The sovereignty objection to international AI governance runs as follows. Decisions about which technologies a country may develop and deploy are domestic decisions, made by national governments accountable to their own citizens. A binding international treaty restricting AI development would transfer effective authority over those decisions to an international body, or to the veto of other countries. This infringes sovereignty. The argument is stated as though it were original to the AI debate. It is not. Identical arguments were made against the NPT, the Chemical Weapons Convention, the Montreal Protocol, and every other major international treaty that constrained states' freedom of action over dangerous technology. In each case, the argument was eventually overcome. Understanding how is useful.
What the sovereignty argument actually claims
The argument is strongest when applied to military technology. Countries facing real security threats have a concrete interest in maintaining freedom of action over their military capabilities. Binding them to international restrictions on AI development in military contexts requires accepting constraints that adversaries might not accept, in a domain where capability matters directly for national survival. This is the hardest version of the sovereignty objection, and it mirrors the position of nuclear-armed states throughout the Cold War.
The argument is considerably weaker when applied to civilian AI development. No country's national security depends on its ability to run unmonitored commercial AI training. The sovereignty cost of declaring and monitoring large-scale civilian AI development programs is low. The political use of the sovereignty argument frequently extends far beyond the cases where the underlying concern is genuinely legitimate.
Separating the cases matters for treaty design. Governance frameworks that address the specific activities where sovereignty concerns are real, large-scale AI development programs with potential for civilizational-scale impact, can be designed to leave civilian and commercial AI largely untouched, substantially reducing the strength of the objection.
How the nuclear sovereignty argument was resolved
The NPT did not eliminate the sovereignty objection. It reframed it. The treaty created a formal architecture of sovereign choice: countries could join and accept safeguards obligations, or they could remain outside the treaty and face the political and economic consequences of being outside the international system. The five declared nuclear-weapon states were permitted to retain their weapons in exchange for accepting obligations around non-proliferation and disarmament. Non-nuclear states accepted monitoring of their civilian nuclear programs in exchange for guaranteed access to peaceful nuclear technology and security assurances.
The mechanism that made this work was that sovereignty was preserved in form while being substantially constrained in practice. NPT signatories can in principle withdraw from the treaty — Article X allows withdrawal with 90 days' notice, as North Korea invoked in 2003. But the political and economic cost of withdrawal is substantial. The sovereignty argument said countries could not accept external constraints on their nuclear choices. The NPT resolved this by making the choice formal and visible, while making the cost of the wrong choice high enough that almost every country concluded joining was preferable.
The result was a treaty that today has 191 states parties. Nine countries have nuclear weapons. Only three countries, India, Pakistan, and Israel, have never joined the NPT. North Korea withdrew and now has weapons. The sovereignty objection, which seemed powerful enough to prevent any binding nuclear governance in the early 1960s, produced a world in which 191 governments accepted binding monitoring obligations for their nuclear programs.
How the chemical weapons sovereignty argument was resolved
The Chemical Weapons Convention (1993) presents a different model. Unlike the NPT, which creates a formal two-tier world of nuclear haves and have-nots, the CWC bans chemical weapons for everyone, including the major powers that negotiated it. The United States and Soviet Union, which had the largest chemical weapons stockpiles in the world, accepted binding obligations to declare and destroy them under international supervision.
The sovereignty accommodation in the CWC is the right of national chemical industries to continue operating for civilian purposes, subject to monitoring. Countries can maintain chemical industries. They cannot maintain chemical weapons. The Organisation for the Prohibition of Chemical Weapons inspects declared facilities and investigates allegations of violation. Countries retain sovereignty over their civilian chemical sectors; they have surrendered sovereignty over one specific application of those sectors.
This model may be more applicable to AI than the NPT model, because AI development is predominantly civilian and dual-use. A treaty that restricted the development of the most capable frontier AI systems — those above specified capability or compute thresholds — while leaving commercial and consumer AI entirely unregulated would follow CWC logic rather than NPT logic. The sovereignty cost of this kind of treaty is lower than the NPT's because it does not create a permanent hierarchy of permitted and prohibited capabilities based on who had what when the treaty was signed.
