Europe spent three years writing the world's strictest AI law. It just spent a few months walking it back.

On Nov. 13, the European Parliament voted 423-57 to approve the AI Act Omnibus, a package that delays the heaviest parts of the AI Act before they ever took effect. The high-risk rules that were supposed to kick in on Aug. 2, 2026, the ones covering AI used in hiring, education, credit scoring, biometrics and policing, are now pushed to December 2027. The version for AI built into regulated products like medical devices and cars slips to August 2028.

Officially, the delay is about missing technical standards. The harmonized rulebook companies were supposed to follow to prove compliance isn't actually finished. Only a handful of standards will be ready by mid-2026, with the rest trickling out through 2027. "These standards are not ready, and that's why we allowed ourselves in the AI omnibus to give us a bit more time," European Commission Deputy Director-General Renate Nikolay said.

The unofficial reason is Anthropic. When the company pulled its models out of Europe in June over US export rules, it set off a sovereignty panic across the bloc and handed every "Europe is falling behind" argument a fresh data point. Within weeks, the case for going easier on European AI companies basically wrote itself.

What changed in the law:

  • High-risk rules delayed: Annex III systems (hiring, credit, policing, education) pushed from August 2026 to December 2027. Annex I systems (regulated products) pushed to August 2028.
  • Industrial carve-out: A new exemption for AI used inside industrial machinery and certain regulated products, the only piece almost everyone agrees is a real substantive cut.
  • Registration weakened: The public database where high-risk AI systems had to be registered gets narrowed. The Commission's own estimate says this saves companies about €100 per registration.
  • Transparency softened: Several disclosure obligations for general-purpose models get loosened.

The people who actually built this law are not thrilled. Gabriele Mazzini, the lead author of the AI Act and now a research affiliate at MIT, said the EU's model has "grown too complex, too broad and too rigid and bureaucratic to adjust." That's the architect of the thing saying the thing doesn't work. Axel Voss, the German MEP who shadowed the amendments, was blunter: the package's "only substantial deregulatory measure was the industrial carve-out," and the rest is mostly cleanup.

The framing fight underneath all this is about who pushed for the delay and why. Julie Gabriel, who runs an AI Act compliance startup in Brussels, told Europe Says that "this whole AI Act is bad for innovation narrative has been pushed and paid for by American hyperscalers and Big Tech companies." Luka Ignac at the Center for Future Generations made the related point that Europe's AI gap is really a capital and market problem, not a regulation problem. The Software & Information Industry Association, which represents a lot of those American companies, has been warning Washington that export controls like the one that hit Anthropic look "antithetical to promoting U.S. technology around the world." Different pressure, same direction.

Civil society is not having it. A joint statement from EDRi, Access Now, Amnesty International, AlgorithmWatch and five other groups called the Omnibus a law that "weakens the AI Act, emboldens industry lobbying, and undermines the EU's credibility as a serious digital regulator." Their argument is simple: the things being delayed and trimmed are the things meant to protect people from AI used on them, not the things that make life hard for engineers.

INTO THE VALLEY

For a decade, Europe's whole pitch was that it would regulate what America wouldn't, and the rest of the world would follow. The Omnibus is the first real moment that pitch cracked, and it cracked not because the law was proven wrong but because one US company pulled its models and everyone in Brussels suddenly remembered they don't have a frontier lab of their own. The next two years won't be about whether Europe enforces the AI Act. They'll be about whether anyone outside Europe still thinks the AI Act is the template to copy.