When the NPT opened for signature in 1968, the fear was concrete: President Kennedy had warned that by the 1970s there might be fifteen, twenty, or twenty-five nuclear-armed states. That world did not arrive. Nine states have nuclear weapons today, not twenty-five. The NPT is a large part of why — and understanding its mechanism matters, because frontier AI presents the same structural problem: a small number of actors hold a dangerous capability, and the question is how to stop it spreading while persuading the have-nots that restraint is in their interest.
The three pillars
The NPT rests on a bargain among three commitments, each pillar leaning on the others. The nuclear-weapon states agreed not to transfer weapons and to pursue disarmament in good faith. The non-nuclear states agreed not to acquire weapons and to accept IAEA safeguards on their nuclear programs. And all parties affirmed the right to peaceful nuclear energy, with cooperation and technology-sharing for those who stayed non-nuclear.
Non-proliferation
The have-nots pledge not to build or acquire weapons and to accept international verification. This is the constraint the treaty exists to impose.
Disarmament
Under Article VI, the haves pledge to negotiate toward eliminating their arsenals. This is the reciprocal concession that makes the constraint politically survivable — the have-nots are not simply being told to stay behind forever.
Peaceful use
Everyone retains access to civilian nuclear technology, with the haves helping the have-nots develop it. This is the benefit that makes joining attractive rather than merely restrictive.
Why the bargain structure matters for AI
The deep lesson of the NPT is that a non-proliferation regime cannot be only a set of restrictions on the weak imposed by the strong. It has to offer the constrained majority something they value — here, energy technology and a promise of eventual disarmament — or they will not join, and a treaty the relevant states have not joined governs nothing. The two-way structure is what produced near-universal membership.
AI faces an identical divide. A treaty on frontier AI will be negotiated between a few states with leading labs and a large majority without. If it looks like the AI 'haves' locking in permanent advantage while everyone else accepts monitoring, it will fail to attract the participation that makes it meaningful. The NPT's answer — shared access to the benefits, plus a credible commitment by the leaders to accept reciprocal constraints — is the template for making an AI bargain acceptable to the states being asked to sign.
The NPT's failures are also lessons
The treaty is not an unqualified success, and its weak points are precisely where AI governance must do better. Article VI's disarmament promise has been honoured slowly and grudgingly, breeding lasting resentment among non-nuclear states who feel the bargain was one-sided. Three nuclear-armed states — India, Pakistan, Israel — never joined, and North Korea withdrew, showing that a determined state can stay outside or exit. And the 'peaceful use' pillar spreads exactly the enrichment technology that brings a state to the nuclear threshold, an ambiguity Iran has exploited for decades.
Each failure carries a design warning. An AI treaty's equivalent of the disarmament promise — whatever the leading states offer in return for others' restraint — has to be credible and honoured, or it will corrode. The treaty needs to make exit costly rather than consequence-free. And it has to grapple, as the NPT never fully did, with the dual-use ambiguity at its core: the same capabilities that deliver enormous benefit are the ones that create the danger.
The NPT worked because it was a deal, not a decree. The nuclear powers did not simply order everyone else to abstain; they offered something in return. Any AI treaty that forgets this will be signed by no one who matters.
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
The template, stated plainly
The NPT gives AI governance a concrete blueprint for the political architecture, distinct from the technical one. Identify the capability to be contained. Ask the states that hold it to accept reciprocal, verifiable limits — not just to impose limits on others. Guarantee the broad benefits of the technology to those who accept constraint. Build in verification through an independent agency. And learn from where the NPT frayed: make the leaders' concessions real, make exit expensive, and confront the dual-use problem head-on rather than papering over it. The most-joined arms control treaty in history is a bargain. The AI version will have to be one too.
Common questions.
Non-proliferation (non-nuclear states agree not to acquire weapons and to accept IAEA safeguards), disarmament (nuclear-weapon states agree under Article VI to negotiate toward eliminating their arsenals), and peaceful use (all parties retain the right to civilian nuclear technology, with cooperation for those who stay non-nuclear). The three are a linked bargain: restraint by the many in exchange for concessions and benefits from the few.
Because AI presents the same structural problem: a small number of actors hold a dangerous capability, and governance depends on stopping its spread while persuading the majority without it that restraint serves their interest. The NPT's core insight — that a non-proliferation regime must offer the constrained majority real benefits and reciprocal commitments, not just impose restrictions — is directly transferable to designing an AI treaty that states will actually join.
Largely, yes. It helped turn a predicted world of twenty-plus nuclear states into one with nine, and it is the most widely joined arms control treaty in history. But it has real failures: slow disarmament by the nuclear powers, several states that never joined or withdrew, and a peaceful-use pillar that spreads threshold technology. These successes and failures are both instructive for AI governance.
Replicating its resentments. Non-nuclear states have long felt the disarmament side of the bargain was never honoured, straining the treaty's legitimacy. If an AI treaty looks like the leading states locking in permanent advantage while everyone else accepts monitoring, it will provoke the same backlash and fail to win broad participation. The concessions offered to the AI 'have-nots' must be genuine and delivered, not rhetorical.