What an AI Treaty Would Actually Do.
A treaty would not stop progress. It would create the floor below which no actor would be allowed to compete, so that progress could be lived with.
The absence of a rule is a rule. It is the rule that whoever moves first writes the rules later.
Treaties are not utopia
An international treaty on frontier artificial intelligence will not be elegant. It will be partial, late, contested, routinely circumvented at the margins, and never popular with the actors it constrains. This is true of every treaty that has ever mattered.
It will still be the difference between a future in which civilization has at least attempted to set the terms of an irreversible technology, and a future in which those terms are set, by default, by whichever firm or government was willing to take the most risk first.
What it would actually contain
A working treaty would do a small number of concrete things, none of them mysterious.
It would define a compute threshold above which any new training run must be registered with an international body before it begins. It would require that any system trained above that threshold be subject to independent pre deployment evaluation by reviewers without commercial ties to the developer. It would establish disclosure obligations for critical safety incidents, with a fixed window measured in days, not quarters. It would prohibit specific categorical uses, beginning with fully autonomous lethal targeting of human beings. It would create a verification regime modeled on the inspection rights that exist, today, for nuclear materials and chemical precursors. It would attach meaningful penalties to non compliance, including restrictions on cross border deployment for firms based in non signatory states.
None of this is technically novel. The legal craft exists. The diplomatic machinery exists. The international bodies that would host it exist. What is missing is political will, which is to say, public demand.
The point of a floor is not to be a roof. The point of a floor is that no one is permitted to dig beneath it.
What it would not do
It would not stop research. It would not freeze development. It would not prevent open source models from existing, although it would shape the conditions under which the largest of them could be deployed at scale. It would not, by itself, redistribute the gains of AI to the workers displaced by it, although it would create the legal scaffolding on which such redistribution becomes possible.
It would also not solve the alignment problem, the labor problem, the synthetic media problem, or the AGI threshold problem on its own. A treaty is a floor. It is not a roof. People sometimes object to the inadequacy of a treaty by listing the problems it does not solve. This is a category error. The point of a floor is not to be a roof.
The objection of inevitability
The most common objection to a treaty is that adversaries will simply ignore it. This argument has been made about every weapons treaty in the modern era, and it is, predictably, made by the actors who would prefer no constraint to exist at all.
Treaties do not require universal compliance to function. They require enough compliance to shift the strategic equilibrium. The Treaty on the Non Proliferation of Nuclear Weapons has not been universally honored. It has, nonetheless, kept the number of nuclear weapons states an order of magnitude smaller than the projections of the 1960s. The Chemical Weapons Convention has not been perfectly enforced. It has, nonetheless, made the use of chemical weapons a near universal occasion for international response. Imperfect floors are floors.
The species has built such instruments before, against the same objections, on shorter clocks, with thinner evidence. It is being asked to do so again, in time.