What an AI Safety Treaty
Negotiation Would Look Like.

In 1968 the Nuclear Non-Proliferation Treaty went from draft language to a signed instrument after years of working groups, deadlocks, and domestic fights. People who say "we need an AI treaty" rarely picture that machinery. They should. The gap between the slogan and a ratified text is where most plans die.

International treaty negotiations run as extended, technical, multi-year processes, conducted by diplomatic missions, legal experts, and technical advisers.

How international treaty negotiations work

Most modern multilateral treaties begin with political authorization: typically a UN General Assembly resolution or a decision by a multilateral body mandating the start of negotiations. This creates the formal basis for a negotiating conference or a standing committee in an existing forum such as the Conference on Disarmament in Geneva. Without this political authorization, discussions about a treaty are preparatory and informal. With it, they become a formal negotiating process with a mandate to produce text.

The negotiating body meets in sessions, typically several weeks per year, with intersessional work conducted by technical working groups between formal sessions. Working groups address specific issues in parallel: definitions, verification mechanisms, institutional arrangements, implementation obligations, financial provisions. Each working group produces draft text that is consolidated into a rolling draft of the full treaty over successive sessions.

The draft treaty is negotiated by consensus, meaning no provision is finalized until all participants agree to it. This is a higher bar than majority voting, but it produces treaties with more durable legitimacy and broader participation, because no country has been outvoted on provisions it finds unacceptable. The cost is that consensus requires more time, and the pace is set by the delegation with the most objections to overcome.

Once the text is finalized, it is opened for signature. Signature is a political act of endorsement, not a legal commitment. Legal commitment comes when a country ratifies the treaty through its domestic constitutional process. Most treaties specify a minimum number of ratifications before they enter into legal force.

What working groups a frontier AI treaty would need

A treaty designed to govern frontier AI development would require working groups organized around questions that have no settled answers in international law. The major ones:

Definitions. The treaty needs to specify which AI systems it governs. This is harder than defining nuclear weapons or chemical agents, because AI is a continuous capability spectrum without a sharp distinction between dangerous and safe versions. Proposals center on compute thresholds as a proxy: a training run above a specified computational scale would trigger treaty obligations. The threshold would need technical justification, regular revision as compute costs fall, and a dispute resolution mechanism for cases near the boundary.

Obligations of covered-system developers. Once a system crosses the defined threshold, what are the required actions? Most proposals include prior notification to a monitoring body before or during the training run; mandatory evaluation of the trained system against specified safety standards; reporting of evaluation results; and restrictions on deployment of systems that fail the standards. Each of these requires definitions of what constitutes adequate compliance, which is itself a technical and legal dispute before a single negotiating session begins.

Verification mechanisms. How would an independent body confirm that declared activities match actual activities? This would be the most technically demanding working group. The proposals in circulation combine hardware-level monitoring through cooperation with chip manufacturers tracking sales above specified specifications, facility inspection through rights to visit large compute clusters, and audit rights over training logs and evaluation records. Each element raises questions of commercial confidentiality and national security classification that are not easily resolved.

Institutional structure. Which body would administer the treaty? A new organization, analogous to the OPCW, established by the treaty itself carries the highest governance value but requires the most political effort and resources to establish. Delegation to an existing body carries lower political cost but may not have the mandate or technical expertise the task requires. The IAEA has experience with dual-use technology monitoring; the ITU has some jurisdictional claims over AI-related issues. Each option has different resource implications and different histories of political independence from major powers.

Entry into force requirements. The treaty would need to specify which countries must ratify before it enters into legal force. A treaty that entered into force without the US, China, or the UK would not bind the organizations responsible for frontier AI development. The NPT required ratification by a defined group of states. An AI treaty would need a similar mechanism, with the obvious difficulty that requiring US and Chinese ratification gives both countries a veto over the treaty's existence.

The negotiating process in stages

The predictable sticking points

Experience with comparable negotiations suggests several disputes that would predictably arise and that need to be designed around from the start.

The verification-confidentiality tension. Countries and companies with frontier AI programs will resist verification mechanisms that require disclosure of commercially or militarily sensitive information. AI capabilities represent competitive advantages. The inspection rights that make a verification system credible will be resisted by the same parties whose compliance makes the treaty meaningful. The nuclear regime resolved this through carefully scoped safeguards agreements that verified the non-diversion of nuclear material without requiring disclosure of weapons design information. ASI governance will require an analogous scoping exercise.

The threshold definition dispute. Whatever compute threshold is used to define covered systems will be contested. Parties just below the threshold will argue it should be lower; parties just above will argue it is set too low. And any fixed threshold will become obsolete as compute costs fall, requiring amendment mechanisms that are themselves politically contested. One approach used in other treaties is to delegate threshold-setting to a technical body rather than embedding fixed numbers in the treaty text itself, allowing revision without treaty amendment.

The China participation problem. A treaty that enters into force without China does not govern the organizations responsible for the most significant AI development outside the US. Getting China to accept binding obligations on its frontier AI development requires offering something credible in return (likely a combination of technology access guarantees and symmetrical constraints on US development) under conditions of deep mutual suspicion. This is a diplomacy problem that treaty design cannot fully solve.

The private actor problem. Frontier AI is developed primarily by private companies. The treaty would bind states, which would then implement obligations through domestic law applied to companies. This introduces regulatory complexity that nuclear and chemical weapons treaties did not face, because those treaties addressed primarily state programs. A frontier AI treaty would need to specify clearly how state obligations translate into company obligations, and how violations by a company are attributed to the state party that hosts it.

"Every major arms control treaty looked impossible to negotiate until it was negotiated. The question is always whether the political conditions have ripened enough to make the cost of continued inaction exceed the cost of the agreement."

A frontier AI safety treaty is the output of a sustained political process that has not yet formally begun, not a single event. The work that is currently happening, AI safety summits, bilateral government discussions, national safety institutes, academic proposals for governance frameworks, is the preparatory phase. Converting that into a formal negotiating mandate, and then into a treaty text, and then into ratification, is the work of years. The urgency is that the capabilities being governed may develop faster than the governance process can run. That urgency is the argument for beginning in earnest now.

The strongest pushback

The fairest objection is that great powers will never bind their frontier labs, so describing the negotiation is academic. Every major arms treaty looked impossible until the cost of inaction rose. The preparatory work exists. The missing piece is political will forced by public demand and industrial levers, not a new vocabulary.