EU Eases Its Landmark AI Act, Delaying the Toughest Rules

The European Union has blinked. On 7 May, negotiators from the European Council, the Parliament and the Commission reached a provisional deal to soften and delay parts of the AI Act, the bloc's flagship rulebook for artificial intelligence and the first serious attempt by any major jurisdiction to regulate the technology comprehensively. It is the first set of amendments since the law was adopted in June 2024, and the direction is unmistakable: less now, later.

The deadlines move

The headline change is timing. Obligations for high-risk AI systems listed in Annex III, the category covering uses such as hiring, credit scoring and access to essential services, were due to take effect on 2 August 2026. They have been pushed back to 2 December 2027, a delay of sixteen months. Rules for AI built into physical products such as medical devices and industrial machinery slip from 2027 to August 2028. Even the lighter transparency duties, which require that AI-generated or manipulated content be labelled, move from August to December of this year, per the European Council and Euronews.

Brussels has framed the package, nicknamed the Digital Omnibus, as a way to cut red tape and give companies room to breathe without abandoning the law's risk-based spine. The heaviest obligations still fall only on systems whose failure could genuinely threaten health or safety.

One area gets tougher

Not everything was loosened. The amendments add two new outright bans, on AI used to generate or manipulate non-consensual intimate imagery and child sexual abuse material. That prohibition takes effect on 2 December 2026. It is a pointed acknowledgement that the most visible harm from generative models so far has not been a rogue superintelligence but ordinary abuse at scale.

Not everyone is satisfied

The carve-outs drew criticism from lawmakers who worry the EU is unpicking its own framework. Green MEP Sergey Lagodinsky warned that exempting machinery sets a precedent. "By having excluded machinery, we're making a first step into fragmenting AI regulation," he said, arguing that regulators will now have to "fill gaps through guidance, codes of conduct, and enforcement action faster than full legislative cycles allow."

The amendments still need formal sign-off, expected in June with publication in July. The shift sits in pointed contrast to the United States, where, as we reported last week, individual states are racing to pass their own AI laws rather than waiting for Washington. Europe is easing off a single comprehensive standard just as America splinters into fifty. Different roads, and it is far from clear yet which one leads somewhere better.