Beyond Regulatory Imitation: A Global Framework for AI Governance

Executive judgment

Global AI governance has a representation problem disguised as a technical debate. The most influential proposals are produced in jurisdictions that possess advanced regulators, mature capital markets, high institutional capacity and incumbent technology firms. Their concerns—frontier safety, privacy, copyright, labour displacement, market concentration and democratic integrity—are real. They are not the complete global risk register.

For much of the world, the costs of insufficient capability are immediate: public services that cannot reach citizens in their language, scarce professional expertise, low agricultural productivity, limited access to health and education, weak administrative data, expensive compute and dependence on foreign infrastructure. A governance framework that counts the harms of deployment but treats forgone development as neutral will systematically prefer the status quo of wealthy societies.

Zwarte Peper’s position: AI should develop and diffuse rapidly. Governance should establish a universal floor of rights, security and accountability, then allow jurisdictions to choose proportionate, sector-specific instruments grounded in their development needs and evidence. International rules should protect interoperability and representation, expand access to compute, data and knowledge, and avoid turning the regulatory preferences of incumbent powers into barriers for everyone else.

This is not an argument for relativism or a race to the bottom. Universal human rights do not become Western because European institutions enforce them. Nor does development justify unreviewable surveillance or automated exclusion. The argument is about completeness and institutional fit: legitimate governance must address both misuse of AI and exclusion from its benefits.

1. The geography of rule-making shapes the definition of risk

UN Trade and Development reported in 2025 that 118 countries—mostly in the Global South—were absent from major AI-governance discussions, while fewer than one third of developing countries had national AI strategies.¹ This is not merely a diplomatic deficit. It changes which facts enter standards, which languages are tested, which institutional costs are assumed and which opportunity costs are visible.

Rules developed in high-capacity administrations may assume accessible regulators, reliable identity systems, specialist auditors, abundant lawyers and firms capable of maintaining extensive documentation. Exported unchanged, those rules can produce nominal protection but little enforcement. They can also raise fixed costs, reserve capable systems for large foreign suppliers and divert scarce public expertise from concrete harms to formal compliance.

Representation cannot be cured by consultation after a framework is substantially settled. Jurisdictions need agenda-setting power, technical participation and evidence-generating capacity. They must be able to show how systems perform in multilingual, informal, rural, low-connectivity and institutionally diverse environments—and to shape standards in response.

2. Comparative models reveal different public priorities

No jurisdiction has solved AI governance. Comparative analysis is nevertheless useful because each model makes different trade-offs visible.

India: governance as an enabler of diffusion

India’s November 2025 AI Governance Guidelines are grounded in seven principles, including “Innovation over Restraint”, fairness, accountability and understandable design. They favour an evidence-led, proportionate and techno-legal approach, rely substantially on existing and sectoral law, and do not propose a new horizontal AI statute at this stage.² The distinctive contribution is not absence of safeguards; it is the proposition that governance should actively enable adoption and inclusive development.

ASEAN: interoperability without premature uniformity

The ASEAN Guide on AI Governance and Ethics provides a voluntary organisational framework that can be tailored to sector, system complexity and risk. Its expanded generative-AI guide addresses accountability, data, trusted development, incident reporting, testing, security, content provenance, safety research and AI for public good.³ This is regulatory cooperation through practical compatibility rather than immediate legislative identity.

African Union: capacity is part of governance

The African Union’s Continental AI Strategy expressly adopts an Africa-centric and development-focused approach. It connects ethical governance with infrastructure, skills, research, locally relevant data, private-sector innovation, cultural renewal and regional cooperation.⁴ Its importance lies in rejecting a false sequence in which countries must first import a complete safety regime and only later build the capability to benefit from AI.

3. Universal rights are a floor, not a complete operating model

A global framework needs a common normative base. The UNESCO Recommendation on the Ethics of Artificial Intelligence—adopted by 193 member states in 2021—places human rights, dignity, fairness, transparency, human oversight, diversity and environmental responsibility at its centre.⁵ These commitments provide a legitimate floor because they arise from a global intergovernmental process and connect AI governance to established international law.

The floor should prohibit or tightly control uses incompatible with human dignity and lawful government; require responsibility for consequential deployment; protect privacy and security; preserve effective remedy; and demand non-discrimination. But a statement of principles does not determine the same administrative mechanism everywhere. A sophisticated conformity-assessment system, sectoral regulator, judicial remedy, licensed professional regime and public-procurement control may each deliver accountability in different contexts.

International alignment should therefore focus on outcomes and evidence before identical procedure. Systems should be able to demonstrate security, traceability, contextual performance and accountable use through mechanisms credible in the relevant legal order. Mutual recognition and equivalence can support cross-border commerce without forcing lower-capacity jurisdictions to replicate the largest rulebook.

4. Development is not an exception to risk analysis

Regulatory impact assessment often compares a proposed AI system with an ideal human process. The proper counterfactual is the actual system: shortages, delay, unaffordable expertise, informal discretion, linguistic exclusion and existing error. AI can introduce new harms; it can also reduce these baseline harms.

