The approval of the European Union’s Artificial Intelligence Act represents far more than the adoption of a new piece of legislation. It illustrates how democratic institutions attempt to govern a technological transformation whose pace often exceeds the capacity of governments themselves to understand, evaluate and regulate it. Rather than being merely another regulatory initiative, the AI Act reflects Europe’s broader vision of how technological innovation should coexist with fundamental rights, legal certainty, economic competitiveness and public trust. In many respects, it marks the beginning of a new phase in the relationship between governments and artificial intelligence, one in which public institutions are no longer simply reacting to technological change but are actively attempting to shape the environment in which that change will unfold.
When the ambassadors of the twenty-seven Member States unanimously endorsed the final version of the AI Act in early 2024, they were not making a political decision overnight. That vote was the culmination of several years of negotiations involving the European Commission, the European Parliament, national governments, technical experts, regulators, industry representatives and civil society organisations. The unanimous endorsement demonstrated that, despite significant disagreements throughout the negotiation process, the Member States ultimately recognised the strategic importance of presenting Europe with a common regulatory framework capable of addressing one of the defining technologies of the twenty-first century.
The legislative journey itself reveals much about the complexity of governing artificial intelligence. Unlike many previous areas of digital regulation, AI is not a single technology, nor is it limited to one industrial sector. Artificial intelligence increasingly influences healthcare, education, finance, transportation, manufacturing, scientific research, public administration, national security and countless other domains. Consequently, any attempt to regulate AI must inevitably deal with an extraordinary diversity of technologies, applications, risks and stakeholders. Designing legislation that is sufficiently robust to protect citizens while remaining flexible enough to accommodate future technological developments became one of the greatest challenges faced by European policymakers.
Although political agreement on the main principles of the legislation had already been reached in December 2023, transforming those political compromises into legally coherent text required weeks of intensive technical work. Legal experts had to refine definitions, reconcile cross-references, clarify obligations and ensure consistency across hundreds of articles and recitals. This phase is often invisible to the public, yet it is precisely where many of the practical implications of legislation are determined. In highly technical regulatory fields such as artificial intelligence, legal precision becomes almost as important as political consensus.
One of the most controversial aspects of the negotiations concerned the regulation of what are now commonly referred to as General Purpose AI models, particularly the increasingly powerful foundation models capable of supporting a wide variety of downstream applications. Systems such as GPT-4 demonstrated that a single model could serve as the technological foundation for thousands of different products and services, dramatically changing the traditional logic of technology regulation. Instead of regulating only individual AI applications, policymakers now faced the challenge of deciding whether the underlying models themselves should also be subject to regulatory obligations.
This question exposed a deeper strategic dilemma for Europe. France, Germany and Italy argued that imposing extensive regulatory requirements on developers of these powerful foundation models might inadvertently weaken Europe’s emerging AI industry. At a time when American companies such as OpenAI, Google and Anthropic were rapidly advancing, and Chinese firms were accelerating their own AI capabilities, European governments feared that excessive regulation could discourage investment, reduce innovation and place European start-ups at a competitive disadvantage. Companies such as Mistral AI in France and Aleph Alpha in Germany were frequently cited as examples of promising European enterprises that required sufficient regulatory flexibility if they were to compete globally.
The European Parliament, however, approached the issue from a different perspective. Parliamentarians argued that excluding the most powerful AI systems from binding regulation would fundamentally undermine the credibility of the entire legislative framework. If the greatest risks potentially originated from the largest and most capable models, then allowing these systems to operate under voluntary commitments while imposing mandatory obligations on smaller AI developers appeared difficult to justify. From their perspective, regulatory proportionality required that obligations increase with capability and potential impact rather than decrease.
The eventual compromise illustrates one of the defining characteristics of European policymaking: the search for balanced solutions rather than absolute victories. Instead of choosing between strict regulation or complete flexibility, negotiators established a layered regulatory model. Basic transparency obligations would apply across all general-purpose AI models, ensuring a common level of accountability throughout the ecosystem. Additional and more demanding requirements would apply only to those models considered sufficiently powerful to create systemic risks, introducing the principle that regulatory obligations should be proportional to technological capability and societal impact.
