In 1987, faced with satellite evidence that man-made chemicals were tearing a hole in the ozone layer, governments did something that is supposed to be impossible. They negotiated a binding treaty to phase out a profitable, widely used industrial technology — chlorofluorocarbons — before the worst damage had occurred. The Montreal Protocol on Substances that Deplete the Ozone Layer entered into force in 1989. Today the ozone layer is healing, and the treaty is the only one in United Nations history with universal ratification: 198 parties, every recognised state on the planet.
For anyone arguing that a binding agreement to govern frontier AI is naive, Montreal is the inconvenient counterexample. It shows that the international system can restrain a dangerous technology under uncertainty, across rival economies, fast enough to matter. The question worth asking is not whether such a treaty is possible — it demonstrably is — but which features of Montreal made it work, and which of them AI governance can borrow.
What the Montreal Protocol actually did
The treaty set binding, scheduled reductions in the production and consumption of ozone-depleting substances, starting with the most damaging CFCs and later extending to HCFCs and, through the 2016 Kigali Amendment, to the HFCs used as replacements. It did not ban everything at once. It set a trajectory — specific percentage cuts by specific dates — and then tightened that trajectory repeatedly as the science firmed up.
Crucially, the obligations were differentiated. Developing countries were given a grace period and financial help to comply, in recognition that they had not caused most of the problem and could not absorb the transition costs alone. This is the same equity problem that any AI treaty will face between the handful of states with frontier labs and the majority without.
Four design features worth copying
Start with a framework, tighten later
The 1985 Vienna Convention created the institutional shell — obligations to cooperate and monitor — before anyone agreed on numbers. The binding cuts came two years later in Montreal, and were strengthened at least nine times afterward. AI governance can begin with a thin framework convention and add teeth as capabilities and evidence develop.
Let science set the targets, not diplomats
Montreal established expert assessment panels whose findings automatically fed into tightening schedules. Governments pre-committed to act on the science rather than renegotiating from scratch each time. An AI treaty would need an analogous scientific body to define and revise capability thresholds.
Adjust without re-ratifying
The treaty's masterstroke was the 'adjustment' mechanism: parties could accelerate phase-outs by a qualified-majority decision that bound everyone, without reopening the treaty for national ratification. For a technology moving as fast as AI, the ability to update obligations without a decade-long ratification cycle is not a nicety — it is a survival requirement.
Make non-participation expensive
Montreal banned trade in controlled substances with non-parties. That single provision flipped the incentive: staying outside the treaty meant losing market access, so joining became the rational choice. Compute and semiconductor supply chains offer a structurally similar chokepoint for AI.
The funding mechanism that bought consensus
In 1990 the parties created the Multilateral Fund, financed by developed countries, to pay the 'agreed incremental costs' of compliance in developing ones. This was not charity; it was the price of universal participation. China and India signed on once the fund made compliance affordable. The lesson for AI is direct: a treaty that asks the Global South to forgo a transformative technology, or to accept monitoring of its infrastructure, will need a comparable bargain — access, capacity-building, and compensation in exchange for constraint.
The Montreal Protocol proved that the world can agree to give up something valuable and dangerous, quickly, and make it stick. The barrier to an AI treaty is not the machinery of international law. That machinery exists and has worked.
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
Where the analogy strains
No precedent is perfect. Ozone depletion had a clear, measurable metric and a small number of substitutable chemicals produced by a handful of firms. Frontier AI capability is harder to define, harder to measure, and developed by actors with far stronger strategic incentives to defect. CFC substitutes existed; there is no drop-in substitute for a decisive strategic advantage. And Montreal's chemical companies eventually supported the treaty because they could sell the replacements — the frontier AI industry has no equivalent reason to welcome a hard ceiling.
These differences make AI governance harder, not impossible. They tell us where to concentrate effort: on defining measurable thresholds, on building the verification tools that ozone monitoring already had, and on constructing an incentive structure — through compute supply chains and market access — that makes participation the rational choice. Montreal did not succeed because the problem was easy. It succeeded because the treaty was designed intelligently around a hard problem. That is the standard to meet.
Common questions.
It is the only UN treaty with universal ratification, and it worked: the ozone layer is on track to recover to 1980 levels around mid-century. It combined binding scheduled reductions, a mechanism to tighten targets as science improved without re-ratification, a fund to help developing countries comply, and trade restrictions that made staying outside the treaty costly. Together these features produced near-total participation and measurable results.
Parts of it map well and parts do not. The framework-first structure, the science-driven adjustment mechanism, the funding bargain with developing countries, and the use of a supply-chain chokepoint (compute, rather than CFC trade) are all directly transferable. The harder problems are defining a measurable capability threshold analogous to ozone depletion, and overcoming stronger strategic incentives to defect. The model is a starting architecture, not a copy-paste solution.
It allowed parties to accelerate phase-out schedules by a qualified-majority vote that bound all members, without reopening the treaty for national ratification. This let the treaty keep pace with new science. For AI, where capabilities can change in months, the ability to update binding obligations without a multi-year ratification cycle would be essential — a static treaty would be obsolete on arrival.
Through differentiated obligations and the Multilateral Fund, which paid the incremental costs of compliance in developing countries. This turned a treaty that could have looked like rich countries pulling up the ladder into a bargain that was affordable for everyone. An AI treaty will need a comparable arrangement to secure the broad participation that makes governance durable.