A health-support tool in a clinician-rich system and one serving a region with severe professional scarcity present different benefit-risk equations. A multilingual service assistant may be imperfect yet materially improve access for people who currently receive no intelligible guidance. Automated translation can distort cultural meaning, but a policy that supports only the highest-resource languages entrenches a different exclusion.

Governance should therefore measure the risk of action and inaction. Where error is reversible and benefit substantial, controlled deployment with monitoring may be preferable to waiting for comprehensive proof. Where a system can deprive liberty, essential income or legal status, stronger evidence and remedy are required. Proportionality must apply to regulatory delay as well as technological intervention.

5. Data, language and knowledge determine whose world the system represents

Global representation is not achieved by translating an English interface. Models encode the availability, classification and authority of knowledge. Important material may be oral, dispersed, written in low-resource languages or governed by community norms that conventional data pipelines ignore.

UNESCO’s global observatory reports that only 31 per cent of surveyed countries have training data in all official languages and 15 per cent in indigenous languages; policies directed to indigenous or minority languages remain rare.⁶ These gaps affect performance and power. A language absent from evaluation can become invisible to procurement. A community whose concepts do not fit a benchmark can be labelled low quality rather than poorly represented.

Public investment should support language datasets, evaluation suites, research institutions and models that markets underprovide. Governance must address provenance, consent, collective interests and benefit sharing without creating property claims so broad that knowledge becomes unusable. Local experts should define material errors and acceptable performance; external benchmarks should be treated as starting points, not neutral truth.

6. Regulatory power and market power must be analysed together

AI infrastructure is highly concentrated. Compute, cloud distribution, frontier models and key developer platforms are controlled by a small number of firms and jurisdictions. Regulation can constrain that power, but complex ex ante duties can also entrench it by increasing the capital, legal and data resources required to compete.

A pro-development framework should preserve access to open models and research, support switching and interoperability, scrutinise exclusionary cloud and distribution practices, and make public support contestable. Safety obligations should attach to specific capability and control, not merely organisational size; exemptions should not create unaccountable harm, but neither should thresholds become permanent licences for incumbents.

UNCTAD’s 2026 work on science and innovation calls for international cooperation to widen access to skills, data and computing resources, and notes the importance of open science and open innovation for developing countries.⁷ Access is not charity. More geographically distributed research and deployment creates evidence, competition and resilience for the global system.

7. A modular architecture for global AI governance

This architecture avoids two extremes: a single global regulator too distant from context, and fragmentation that leaves cross-border systems without common expectations. The global layer should address truly global problems—frontier capability evidence, severe incident coordination, technical standards, infrastructure concentration and access. National and sectoral institutions should govern deployment and remedy.

8. Five institutional reforms for genuine representation

1. Representation tied to decision rights. Developing countries need seats, agenda-setting power, technical staff and resources in international standard-setting and scientific processes.
2. A distributed evidence network. Publicly supported evaluation centres should test languages, sectors and conditions underrepresented in frontier laboratories, using common methods while retaining local definitions of harm.
3. Shared incident infrastructure. A federated system should allow regulators to exchange severe-incident signals without requiring disclosure of personal, security-sensitive or commercially protected information beyond necessity.
4. Compute and knowledge access. Multilateral finance, regional facilities, open research and contestable public procurement should expand the capacity to build and adapt—not only consume—AI.
5. Regulatory additionality review. Major international rules should disclose their compliance cost, effect on entry, administrative requirements and expected benefit for lower-capacity jurisdictions.

The UN Secretary-General’s advisory body proposed a globally inclusive and distributed architecture, an international scientific panel, policy dialogue, capacity development and a global fund.⁸ These proposals will be credible only if participation changes outcomes rather than legitimising decisions made elsewhere.

9. The legitimacy test

A proposed AI rule should answer six questions:

• What concrete harm or coordination failure does it address?
• Which actor controls that risk, and does the duty attach to that actor?
• What benefit, access or learning will be delayed—and for whom?
• Can the outcome be demonstrated through a less restrictive or more context-sensitive mechanism?
• What administrative, technical and financial capacity does implementation require?
• Did affected regions shape the rule, evidence and revision process with actual decision power?

A framework that cannot answer these questions may still be politically influential. It should not be described as globally legitimate.

Conclusion: a rules-based order without regulatory hierarchy

The alternative to Western regulatory dominance is not a world without rules. It is a better rules-based order: one in which universal rights constrain power, evidence travels across borders, jurisdictions retain room to pursue development and institutions outside the traditional centres participate as authors.

India’s pro-innovation techno-legal approach, ASEAN’s practical interoperability, the African Union’s development-centred strategy and UNESCO’s global rights framework each contribute something the dominant debate lacks. None should become a new template imposed everywhere. Together they show that safety, diffusion, capacity and dignity belong in the same analysis.

AI governance will be legitimate when it expands humanity’s capability while making consequential power answerable. That requires rapid innovation, open opportunity and rules capable of seeing the whole world.