In December 2023, European Commission President Ursula von der Leyen hailed the newly finalized EU Artificial Intelligence Act as a law that “transpose European values to a new era.” Three years later, after major revisions to the same legislation, she described it instead as creating “a simple, innovation-friendly environment.”
The shift in Von der Leyen’s rhetoric — from values-driven governance to innovation-focused regulation — reflects a broader change in Brussels' regulatory posture over the past three years. The AI Act has become one of the chief examples of that evolution.
The amendments finalized on May 7 postponed the implementation of rules governing AI systems in high-risk settings until 2027 or 2028, and removed industrial applications from the scope of those requirements altogether.
Crucially, however, the revisions left intact the timetable for rules covering general-purpose AI models developed by companies like OpenAI and Anthropic, which are set to take effect in August. As a result, some analysts and lawmakers argue that while the AI Act's implementation timeline has been relaxed, its core regulatory framework remains intact.
“The fundamental ambition of the legislation has not changed as dramatically as some commentators would say,” said Nicole Lemke, a senior policy researcher working on AI systems at the European think tank Interface. “The AI Act was always, and still is, based on a risk-based architecture.”
What has changed in the EU AI Act?
When the AI Act was initially approved in early 2024, it set out ambitious deadlines for regulating high-risk AI systems deployed in areas like employment, critical infrastructure or biometric identification. It also introduced rules for the general-purpose AI models powering systems like ChatGPT, Claude and Mistral.
Under the revisions, the implementation deadline for high-risk AI systems has been pushed back from August 2026 to December 2027, while requirements for AI integrated into products already covered by sector-specific regulations have been delayed until August 2028.
The extensions will allow the European Commission more time to finalize guidance on identifying high-risk systems after months of holdups, while also providing companies with a longer runway to prepare for compliance.
The other major change concerns AI used in industrial machinery, including applications in car manufacturing and advanced lithography machines that make semiconductor chips. Securing an exemption for industrial uses was a priority for German Chancellor Friedrich Merz, who wanted to ensure Germany’s legacy industries could deploy AI freely.
As a result, companies in these sectors will be subject only to existing sector-specific regulations. According to Julie Gabriel, co-founder and CEO of eyreACT, a company that automates EU AI Act compliance, the carve-out could offer an early hint at how the EU's approach to AI regulation may evolve in the years ahead.
The revisions also introduced several narrower changes, including new bans on using AI to generate child sexual abuse material and on so-called “nudifier” apps that create non-consensual intimate images.
But some lawmakers argue that the EU could have gone much further in simplifying the law. Axel Voss (EPP, DE), a shadow rapporteur on the amendments, said the package's only substantial deregulatory measure was the industrial carve-out. Beyond that, he argued, the law remains largely unchanged, with most modifications merely amounting to delayed implementation deadlines.
Why change the EU AI Act now?
The broader question for policymakers and analysts is why the EU moved to amend the AI Act so soon after its adoption in 2024. Supporters argue the changes were needed to boost European competitiveness in the global AI race. Critics, however, contend this narrative was pushed by lobbying from major U.S. tech firms and that the law had scarcely begun to take effect.
Much of the simplification drive across the EU's digital agenda can be traced to the 2024 Draghi Report, which urged the EU to accelerate innovation in strategic technologies.
That message has since been embraced by influential technology lobbying groups, including AmCham EU, the U.S. Chamber of Commerce to the EU, and the ACT Association, a group representing small and mid-sized technology companies.
A joint letter by American and European industry groups published in April 2026 stated they “share[d] the Commission’s commitment to a clear, simple and innovation-friendly implementation of the AI Act.”
Additional layers of regulation would further push Europe out of the global AI race, said Vicente Domecq, a tech and digital policy lead with the European Conservatives and Reformists in the European Parliament and a supporter of the changes to the AI Act. “The challenge is not only in safety and ethics,” he said. ”But the real challenge should be in innovation and investment.”
Yet some analysts dispute the claim that the AI Act hindered innovation. After all, the AI Act has largely still not been implemented, said Risto Uuk, head of European policy and research at the Center of Life Institute in Brussels. No fines have been issued under the law, while the key provisions governing high-risk AI systems and general-purpose AI models have yet to enter into effect.
Rather, said Luka Ignac, interim director for AI at the Center for Future Generations, Europe's lack of AI innovation is more closely tied to a lack of capital and market fragmentation.
“[The innovation] argument often ignores the fact that the EU's leverage is in regulatory frameworks,” he said, pointing to the so-called Brussels Effect coined by Anu Bradford. “How Europe positions itself in this environment is also thinking about setting governance standards that others can join.”
Instead, critics of the AI Act changes say that lobbying by U.S. Big Tech played a central role in securing the delays. Domecq said that he had spoken with representatives from several American tech companies about the AI policy. “Of course,” they played a role, he said. “I’ve received powerful insights and recommendations from companies that have their own interests, but of course they know the market.”
Gabriel begged to differ. “This whole AI Act is bad for innovation narrative has been pushed and paid for by American hyperscalers and Big Tech companies.”
Delays at EU AI Office stall high-risk AI rules
Another driver of the amendments was the EU AI Office's failure to produce key guidelines on time. Even though the EU AI Act initially passed in 2024, lawmakers left parts of the regulation for the EU AI Office to finalize. Guidelines for how to determine whether a system should be classified as high-risk was only published by the EU AI Office in May this year.
“I’ve been in some of the technical standards meetings where the three hours of discussions revolved around the term ‘ethics,’ what do we mean by that?” Uuk said.
Without an agreed definition of what constitutes a high-risk system, that portion of the law cannot be enforced. Gabriel said part of the problem stemmed from the legislation's original drafting process, which, she said, was done by non-technical MEPs who didn’t understand the challenges of coming up with complex technical standards from scratch.
That was echoed by Voss, the shadow rapporteur on the AI simplification. "We are not understanding the complexity of... being compliant with the law, because we are not going into the technical details.”
A final challenge, according to Uuk and Ignac, is making sure that the AI Office has enough backing to get through the huge amount of technical work needed to properly regulate the models. In the past, the AI Office has struggled to hire enough experienced technical staff.
What's next for the EU AI Act?
While the AI Act has just received a sweeping overhaul, the steepest challenge yet for the EU’s efforts to regulate AI will come at the start of August, when the enforcement provisions for general-purpose AI models come into effect.
Starting Aug. 2, the models provided by U.S.-based companies like OpenAI, Anthropic and Google and European firms like Mistral will have to comply with regulations guided by the Code of Practice, a document finalized last year that defines transparency, copyright and safety guidelines for models. If companies don’t comply, the AI Office can order model recalls, impose fines and mandate changes — a first-of-its-kind regulatory mandate that covers any firm offering its model in the EU.
Then there the question of the longer-term trajectory of the EU’s efforts to regulate a fast-moving technology. New AI applications, such as AI agents, are not directly covered by the AI Act
Voss said that there’s a feeling that the AI Act is already “old” given the rapid advances in AI. He argued that EU institutions should regularly revisit and update the framework to ensure it keeps pace with new technological developments and emerging applications.
“The hard question is: will Europe build the institutional capacity to regulate the technology at the pace it’s developing?” said Ignac. “That’s the biggest challenge for where we go from here, building institutional capacity, and in a way that doesn’t grow bureaucracy, but instead real technical expertise.”
Sign up to The Parliament's weekly newsletter
Every Friday our editorial team goes behind the headlines to offer insight and analysis on the key stories driving the EU agenda. Subscribe for free here.