Achieving this compromise required significant political negotiation. Even after the provisional agreement had been announced, several Member States continued attempting to influence elements of the legislative text during the subsequent technical drafting process. However, by that stage, the essential legal architecture had already been established, leaving relatively little room for fundamental changes. When Belgium assumed the rotating Presidency of the Council in January 2024, it deliberately maintained a disciplined negotiation schedule, recognising that reopening major political compromises could jeopardise the entire agreement with the European Parliament.
France remained particularly active in seeking additional adjustments, attempting to gather sufficient support among other Member States to renegotiate aspects of the compromise. Yet political momentum gradually shifted. Germany eventually withdrew its reservations after internal discussions within the governing coalition, while Italy chose not to prolong the dispute, especially given its forthcoming leadership of the G7, where artificial intelligence had already become one of the principal geopolitical issues. Even within the French government, different ministries adopted contrasting positions. While economic policymakers remained concerned about maintaining industrial competitiveness, other ministries recognised the importance of protecting intellectual property rights and providing law enforcement authorities with carefully defined operational exceptions. These internal differences ultimately contributed to France accepting the final compromise, albeit while emphasising several implementation priorities.
Indeed, one of the most important lessons from the AI Act is that passing legislation does not conclude the regulatory process; rather, it initiates a much longer phase of implementation and institutional development. European legislation often establishes broad legal principles while leaving many operational details to secondary legislation, technical standards and regulatory guidance. In the case of the AI Act, the European Commission is expected to develop approximately twenty implementing measures that will clarify how many provisions should operate in practice. These secondary rules may ultimately prove almost as influential as the primary legislation itself.
Alongside these implementing measures, Europe has also begun constructing entirely new institutional capacities. The creation of the AI Office within the European Commission reflects the recognition that effective AI governance requires specialised expertise capable of keeping pace with rapidly evolving technologies. Rather than relying exclusively on existing regulatory bodies, the European Union is building dedicated institutions designed specifically to supervise advanced AI models, coordinate national authorities and support consistent implementation across Member States. National experts seconded from governments throughout Europe are expected to play an important role in developing this emerging regulatory ecosystem.
Even after political agreement had been secured, several Member States formally recorded concerns that they hoped would be addressed during implementation. Slovakia requested further clarification of key legal concepts, stronger international coordination and greater flexibility regarding AI systems used in private, non-professional contexts. Austria highlighted questions surrounding data protection, consumer rights and the use of remote biometric identification technologies, particularly where law enforcement authorities are involved. These statements demonstrate that legislative consensus does not necessarily eliminate all disagreement; instead, it often transfers many debates from the legislative phase to the implementation stage.
The AI Act also introduces a gradual implementation timetable, recognising that governments, companies and regulatory authorities require time to adapt to the new legal framework. Following formal adoption by the European Parliament and subsequent approval by the Council, the regulation entered the final stages before publication in the Official Journal of the European Union. Twenty days after publication, the regulation formally entered into force, although its various provisions became applicable according to different schedules. The prohibition of certain unacceptable AI practices begins after six months, obligations for general-purpose AI models apply after one year, while most remaining provisions take effect after two years. Certain high-risk AI systems subject to additional conformity assessments receive an additional year before full compliance becomes mandatory.
This phased approach reflects a broader philosophy of regulatory transition. Rather than imposing immediate obligations across every sector simultaneously, European lawmakers sought to provide sufficient time for organisations to develop compliance capabilities, establish governance mechanisms and adapt internal processes. In doing so, the AI Act acknowledges that regulating artificial intelligence is not simply about enforcing legal obligations but about supporting the gradual construction of an entirely new governance ecosystem.
Ultimately, the significance of the AI Act extends well beyond Europe itself. As the first comprehensive legislative framework dedicated specifically to artificial intelligence, it is already influencing policy discussions around the world. Governments, international organisations and regulatory agencies are closely examining the European model, not necessarily to replicate it in every detail, but to understand how democratic societies might balance technological innovation with public accountability. Whether the AI Act ultimately succeeds will depend not only on the quality of its legal provisions but also on the effectiveness of its implementation, the adaptability of its institutions and its ability to evolve alongside one of the fastest-moving technological revolutions in modern history.