Europe guts its own ai act: deadlines pushed, deepfake nudifiers banned
Strasbourg just blinked. Two EU committees voted Wednesday to defang the bloc’s flagship AI Act, pushing compliance dates two years into the future and quietly handing startups a get-out-of-jail card worth billions in avoided penalties.
The package—dubbed the AI Act Omnibus—delays oversight of high-risk systems such as biometric border gates, predictive-policing algorithms and AI résumé screeners until December 2027. Safety components embedded in medical devices or driverless cars get an even softer landing: August 2028. The watermarking rules for synthetic media slide too, giving frontier labs breathing room they never expected.
Why the retreat matters outside brussels
Europe wrote the first global rulebook on artificial intelligence; now it is rewriting the footnotes in real time. Venture capitalists from Paris to Tallinn are already reallocating dry powder, betting that the Continent’s “Brussels effect” has turned into a Brussels deferral. The signal is unmistakable: innovate first, comply later.
Startups win. Citizens wait. Digital-rights outfits warn that every extra month of lax enforcement is another month for biased hiring tools and welfare algorithms to hard-wire discrimination. “Postponing compliance postpones protection,” snapped the Center for Democracy & Technology’s Europe arm.
The one thing they did ban outright
While the calendar loosens, the no-go list tightens. So-called nudifiers—apps that strip clothes off real people’s photos—are now classified as unacceptable risk, lumped alongside subliminal manipulation and social-scoring schemes. Fines can hit €35 million or 7 % of global turnover, whichever hurts more. No grace period, no grandfather clause. If your model can mint non-consensual porn, you pull the plug or pull out of Europe.
Yet even here the drafters left a trapdoor: systems that inject “effective mitigation” remain legal. Expect lawyers to spend the next two years arguing over what “effective” means while the images keep circulating.
What happens next
The full Parliament still needs to rubber-stamp the package on 26 March, but the real fight moves to the Council where France and Germany lobby to keep their AI champions unshackled. Meanwhile, technical standards—the granular specs engineers actually code against—won’t be ready until late 2026. Without them, the August 2026 deadline was a mirage; the Omnibus simply admits it out loud.
Europe wanted to lead the world on ethical AI. Today it leads in regulatory acrobatics, juggling innovation sovereignty and citizen safety with the same pair of hands. One thing is certain: the next 36 months will be a stress test not for algorithms, but for trust.