAI models scheduled for August 2026 do hold.

AI Act: what changes on 2 August 2026 (and what Europe is putting off)