As Washington Fights the States Over AI Rules, the EU Delays Its High-Risk AI Act Obligations to December 2027

The United States has no federal AI law, and the administration’s main AI fight in 2026 is with American states that wrote their own. Across the Atlantic, the EU has the opposite problem: one detailed law for 27 countries, which it has just slowed down.

The EU’s AI Act was meant to bite hardest on 2 August 2026. That was the date when rules for high-risk AI, the systems that screen job applicants, score credit, grade exams or support policing, were due to apply. They didn’t. Brussels moved the date, by law, a month before.

The change came through the Digital Omnibus, which the Commission proposed in November 2025. Negotiators struck a deal in May 2026, Parliament endorsed it in June by 423 votes to 57 with 174 abstentions, and the Council signed off on 29 June. It entered into force on 27 July 2026 as Regulation (EU) 2026/1744.

Here’s what moved, and what didn’t.

The new calendar

High-risk systems listed in Annex III of the Act now have until 2 December 2027. AI built into products that already have EU safety rules, like machinery or medical devices, moves from August 2027 to 2 August 2028.

Several things stayed put. The bans on prohibited practices have applied since February 2025. Obligations for general-purpose AI models have applied since August 2025, and the Commission’s power to fine their providers, up to 3% of worldwide turnover or €15 million, started on 2 August 2026. Transparency duties also started that day: telling people when they’re dealing with an AI system, marking AI-generated content in a machine-readable way and disclosing deepfakes. National authorities can now fine up to €35 million or 7% of turnover for prohibited practices.

The Omnibus also added two bans, on AI that generates non-consensual intimate images and on AI that generates child sexual abuse material, from 2 December 2026. And it handed the AI Office exclusive supervision of AI systems built on general-purpose models, plus AI inside the very large platforms covered by the Digital Services Act.

So the EU delayed the part of the Act that touches the most ordinary businesses, kept the part aimed at the big model makers, and gave itself sharper tools against them. That’s a fairly clear statement of priorities.

The voluntary code for general-purpose models, published in July 2025, shows the same split. Google, Microsoft, OpenAI and Anthropic were among 26 signatories. Meta refused, saying the code went beyond the Act.

Britain still has no AI bill

The UK took a different road from the start: no horizontal AI law, existing regulators applying existing rules. The King’s Speech of 13 May 2026 kept it that way. There was no general AI bill. There was a Regulating for Growth Bill with statutory sandboxes, and a police reform bill that creates a regulator for facial recognition.

Where Britain has legislated, it has gone after specific harms. The Crime and Policing Act 2026, which got Royal Assent in April, creates offences for tools that generate child abuse images and for deepfake intimate image apps. A backbench bill to ban superintelligence is due for second reading in November without government support.

China regulates the output

China’s approach starts from content. Labelling rules issued in March 2025 by the Cyberspace Administration and three other agencies took effect on 1 September 2025, alongside a mandatory national standard. AI-generated text, images, audio, video and virtual scenes need visible labels, plus hidden labels in file metadata.

That’s a year of enforcement already. The EU’s own marking duty only started in August 2026, with a four-month grace period for generative systems already on the market.

And Washington is fighting its own states

The United States has no federal AI statute either. Its fight is internal: whether federal policy should override state AI laws like Colorado’s and California’s.

Put the four side by side and each has picked a different lever. The EU regulates by risk category and is now pacing itself. Britain regulates by harm, one law at a time. China regulates what users see. The US is still arguing over who gets to regulate at all.

For a company selling into all four, the 2026 changes mostly buy time in Europe. They don’t reduce the number of rulebooks.