European AI Act: Towards Global Harmonization of AI Standards?
The European Union reached a decisive milestone in August 2024 with the entry into force of its artificial intelligence regulation. This legislation, the first of its kind globally, does not merely regulate AI systems within its territory: it outlines a model of governance through technical standards likely to transform regulatory practices far beyond its borders. But the ambition of global harmonization faces complex geopolitical and legal realities.
The European approach: an exportable model?
The European framework is based on a unique architecture that combines regulatory constraints and voluntary technical standards. The AI Act establishes a classification of AI systems according to their risk level: unacceptable, high, limited, or minimal. Specific requirements apply to each category.
What distinguishes the European approach is its reliance on “harmonized standards” developed by the standardization bodies CEN, CENELEC, and ETSI. These standards, once recognized by the European Commission, grant companies a presumption of conformity. In other words, complying with these technical standards is legally equivalent to fulfilling the legal obligations of the regulation.
| AI Act Risk Category | Main Requirements |
|---|---|
| Unacceptable | Prohibition |
| High | Strict compliance |
| Limited | Transparency |
| Minimal | Minimal obligations |
This strategy offers several advantages. It allows for progressive adaptation to technological developments, as standards can be revised more quickly than legislative texts. It also draws on the technical expertise of industrialists and researchers participating in standardization committees. Above all, it establishes bridges with international standards ISO, IEC, and IEEE, potentially facilitating their adoption outside Europe.
The Brussels Effect in action
The European single market, with its 450 million consumers, exerts considerable attraction. International companies wishing to access it must comply with European requirements, thus creating what researchers call the “Brussels Effect”. This dynamic has already worked for the GDPR, which has become a global benchmark for personal data protection.
In the field of AI, the mechanism could be repeated. A technology player developing a system compliant with European standards already has a competitive advantage in regulated markets. Rather than maintaining several versions of the same product, there is a strong temptation to adopt the most demanding standard as a global reference.
“The European Union holds a global normative leadership role in terms of digital regulation, restricting corporate freedom to increase individual freedom.”
Obstacles to international harmonization
Despite this potential for influence, convergence towards a global normative framework faces structural resistance. Divergent priorities among major powers constitute the primary major obstacle.
Traditionally, the United States favors an approach based on innovation and market self-regulation. The recent repeal by the Trump administration of the presidential decree on AI signed by Joe Biden illustrates this orientation towards deregulation. Conversely, China is developing a model centered on digital sovereignty and state control of strategic technologies.
Divergent conceptions of key concepts
Beyond regulatory philosophies, the very interpretations of fundamental notions vary considerably. What exactly is meant by “transparency” of an AI system? For Europe, this implies the explainability of automated decisions and the right to information for users. For other jurisdictions, transparency may be limited to disclosing the existence of an AI without detailing its operation.
The question of responsibility raises similar issues. Who is liable for damages caused by AI: the developer, the deployer, the end-user? National legal traditions profoundly influence these arbitrations, making it difficult for an international consensus to emerge.
The cost of normative fragmentation
For companies operating globally, the proliferation of incompatible regulatory frameworks represents a considerable burden. Developing, testing, and certifying AI systems under several distinct normative regimes substantially increases compliance costs.
This double conformity particularly affects small and medium-sized enterprises, which are less equipped than technology giants to navigate regulatory complexity. The risk is to create an additional barrier to entry, further concentrating the market in the hands of a few dominant players.
Fragmentation also threatens technical interoperability. If incompatible standards are imposed in different regions, AI systems developed for one market might not function correctly in another, hindering trade and collaborative innovation.
Levers for gradual convergence
Faced with these challenges, several mechanisms could foster gradual harmonization of technical standards on AI. International standardization organizations, particularly the ISO/IEC JTC1 SC 42 committee dedicated to artificial intelligence, constitute natural forums for convergence.
Europe actively participates in these efforts, seeking to infuse its ethical priorities into global standards. This strategy of influence through participation gradually integrates European requirements into international references, which other jurisdictions then adopt.
Bilateral agreements as bridges
Beyond multilateral bodies, bilateral mutual recognition agreements offer a pragmatic path to interoperability. Two jurisdictions can agree that their respective certification regimes are equivalent, allowing companies certified in one to operate in the other without additional procedures.
Canada, for example, has developed a regulatory approach with similarities to the European model through Bill C-27 and its Artificial Intelligence and Data Act (AIDA). Although this project is currently on hold, the shared conceptual foundations potentially facilitate future normative convergences.
Governance and participation challenges
Global harmonization of technical standards cannot succeed without ensuring equitable participation in standardization processes. Currently, the representation of developing countries in international technical committees remains limited, creating a risk that global standards primarily reflect the interests and constraints of advanced economies.
This question of legitimacy becomes crucial when these standards acquire a quasi-regulatory scope. If technical standards become the preferred way to regulate AI, their development must involve all stakeholders: states, businesses, civil society, and the scientific community.
Current geopolitical tensions complicate this inclusive ambition. Technological rivalries between major powers, national security concerns, and digital sovereignty strategies fragment spaces for international cooperation. Technology becomes a field of strategic competition as much as an object of shared regulation.
The role of technology companies
Large technology groups play an ambivalent role in this process. On the one hand, their technical expertise is essential for developing realistic and applicable standards. On the other hand, their influence on standardization processes raises questions about regulatory capture and the preservation of the public interest.
The balance between industry participation and standard independence is a constant challenge. Transparency mechanisms, rules for managing conflicts of interest, and the diversity of represented actors become decisive for the credibility of the standards produced.
Towards what global architecture?
Rather than complete harmonization, perhaps a multi-level architecture of AI standards should be considered. A common minimal foundation could establish shared definitions and fundamental principles, leaving jurisdictions the freedom to add specific requirements reflecting their values and priorities.
This “modular standards” model would allow for basic interoperability while preserving a certain regulatory diversity. AI systems compliant with the common core would circulate freely, while additional layers of certification would apply depending on the targeted markets.
The European experience with the AI Act will provide valuable lessons for refining this architecture. The first years of application will reveal the strengths and weaknesses of the technical standards regulation model, guiding necessary adjustments and potentially inspiring other jurisdictions.
For economic actors as well as regulators, the issue is no longer whether AI will be regulated, but how to ensure that this regulation promotes responsible innovation while preserving fundamental rights. Between inefficient fragmentation and rigid uniformity, the path of progressive and pragmatic convergence remains to be charted. The European approach, with its strengths and limitations, constitutes a full-scale experiment whose results will shape international debates in the coming years.
Recent developments, from the impact of AI copilots on software architecture to the transformation of the banking sector by generative AI, demonstrate that regulation must adapt to an unprecedented pace of innovation. The ability of normative frameworks to evolve as quickly as technology will determine their relevance and global adoption.
