After exploring the mechanisms of algorithmic manipulation and strategies for defending against disinformation, it is time to address the fundamental question: how to effectively regulate AI on a global scale? Faced with a fragmented and rapidly evolving regulatory landscape, this article maps existing approaches and outlines paths towards balanced governance.

Introduction: The Challenge of Regulation Without Borders

Artificial intelligence knows no borders. A model developed in San Francisco can be deployed in Singapore, cause harm in Europe, and be trained on data hosted in Canada. Yet, regulatory responses remain stubbornly national or regional.

The proliferation of regulatory announcements makes navigating this rapidly evolving landscape extremely difficult. Understanding AI regulation increasingly means comparing apples and oranges. On one hand, the UK adopts a "pro-innovation approach" with no legally binding requirements. On the other, the EU imposes fines of up to €35 million or 7% of global turnover.

This fragmentation creates three major challenges:

1. Semantic ambiguity: the term "AI regulation" covers radically different realities depending on the jurisdiction
2. International fragmentation: divergent frameworks risk compromising cooperation and legal interoperability
3. Unequal access: an opaque understanding of what regulation implies can hinder civil society participation in the legislative process

Faced with these challenges, how can we design effective regulation that preserves innovation while protecting fundamental rights?

1. Overview of Regulatory Models Worldwide

🇪🇺 The European Model: Risk-Based Dominant Legislation

The European Union led the way with its Regulation on Artificial Intelligence (AI Act) , adopted in 2024, which constitutes the first comprehensive and horizontal legislation in the world. Its approach is based on a risk-based classification:

• Unacceptable risk: social scoring systems or real-time facial recognition for mass surveillance → outright prohibition
• High risk (medical diagnosis, credit scoring, recruitment) → strict obligations for transparency, security, traceability, and data quality
• Limited risk (chatbots) → reduced transparency obligations
• Minimal risk (voice assistants) → no specific constraints

Sanctions are deterrent: up to €35 million or 7% of global turnover for violations of rules on prohibited practices.

This model has inspired other Asian countries. South Korea, after favoring flexible approaches, adopted its Framework Act on AI in December 2024, which is directly inspired by the European model with risk classification and enhanced transparency obligations for generative AI.

The United Kingdom, with its Online Safety Act effective March 2025, adopts a service-category approach, imposing differentiated obligations based on platform size and nature. Fines can reach £18 million or 10% of global turnover.

🇺🇸 The American Model: Market and Self-Regulation

The United States adopts a radically different approach, marked by the absence of specific federal legislation and high political volatility. The Biden administration issued an executive order in 2023 on the "safe, secure, and trustworthy development of AI," but it was revoked by the Trump administration in January 2025.

The American model rests on four pillars:

1. Good practice guides: the NIST AI Risk Management Framework provides voluntary recommendations to reduce risks and biases
2. Administrative initiatives: despite the revocation of Biden's order, some federal agencies maintain sectoral requirements
3. Industrial self-regulation: organizations like the Association for the Advancement of AI establish ethical standards that their members commit to following
4. Application of existing law: the FTC has already sanctioned companies for AI-related violations, such as the $25 million fine imposed on Amazon for non-compliance with child data protection with Alexa

Meanwhile, a mosaic of legislation is emerging at the state level: California, Utah, Arkansas, Maryland, Illinois, and others have adopted their own laws on specific aspects of AI.

🇨🇳 The Chinese Model: Centralized Control and National Security

China has developed a pragmatic and reactive approach, with a series of texts targeting specific technologies:

• Measures for the Management of Generative AI Services (2023)
• Provisions on Algorithmic Recommendation Services (2022)
• Regulations on the Administration of Deep Synthesis (Deepfakes) Internet Information (2022)

These texts are part of a broader framework including the Cybersecurity Law and the Personal Information Protection Law. The Chinese approach is characterized by strong state control, with content requirements aligned with fundamental socialist values and close oversight of service providers.

🇯🇵 The Japanese Model: Ethics and Flexibility

Japan favors an ethical and non-binding approach, centered on promoting shared principles rather than legal obligations:

• Social Principles of Human-Centric AI (2019): seven fundamental principles including human-centricity, education, privacy protection, safety
• Guidelines for AI Operators (2024): detailed framework for developers, providers, and users
• Act on Promoting Research, Development, and Application of AI-related Technologies (2025): programmatic text without sanctions, defining strategic orientations

Japan relies on "governance by principles" , where adherence to ethical values is supposed to naturally guide actors. In case of breach, market mechanisms (reputation) and common law apply.

🇨🇦 The Canadian Model: In Transition to Legislation

Canada was a pioneer with its Pan-Canadian AI Strategy (2017), the world's first national strategy. Since then, the country has developed voluntary guides such as the Voluntary Code of Conduct for the Responsible Development of Advanced Generative AI Systems (2023).

However, the Artificial Intelligence and Data Act (AIDA) , proposed in 2022, has not yet been adopted. Canada is therefore in a transition phase, with an approach still largely non-binding but an expressed intention to move to a legislative model.

2. Common Challenges and Blind Spots of Regulation

Comparative analysis of existing frameworks reveals several cross-cutting challenges.

