AI safety is a global challenge that transcends borders—dangerous AI developed anywhere threatens everyone everywhere. Yet the world is organized into sovereign states, each claiming supreme authority within its territory. This creates a fundamental tension: effective AI safety governance may require circumscribing the very sovereignty that constitutes the international system. Understanding sovereignty—its origins, its logic, and how it has been limited before—is essential for designing AI safety governance that can actually work.

The Sovereignty Problem for AI Safety

AI safety governance faces a structural challenge: there is no global sovereign. The international system consists of states, each claiming "supreme authority within a territory." No higher authority exists that can make and enforce rules for all.

This creates several problems:

  • Coordination failure: each state decides its own AI policies, potentially racing toward unsafe development
  • Enforcement gaps: no global authority can enforce safety standards against unwilling states
  • Regulatory arbitrage: unsafe AI development can relocate to permissive jurisdictions
  • Information asymmetry: states may hide AI capabilities from external oversight

The very structure of the international system—the sovereignty of states—may make adequate AI safety governance impossible.

What Is Sovereignty?

Sovereignty, as it emerged in early modern Europe and spread globally, has three components:

1. Authority (Not Just Power)

Sovereignty is not merely coercive power—the ability to force compliance. It is authority: "the right to command and correlatively the right to be obeyed." Sovereignty claims legitimacy, deriving from some acknowledged source—natural law, divine mandate, constitution, or international recognition.

For AI safety: a global AI governance body would need authority, not just power. But where would its authority come from?

2. Supremacy

The sovereign holds authority superior to all other authorities within its domain. The U.S. Constitution is superior to Pennsylvania law; the sovereign state is superior to subnational authorities within its territory.

For AI safety: this suggests any effective global governance would need supremacy over national AI policies—but this would directly contradict state sovereignty.

3. Territoriality

Authority is defined by geographic boundaries. People belong to a state by virtue of location within its borders, regardless of identity, kinship, or consent.

For AI safety: territoriality is problematic. AI development, compute resources, talent, and information flow across borders. Territorially-defined authority may be inadequate for globally-networked technology.

Internal and External Sovereignty

Sovereignty has two dimensions:

Internal sovereignty: supreme authority within borders. The state's prerogative to govern its territory without internal challenge.

External sovereignty: recognition by other states of independence from outside interference. "Constitutional independence"—freedom from external influence on basic prerogatives.

AI safety governance challenges both:

  • Internal: AI safety regulation may require external oversight of internal development—contradicting internal sovereignty
  • External: international AI agreements require states to accept external constraints on their AI policies—contradicting external sovereignty

Historical Circumscription of Sovereignty

Yet sovereignty is not absolute. It has been circumscribed before in ways relevant to AI safety.

1. Human Rights

The Universal Declaration of Human Rights (1948) and subsequent covenants committed states to respecting individual rights within their territories. While initially aspirational, human rights law has evolved:

  • Legal commitments: states legally bound to respect civil, political, economic, and social rights
  • Judicial oversight: regional courts (European Court of Human Rights) can rule against states
  • Enforcement mechanisms: international pressure, sanctions, even intervention for severe violations

The key insight: states accepted limits on their sovereignty in exchange for membership in an international community that respects certain norms.

2. The Responsibility to Protect (R2P)

R2P, adopted in 2005, reconceptualizes sovereignty from control to responsibility:

  • States have a responsibility to protect their populations from genocide, war crimes, ethnic cleansing, and crimes against humanity
  • If a state fails to protect its population, the international community has a responsibility to assist
  • If a state is manifestly failing to protect, the international community must take timely and decisive action, including military intervention if necessary

This inverts sovereignty: rather than shielding states from external interference, sovereignty becomes conditional on fulfilling responsibilities. Failure to protect can justify overriding sovereignty.

Application to AI safety: could catastrophic AI risk justify overriding sovereignty to prevent development of dangerous systems? The R2P model suggests: if states fail to protect their citizens (and the world) from AI risks, outside intervention could be legitimate.

3. European Integration

The European Union demonstrates that states can voluntarily pool sovereignty:

  • Supranational institutions: European Commission, Court of Justice, Parliament have authority over member states
  • Policy areas: trade, competition, currency (Eurozone), human rights, some foreign policy
  • Legal supremacy: EU law can override national law in covered areas

States accepted limits on sovereignty because benefits of coordination exceeded costs of autonomy. This may be the most promising model for AI safety: voluntary pooling of sovereignty where the benefits of coordination (preventing catastrophic AI) exceed the costs (constraining national AI policy).

4. International Criminal Law

The International Criminal Court can prosecute individuals for genocide, crimes against humanity, and war crimes—regardless of their official position. Even heads of state are not immune. This establishes that some actions exceed sovereign prerogatives.

Application to AI safety: could individuals who develop and deploy catastrophically dangerous AI be held criminally liable under international law?

