Regulate AI through multiple competing jurisdictions with clear, enforceable rules rather than comprehensive global frameworks.
Lee Kuan Yew shows predictable rules attract more investment than perfect ones. Ostrom demonstrates overlapping authorities govern complex systems better than single regulators. Hayek proves markets discover AI risks faster than bureaucrats can anticipate them. Keynes argues tail risks require insurance mechanisms that markets underprice.
The council splits on intervention level but agrees on institutional design.
- 1Establish clear enforcement boundaries between local data privacy, national competition oversight, and international safety coordination authorities.
- 2Preserve regulatory exit options across jurisdictions to enable policy learning and prevent capture by incumbent firms.
- 3Create insurance mechanisms for AI tail risks that markets systematically underprice, particularly democratic and social stability threats.
AI companies need rules that work, not rules that sound good.
The 1997 Asian financial crisis taught me that global markets punish countries with weak institutions faster than they reward countries with good intentions. AI regulation will follow the same pattern. Companies will move to jurisdictions that combine predictable rules with competent enforcement. Singapore attracted multinational corporations not by having the most liberal policies, but by having the clearest ones. When we regulated financial services, we didn't copy London or New York wholesale — we adapted their frameworks to our institutional capacity and strategic needs.
Regulators who cannot distinguish between innovation that builds capability and innovation that creates systemic risk will lose both. The question is not whether to regulate AI, but whether your regulatory institutions can execute what they promise. Countries that over-regulate will drive investment away. Countries that under-regulate will face crises they cannot manage. The winners will be those that regulate precisely what they can enforce effectively.
Effective AI governance requires multiple overlapping authorities, not single comprehensive frameworks.
My decades studying commons governance revealed that complex systems are better managed by polycentric institutions than by single hierarchical authorities. AI governance faces the same challenge: no single regulator possesses the knowledge, legitimacy, and enforcement capacity to govern AI across all domains. When irrigation systems in Spain succeeded for centuries, it was because local water courts, regional authorities, and national laws created overlapping jurisdiction with clear boundaries. Each level governed what it could govern best.
AI development spans local data privacy, national security, international competition, and global safety concerns. Trying to house all of this in one regulatory framework guarantees either paralysis or capture. Instead, we need nested institutions: local authorities governing AI in municipal services, national regulators overseeing competition and safety, international bodies coordinating on global risks. The redundancy isn't inefficiency — it's resilience. When one level fails, others can adapt.
Central authorities cannot possess the knowledge required to regulate technological innovation effectively.
The knowledge problem I identified applies with special force to AI regulation. Innovation proceeds through the discovery of possibilities that cannot be known in advance. When bureaucrats in Brussels or Washington attempt to write rules for technologies they do not understand, applied to use cases they cannot foresee, they inevitably produce regulations that are either too broad (stifling beneficial innovation) or too narrow (missing genuine risks). The 1990s internet boom occurred precisely because no central authority was regulating it.
Markets discover both opportunities and dangers through the price system and competitive selection. Companies that build unsafe AI products will lose customers and capital. Companies that build valuable AI products will attract both. This process generates information about what works and what doesn't faster and more accurately than any regulatory committee. The regulators' role should be limited to enforcing property rights, preventing fraud, and ensuring competitive markets — not attempting to direct the innovation process itself.
Under genuine uncertainty about AI impacts, insure against worst cases rather than optimize for expected outcomes.
The AI governance challenge exemplifies what I called genuine uncertainty — situations where we cannot assign meaningful probabilities to future scenarios. We know AI will transform economic and social arrangements, but we cannot calculate the likelihood of specific outcomes. In such circumstances, the correct approach is not to optimize for the most likely scenario but to insure against the worst plausible ones. This is why Britain created the Bank of England's lender-of-last-resort function before we fully understood financial crises.
Hayek's market discovery process works well for normal technological innovation, but AI presents tail risks that markets systematically underprice. A social media algorithm that destabilizes democratic elections imposes costs that never appear in the company's profit-and-loss statement. The private returns to AI development exceed the social returns when negative externalities are large and uncertain. Government intervention is not about directing innovation — it is about ensuring that innovators internalize the full social costs of their decisions.
Regulatory frameworks should preserve exit options and accept that beneficial innovation requires initial ambiguity.
My hiding hand principle applies directly to AI governance: if regulators knew in advance how difficult it would be to govern AI effectively, they might not start. But this ignorance can be productive. The internet's early development benefited from regulatory ambiguity that allowed experimentation. Similarly, AI governance should preserve reversibility — the capacity to adjust rules as we learn what works and what creates genuine harm.
The voice-and-exit framework suggests that effective AI governance requires multiple jurisdictions with different approaches. Companies and users need exit options when regulatory frameworks become counterproductive. Singapore's approach will differ from California's, which will differ from Estonia's. This diversity is not a problem to be solved through harmonization — it is a feature that allows policy learning. The jurisdictions that get AI governance right will attract investment and talent. The jurisdictions that get it wrong will face exit pressures that force correction.
The convergence note
Where the council converges: Innovation requires predictable rules, not perfect ones. Regulatory uncertainty can be more harmful than imperfect regulation.
Where it divides: LKY and Keynes believe active government intervention is necessary for managing AI risks. Hayek argues markets discover risks and opportunities more effectively than regulators. Ostrom and Hirschman favor multiple experimental approaches over unified frameworks.
What only the policymaker can resolve: Whether to prioritize innovation speed or safety margins when both matter but cannot be maximized simultaneously.