🧩 The Four Major Challenges Identified by Research

A systematic study of over 40 international regulatory documents identified four main challenges in implementing AI governance frameworks:

1. The human supervision paradox: compliance requirements regarding human control conflict with the autonomous nature of AI systems. How can a human effectively supervise a system that makes decisions at a speed and scale exceeding their capabilities?
2. The lack of operational guidance: despite detailed structural requirements, regulatory frameworks rarely provide sufficient guidance for concretely implementing risk assessments.
3. The absence of real-time monitoring systems: copyright compliance for training data relies entirely on reactive enforcement, with no continuous control mechanism.
4. The fragmented enforcement architecture: a single AI decision can simultaneously violate multiple regulations, creating a cumulative exposure to sanctions that organizations struggle to anticipate.

📊 Key Dimensions for Comparing Regulations

A recent taxonomy proposes eleven metrics to analyze and compare AI regulations:

• Adoption status
• Novelty of the framework
• Maturity of the digital legal landscape
• Scope (extraterritorial or not)
• Enforcement mechanisms
• Sanctions provided
• Operationalization
• International cooperation
• Stakeholder consultation
• Regulatory approach (technology vs. application)
• Focus

This analysis grid allows moving beyond simplistic comparisons and identifying strengths and weaknesses of each approach.

⚖️ The Tension between Freedom of Expression and Combating Disinformation

The Council of Europe, through its European Court of Human Rights, has developed nuanced case law on the issue of disinformation. The European doctrine of freedom of expression is based on a democratic justification: freedom of expression is essential to the functioning of a democratic society.

This approach leads to fundamental differences with the United States, where the First Amendment doctrine makes any legal action against disinformation much more difficult. As experts point out, the challenge posed by disinformation lies precisely in the fact that digital technologies and platforms have significantly subverted the traditional mechanisms of the public sphere.

3. Paths to Effective Regulation

Faced with these challenges, what paths are emerging for balanced and effective regulation?

1️⃣ Hybrid Approach: Combining Legislation and Self-Regulation

Analysis of emerging practices shows that the binary opposition between state regulation and self-regulation is outdated. The most promising models combine:

• A binding legislative foundation for the most serious risks (like prohibiting social scoring in the AI Act)
• Market and self-regulation mechanisms for areas where rapid innovation demands flexibility
• Positive incentives to encourage the adoption of good practices beyond the legal minimum

South Korea illustrates this evolution, moving from a flexible approach to binding legislation while retaining spaces for flexibility. The United Kingdom, with its Online Safety Act, also combines legal obligations and market mechanisms like "regulatory sandboxes."

2️⃣ International Harmonization and Interoperability

Current regulatory fragmentation creates considerable compliance costs and legal uncertainties. Several paths to harmonization are emerging:

• Mutual recognition: agreements between jurisdictions to recognize the equivalence of their respective standards
• Common international standards: the work of ISO (notably ISO/IEC 42005:2025) provides a technical basis for harmonization
• Common principles: the OECD AI Principles and UNESCO Recommendations on AI Ethics offer a foundation of shared values

The G7 Hiroshima AI Process initiative illustrates this desire for convergence, with the adoption of international guiding principles for advanced AI systems.

3️⃣ Regulation through Transparency and Data Access

Rather than seeking to directly control AI content, a promising approach consists of imposing transparency obligations and granting data access to researchers and civil society.

The European Digital Services Act requires very large platforms to grant accredited researchers access to their data. This approach allows democratic control without prior censorship.

Tools like AdAnalyst or CheckMyNews (developed for Meta and YouTube) show that it is possible to collect research data while respecting privacy.

4️⃣ Human Rights-Based Regulation

The United Nations Secretary-General's report insists on a crucial point: the fight against disinformation must not come at the expense of fundamental freedoms. Key principles include:

• Protecting, respecting, and promoting freedom of expression
• Avoiding regulation based on vague definitions
• Refraining from internet shutdowns and website blocking
• Involving civil society in policy design

This approach, advocated by the Council of Europe, reminds us that remedies against disinformation can sometimes be worse than the evil they claim to fight.

5️⃣ Regulation through Education and Societal Resilience

Finally, no regulation will be effective without a massive effort in media and information literacy. Research shows that prebunking (preventive inoculation) strategies can be very effective.

Singapore, in its analysis of anti-disinformation measures, identifies three main categories of government actions:

1. Legislation and regulation
2. Specialized task forces
3. Public education

Explicit and consistent communication with the public, as well as strengthening information culture from school onwards, are indispensable complements to purely regulatory approaches.

Conclusion: Towards Variable Geometry Governance

Effective AI regulation will not come from a single model, but from an intelligent combination of approaches adapted to the cultural, legal, and political contexts of each region.

Comparative analysis of existing frameworks reveals that the fusion of models is underway:

• Legislative approaches (Europe, Korea) integrate elements of flexibility and market
• Market approaches (United States) are developing control mechanisms via existing law
• Ethical approaches (Japan) are gradually acquiring legislative bases

The way forward consists of building a multi-level governance system:

• At the top: shared ethical principles at the international level
• At the intermediate level: market and self-regulation mechanisms for rapid innovations
• At the base: a binding legislative foundation for serious risks, with deterrent sanctions

Only under this condition can we build AI governance that is protective of fundamental rights, conducive to innovation, and adapted to the global and fluid nature of this technology.