The Challenge of Non-Absolute Sovereignty

These precedents show sovereignty can be limited. But each circumscription faces resistance:

  • Human rights: states resist enforcement, claim cultural exception, withdraw from jurisdictions
  • R2P: contested applications, accusations of selective enforcement, great power vetoes
  • EU integration: Brexit, rise of Euroskepticism, democratic deficit critiques
  • ICC: non-membership by major powers, accusations of bias, withdrawal threats

The lesson: sovereignty can be limited, but not easily or permanently. States resist constraints on their autonomy, especially powerful states.

Implications for AI Safety Governance

1. Recognize the Sovereignty Constraint

Effective AI safety governance must work with, not against, the sovereignty structure of the international system. Proposals that assume away sovereignty are unrealistic.

This means:

  • Voluntary cooperation as the primary mechanism
  • Incentives for participation (benefits of coordination, costs of exclusion)
  • Gradual pooling of sovereignty where benefits are clear
  • Residual sovereignty retained in other areas

2. Build on R2P Logic

The R2P model—sovereignty as responsibility—suggests:

  • State responsibility: states have a responsibility to ensure AI developed in their territory is safe
  • International oversight: if states cannot or will not ensure safety, international mechanisms may intervene
  • Last resort: only when states manifestly fail would intervention be legitimate

Challenges: defining "manifest failure," gaining international consensus, avoiding abuse.

3. Create Benefits for Sovereignty Pooling

The EU model succeeded because states gained from coordination. For AI safety:

  • Shared safety infrastructure: joint research, evaluation facilities, expertise
  • Economic benefits: access to coordinated markets, reduced duplication
  • Security benefits: protection from unsafe AI developed elsewhere
  • Reputation benefits: legitimacy from participation in responsible AI governance

States must see participation as beneficial, not merely constraining.

4. Address Free-Riding and Defection

Sovereignty enables states to opt out of agreements, free-ride on others' restraint, or defect when convenient. Countermeasures:

  • Conditional benefits: participation required for access to certain technologies, markets, or institutions
  • Transparency requirements: verification mechanisms that detect defection
  • Reputational costs: naming and shaming defectors
  • Last resort enforcement: sanctions or other consequences for severe violations

5. Design for Variable Geometry

Not all states will participate equally. Design governance for:

  • Core participants: leading AI states with strongest commitments
  • Associate members: states accepting some but not all obligations
  • Observers: states monitoring but not bound
  • Non-participants: states outside the framework

Governance must work even without universal participation, while creating incentives for joining.

6. Prepare for Sovereignty Assertions

States periodically reassert sovereignty against perceived external constraints. AI safety governance should:

  • Build domestic support: legitimacy within states, not just internationally
  • Preserve flexibility: allow states to adapt commitments to domestic contexts
  • Provide exit options: clear procedures for withdrawal, reducing incentive for unilateral defection
  • Demonstrate value: continually show that participation benefits exceed costs

The Democratic Deficit Question

Pooling sovereignty raises democratic questions: who governs the governors? If states transfer authority to international AI safety bodies, how are those bodies accountable?

Options:

  • Intergovernmental: states retain control through councils of ministers (UN model)
  • Supranational: independent institutions with direct authority (EU model)
  • Hybrid: combination of state control and independent institutions

Each has trade-offs. Intergovernmental preserves sovereignty but may be slow and gridlocked. Supranational is effective but raises democratic deficit. Hybrid attempts balance but is complex.

The Exception Problem

Carl Schmitt argued that "sovereign is he who decides on the exception"—the emergency that suspends normal rules. AI safety governance faces this problem:

  • What happens when a state claims AI safety requirements don't apply due to national emergency?
  • Who decides what counts as an exception?
  • Can international governance override national emergency claims?

This may be the hardest sovereignty challenge: the final authority to suspend rules in emergencies.

Conclusion

Sovereignty is not an absolute barrier to AI safety governance—but it is a profound constraint. The international system's structure—states claiming supreme authority within territories—creates fundamental challenges for governing a globally-networked technology with catastrophic potential.

Historical precedents show sovereignty can be circumscribed:

  • Human rights law limits what states can do to their own people
  • R2P makes sovereignty conditional on protecting populations
  • European integration shows states can voluntarily pool sovereignty
  • International criminal law holds individuals accountable regardless of position

But each circumscription faces resistance, especially from powerful states. AI safety governance must:

  • Work with sovereignty, not assume it away
  • Create benefits for participation that exceed costs of constraint
  • Build on R2P logic—sovereignty as responsibility
  • Address free-riding, defection, and variable participation
  • Prepare for sovereignty reassertion
  • Solve the democratic deficit and exception problems

The goal is not to eliminate sovereignty but to reshape it: from absolute authority to responsible authority, from control within borders to participation in global coordination, from freedom from external constraint to acceptance of mutual obligations.

Effective AI safety governance requires sovereignty transformed, not sovereignty transcended.

References

  • Bodin, Jean (1576). Les Six livres de la République.
  • Hobbes, Thomas (1651). Leviathan.
  • International Commission on Intervention and State Sovereignty (2001). The Responsibility to Protect.
  • Philpott, Daniel (2001). Revolutions in Sovereignty. Princeton University Press.
  • Stanford Encyclopedia of Philosophy (2023). "Sovereignty." https://plato.stanford.edu/entries/sovereignty/