The ozone layer model and scientific consensus
The Montreal Protocol on Substances that Deplete the Ozone Layer (1987) resolved the sovereignty objection through a different mechanism: scientific consensus made non-participation politically untenable. Once it became clear that chlorofluorocarbons were causing measurable damage to the ozone layer, and that ozone depletion would cause measurable harm to every country on earth, the sovereignty argument against joining the protocol became difficult to sustain in domestic politics. Governments that invoked sovereignty to avoid CFC restrictions faced the question of what sovereignty they were protecting when the consequence was harm to their own citizens from ultraviolet radiation.
The Montreal Protocol also included a technology transfer mechanism. Developing countries that would bear the economic cost of phasing out CFCs received assistance, both financial and technological, to facilitate compliance. This addressed the version of the sovereignty argument that international environmental agreements burden developing countries to pay for problems created by wealthy ones. By 2009, the Montreal Protocol had achieved universal ratification: 197 parties, every UN member state.
The AI parallel is imperfect because the scientific consensus on AI existential risk is not as unified as the ozone science was by 1987. But the Montreal Protocol model suggests that a clearer and more public expert consensus on AI risk would change the political dynamics of the sovereignty objection significantly. The more concretely the risk is understood, the less sustainable the sovereignty argument becomes.
What remains genuinely difficult
The sovereignty objection is not simply an excuse for governments to avoid inconvenient obligations. It reflects real political constraints. Governments answer to domestic constituencies who expect them to maintain national capability and freedom of action. Treaties that require a country to accept constraints that rivals have not accepted face sustained domestic opposition, as the SALT II treaty did in the United States after Soviet intervention in Afghanistan in 1979, when the Senate declined to ratify a treaty that was already signed.
For AI, the constraint is sharpest in the US-China relationship. An AI governance treaty that binds the United States but not China, or that China ratifies without meaningful enforcement, would function as a unilateral constraint. This concern is frequently raised under the banner of sovereignty, but it is more precisely a participation problem. The sovereignty objection can be addressed by treaty design. The participation problem can only be addressed by diplomacy, by creating the conditions under which both the US and China conclude that joining a binding framework is preferable to remaining outside one.
"The sovereignty argument against arms control has been made before every major treaty in history. It has never been the final word."
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
History suggests that the sovereignty objection to AI governance is a real political constraint to be designed around, not a fundamental barrier. The treaty architectures that have succeeded have done so by preserving the form of sovereign choice while making the cost of the wrong choice sufficiently high. The same approach is available for AI governance. The challenge is building the political conditions in which governments conclude that joining is the right choice — which is a question of public understanding, scientific consensus, and sustained diplomatic effort, not a question of whether binding international governance is compatible with sovereignty.
Common questions.
All international treaties involve sovereignty constraints. A treaty that bound parties to nothing would not be a treaty. The question is whether the governance benefits justify the constraint. Countries accept sovereignty trade-offs constantly through trade agreements, alliances, and international institutions. For existential risk from frontier AI, the case for accepting governance constraints is strong on exactly the same grounds that justified accepting constraints on nuclear and chemical weapons development.
In specific cases, yes. The Comprehensive Nuclear-Test-Ban Treaty has not entered into force because it requires ratification by 44 specific countries, and eight of them, including the United States, China, India, and Pakistan, have not ratified. However, a de facto moratorium on nuclear testing has been largely observed, showing that treaty non-ratification does not necessarily mean the underlying norm has failed. The sovereignty objection has delayed treaties; it has not permanently blocked underlying governance goals.
Yes. Treaty design choices significantly affect how sovereignty-constraining a treaty is in practice. A framework focused on transparency and reporting obligations rather than outright prohibitions is less invasive than a broad ban. A treaty with a clear scope, covering only frontier AI development above specified thresholds, is less invasive than one governing all AI. Graduated participation mechanisms, similar to the Montreal Protocol's differentiated obligations for developed and developing countries, let countries join at a commitment level proportionate to their capacity.
Because the alternative is often assessed as worse. Countries accept binding restrictions on military and dual-use technology when they conclude that a world in which all major actors are constrained is safer than a world in which none are. NPT non-nuclear signatories accepted restrictions on nuclear development because the alternative was a world with 30 or more nuclear-armed states. The calculation for AI governance is analogous: binding international AI safety constraints may be safer for every country, including the leading AI-developing powers, than an unconstrained race.