The European Parliament and Council reached a provisional deal last week to delay and trim the AI Act, the bloc's flagship legislation governing artificial intelligence. Under the agreement, announced in a joint Council press release, the high-risk rules that originally applied from August 2026 will not take effect for standalone systems until December 2, 2027, and not until August 2, 2028 for AI embedded in regulated products like medical devices and vehicles.
The package, branded the AI omnibus, is being sold by negotiators as simplification rather than retreat. Smaller and mid-sized companies, including a new category of small mid-caps with up to 500 employees, get longer transition periods and lighter documentation requirements. Several pieces of overlapping reporting between the AI Act and the GDPR are being consolidated. The Commission says the changes will save European businesses an estimated 5 billion euros in compliance costs over the next five years, according to Tech Policy Press.
Civil society pushes back
Civil-society groups read the same announcement very differently. Algorithm Watch and EDRi, both quoted in analysis from Hogan Lovells, called the delay a two-year holiday for high-risk AI systems and warned that affected populations, including job applicants and benefits claimants, will spend that time inside systems with no enforced safeguards. The fact that the deal also moves the AI Office's enforcement powers later in the timeline drew particular concern.
One area did get tougher
The omnibus introduces an explicit, EU-wide ban on non-consensual sexual or intimate AI-generated imagery, and on AI-generated child sexual abuse material. That provision is fast-tracked and applies from December 2, 2026, a year ahead of the broader high-risk rules. It closes what civil-society groups had been calling the most glaring gap in the original Act. The change responds to a wave of deepfake-related incidents in European schools and workplaces over the last eighteen months.
The political math behind the delay is not subtle. American and Chinese AI firms have been operating without anything close to the AI Act's compliance overhead, and European companies including Mistral, Aleph Alpha and SAP have been lobbying for breathing room. The Commission has also been under pressure from member-state governments, particularly Germany and France, to avoid choking the bloc's own model developers before they reach scale.
Adoption of the final text is targeted before the August 2026 recess, with the European Commission opening a separate consultation on transparency-obligation guidelines earlier this month. Until then the legal status quo holds. After that, expect a slower, more fragmented rollout than the headline 2026 date implied, and a sharper political fight over enforcement than the original 2024 negotiations produced.
Commentarii · 0