EU AI Act faces pause barely a year after becoming law

The European Union's Artificial Intelligence Act is barely a year old. It entered into force in August 2024 after three years of negotiation, hailed as the world's first comprehensive rulebook for AI. Yet the political momentum that carried it over the line has already reversed. On 6 June, the Commission's tech sovereignty chief Henna Virkkunen told a meeting of EU ministers in Luxembourg that the bloc should not rule out postponing parts of the law if the technical standards and compliance guidelines are not ready in time. The high-risk AI obligations, the core of the regulation, are not even due to apply until 2026 or 2027.

A shift from safety to competitiveness

When the AI Act was agreed, EU officials presented it as a global gold standard designed to keep humanity safe from dangerous technology. The focus has since moved decisively toward competitiveness. The new Commission, installed in December 2024, has made simplification and industrial policy its twin priorities. President Ursula von der Leyen's 'competitiveness compass' and the Draghi report on European competitiveness both argue that regulatory burden is holding back the continent's ability to produce AI champions capable of rivalling the United States and China.

That argument has found a receptive audience in national capitals. France and Germany have both signalled concern that the AI Act could disadvantage homegrown companies such as Mistral AI and Aleph Alpha. The Polish presidency of the Council, which took over in January, has made digital simplification a flagship theme. The result is a Commission that now speaks the language of 'regulatory relief' where it once spoke of 'fundamental rights'.

Industry divided over certainty versus delay

Business groups are not speaking with one voice. The Computer & Communications Industry Association (CCIA), whose members include Amazon, Google, Meta and Microsoft, submitted formal feedback to the Commission in April arguing that a postponement is 'essential to give companies enough time to prepare'. The US government echoed that position in its own submission. But Tomasz Snażyk, chief executive of the Polish business group AI Chamber, told me the indecision is worse than either outcome. 'If you actually want to impose something, impose it, if you want to stop the clock, let's stop the clock,' he said. 'People want to be very certain of what is going on.'

Snażyk's members, mostly smaller European AI developers, need clarity to plan product roadmaps and raise capital. A pause that lasts months or years creates a regulatory limbo in which investors cannot assess compliance costs. That, in turn, makes European AI startups harder to fund than their US counterparts, who operate under a lighter, voluntary framework. The irony is that the very companies the Commission says it wants to help may be the ones most damaged by uncertainty.

Safety campaigners warn of dismantled protections

Digital rights advocates are alarmed. Blue Duangdjai Tiyavorabun, a policy adviser at EDRi, described the move as 'dismantling hard-won legal protections in a matter of months'. The AI Act prohibits certain practices outright, social scoring, real-time biometric identification in public spaces, subliminal manipulation, and those bans took effect in February. But the obligations on high-risk systems, covering everything from medical devices to recruitment algorithms, remain on the horizon. Campaigners argue that delaying them leaves a vacuum in which unsafe systems can be deployed without oversight.

Kim van Sparrentak, a Dutch Greens MEP who helped negotiate the Act, takes a different tack. She argues that enforcement, not delay, is the answer. 'As long as we don't have an AI Act that is being enforced, we keep systems on the market of which you don't know whether these function well or if these are safe,' she said. 'That to me is essential if we want to ensure that a lot of companies start using AI.' She cites a figure that only 13% of European companies properly use AI, a statistic that, if accurate, suggests the barrier to adoption is not regulation but something else: skills, data access, or capital.

The missing guidelines and the simplification review

Part of the problem is self-inflicted. The AI Act relies on harmonised technical standards and detailed guidance to turn legal requirements into engineering specifications. Those standards are being developed by European standardisation bodies CEN and CENELEC, but the process is running late. The Commission's own AI Office, led by Kilian Gross, has not yet published the promised guidance on transparency obligations for general-purpose AI models or on the 'human in the loop' requirements for high-risk systems. Industry groups have seized on that gap to argue that compliance is impossible on the current timetable.

Into that vacuum steps the Commission's digital simplification review, due to be presented in December. Virkkunen has said it will focus on AI, cybersecurity and data rules. Gross has indicated that 'targeted' changes to the AI Act are on the table. That phrasing, targeted changes, is doing a lot of work. It could mean technical corrections. It could mean substantive carve-outs for general-purpose AI. The Commission spokesperson Thomas Regnier insisted the institution remains 'fully committed to the main goals of the AI Act', but added that 'all options remain open for consideration'.

Parliament prepares to grill the Commissioner

The European Parliament is not a passive observer. On Tuesday, Virkkunen appears before the parliamentary working group that monitors AI Act implementation, co-chaired by Irish Renew MEP Michael McNamara. He has made clear the session will be dominated by the pause question. 'Just to throw everything out, there would be a risk of throwing out the baby with the bathwater,' McNamara said. He wants specifics: what exactly would be postponed, and for how long. Van Sparrentak framed the hearing as a test of whether the Commission will 'stand up that we have rules which will ensure that AI systems in Europe are safe' or whether it will 'start listening to Trump', referencing the US president's public opposition to the Act.

The Trump administration's intervention is unusual. The US government rarely weighs in on EU legislative implementation with such directness. Its April submission to the Commission called for a delay, aligning with the Big Tech lobby. That has given European critics of the Act a powerful external ally, and given defenders a convenient villain. Van Sparrentak's line, 'It's super unwise to start listening to Trump', captures the political dynamic: a European law, negotiated by European institutions, is being reconsidered under pressure from a foreign administration that has made no secret of its desire to weaken European tech regulation.

The Brussels effect at stake

The broader consequence is the credibility of the 'Brussels effect', the idea that EU rules become global standards because companies comply once for the single market and then apply the same standards worldwide. The GDPR set that precedent. The AI Act was supposed to extend it to artificial intelligence. If the EU rewrites its own law before the ink is dry on the implementation timetable, other jurisdictions will take note. Brazil, Canada, South Korea and Japan are all drafting AI legislation. They will watch to see whether the EU stands by its rules or treats them as opening bids.

There is also a domestic political risk. The AI Act passed with a broad coalition: socialists, liberals, greens and the centre-right European People's Party all voted for it. Reopening the text risks fracturing that coalition. The EPP, which backed the simplification agenda in the election campaign, may find its MEPs reluctant to dismantle a law they negotiated. The Greens and the Left will oppose any weakening. Renew sits in the middle, with McNamara and van Sparrentak already signalling resistance. A parliamentary vote on any amendment would be unpredictable.

What the numbers actually tell us

The 13% adoption figure cited by van Sparrentak deserves scrutiny. It comes from a Eurostat survey on ICT usage in enterprises, though the precise methodology and date are not always clear in political citations. What is clear is that European AI adoption lags the US. The Draghi report put EU private investment in AI at roughly €10 billion in 2023, compared with €60 billion in the US. The regulation is one variable among many: talent retention, compute access, fragmented digital markets, and risk-averse venture capital all play larger roles. Blaming the AI Act for the adoption gap is convenient but not obviously supported by evidence.

Equally, the compliance cost estimates vary wildly. The Commission's own impact assessment suggested average compliance costs of €160,000 to €330,000 per high-risk AI system for SMEs. Industry estimates run higher. But those costs are incurred only when the rules apply, in 2026 or 2027. A pause now does not save money today; it defers the spending decision while extending the period of uncertainty.