Case study · Europe, with global reach · 2021–present
The EU AI Act and the contest to govern artificial intelligence
The first comprehensive AI law shows a regional actor exporting rules through market access, while the capability it regulates sits with a few private firms and rival powers pursue different models.
The case
The European Union’s Artificial Intelligence Act, Regulation (EU) 2024/1689, is the first comprehensive law on artificial intelligence adopted by a major jurisdiction. Proposed by the European Commission in April 2021 and agreed politically in December 2023, it entered into force on 1 August 2024 and applies in stages. It sorts AI systems by risk. Practices judged to carry unacceptable risk, such as social scoring and certain uses of biometric categorisation and emotion recognition, have been prohibited since 2 February 2025. Providers of general-purpose AI models have had obligations on documentation, transparency and copyright since 2 August 2025, supported by a voluntary code of practice published in July 2025, which most leading developers signed and Meta declined to sign. High-risk systems, such as those used in employment, education, credit and law enforcement, face the most demanding obligations.
The Act applies to any provider placing AI systems on the EU market, wherever it is based, which gives it reach beyond Europe. Its timetable has nonetheless slipped. In November 2025 the Commission proposed a ‘digital omnibus’ to simplify the rules, and on 7 May 2026 the Council and Parliament reached a provisional agreement postponing obligations for stand-alone high-risk systems from August 2026 to December 2027, and for high-risk AI embedded in regulated products from August 2027 to August 2028.
Approaches elsewhere diverge. The United States revoked its 2023 executive order on AI safety in January 2025 and published an action plan prioritising innovation in July 2025. China regulates generative AI through administrative measures in force since 2023. The UN adopted a Global Digital Compact in 2024.
Actors and stakeholders
- European Union (Commission and its AI Office, Parliament, Council)Supranational regional organisation
Role
Legislated the Act, enforces the rules on general-purpose models through the Commission’s AI Office and coordinates national authorities that supervise the rest.
Power and its sources
Normative and market power: firms comply to keep access to one of the world’s largest consumer markets, which can carry EU standards abroad (the ‘Brussels effect’).
- Developers of general-purpose AI modelsTransnational companies
Role
Mostly based in the United States. Most signed the 2025 code of practice; Meta refused, arguing that it created legal uncertainty and went beyond the Act.
Power and its sources
Control over compute, data, models and the platforms through which AI reaches users, which is Strange’s production and knowledge structures. They can lobby, litigate, delay launches in Europe and shape technical standards.
- United States governmentState
Role
Since 2025 has favoured deregulation at home and criticised EU digital rules as burdens on American firms, while using export controls on advanced chips to shape who can build frontier systems.
Power and its sources
Structural power over chips and cloud infrastructure; diplomatic and trade leverage over allies.
- ChinaState
Role
Regulates algorithms and generative AI through administrative rules that emphasise content control and security, and promotes its own vision of global AI governance.
Power and its sources
A large domestic market and the state capacity to make and enforce rules quickly; growing capability despite export controls.
- Civil society and digital rights organisationsOrganised civil society
Role
Campaigned for bans on biometric mass surveillance, criticised exemptions for national security and law enforcement, and oppose the delays agreed in 2026.
Power and its sources
Expertise, publicity and access to EU legislators, but no veto.
- People subject to AI decisionsAffected stakeholders
Role
Job applicants, students, borrowers, migrants and others assessed by automated systems; the Act gives them rights to complain and, for certain high-risk decisions, to an explanation.
Power and its sources
Limited direct power, exercised mainly through civil society, regulators and courts.
Through the key concepts
The Act is a study in normative and structural power. The EU has few frontier AI firms of its own, but it controls access to a rich market, and by writing rules first it can shape practice abroad, as its data-protection rules did. The firms hold a different power: the capability itself, and the knowledge on which regulators depend. Barnett and Duvall’s productive power appears in the contest over definitions, such as what counts as ‘high-risk’ or as a general-purpose model with systemic risk, which decides who carries obligations.
The EU presents the Act as digital sovereignty: the claim to set rules for technologies that shape its society even when they are built elsewhere. Its extraterritorial reach shows sovereignty exercised through market access rather than territory. Critics, including the US government, read the same reach as an imposition on others’ firms. The 2026 delay shows that regulatory sovereignty is itself constrained by competitiveness concerns and by dependence on foreign technology.
Legislation is an attempt to reassert public authority over private power: developers and platforms exercise authority over information and decisions without any legitimacy claim of their own. The Act’s input legitimacy is comparatively strong, having passed an elected Parliament and the Council. Its output legitimacy is contested, as industry warns of stifled innovation and civil society of weakened protection, and reliance on voluntary codes raises the question of whether co-regulation with the regulated is legitimate.
AI governance is a networked interdependence problem. Compute, chips, data and models are concentrated in a few firms and states; models cross borders instantly while regulators are national or regional. The Act therefore depends on cooperation it cannot command, and competing models of governance, rights-based in Europe, market-led in the United States and state-directed in China, risk fragmenting the rules for a technology that is global by design.
Perspectives
EU institutions and supporters
Trustworthy AI requires rules proportionate to risk, a single market needs one set of rules rather than twenty-seven, and setting standards early lets democratic institutions, not companies, define acceptable uses.
Competitiveness critics
Complex, front-loaded regulation raises costs for European firms that already trail American and Chinese competitors. The 2024 report on European competitiveness led by Mario Draghi warned that regulatory burdens hold back innovation, and the 2026 delay concedes the point.
The US administration
European rules discriminate against American firms and slow innovation. Leadership in AI is a matter of national security, best pursued through investment, export controls and light regulation.
Digital rights organisations
The Act contains too many exemptions, particularly for policing, migration and national security, and the omnibus weakens protections before they have taken effect.
Evaluation
The AI Act tests whether a regional actor can govern a technology it does not dominate. Its achievements are real: it turned contested principles into binding law, introduced bans that already apply and created an enforcement body for the most powerful models.
The case nonetheless cuts against an easy story of the Brussels effect. The capability remains with a few foreign firms, one major developer refused the voluntary code, the United States openly opposes European digital rules, and the EU itself agreed to postpone its high-risk obligations under competitiveness pressure. The most defensible judgement is that the Act has made the EU a norm entrepreneur at the emergence stage of AI governance, but whether its rules cascade depends on factors it does not control: the behaviour of the firms, the stance of the United States and China, and whether its own rules survive revision.
Questions to take further
- Can a regulator govern a technology whose capability sits mostly with foreign private firms?
- Is the Brussels effect best described as soft power, structural power or normative power?
- Does the 2026 delay show regulatory pragmatism or regulatory capture?
- Who should have a say in defining which uses of AI count as high-risk?
Sources
- 01Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) (opens in a new tab), 12 July 2024
- 02European Commission, General-Purpose AI Code of Practice, July 2025
- 03Gibson Dunn, ‘EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes’ (opens in a new tab), May 2026
- 04Sidley Austin, Data Matters, ‘EU Lawmakers Reach Provisional Agreement to Delay Key EU AI Act Obligations’ (opens in a new tab), 22 June 2026
- 05Anu Bradford, The Brussels Effect: How the European Union Rules the World (Oxford University Press, 2020)
- 06United Nations, Global Digital Compact (annex to the Pact for the Future), September 2024