Can the US and China agree to limit AI?

Could the two leading AI powers agree on limits to frontier AI development, verify them, and keep them?

Why it matters

Most pause proposals fail if the other leading power continues. “If we slow down, China wins” is one of the most common arguments against slowing down in the United States. See Can AI progress be paused.

The case for yes

• Both governments prize control and stability. A system that its builders cannot control threatens any regime.
• Rivals have limited dangerous weapons during intense competition, as the United States and the Soviet Union did with nuclear arms control.
• Frontier AI depends on a concentrated chip supply chain, which makes some limits checkable.
• Unilateral restraint has sometimes come first: the United States renounced biological weapons in 1969, before the 1972 Biological Weapons Convention.

The case for no

• Each side may see AI as decisive for military and economic power, making a tie unattractive to whoever expects to lead.
• Mutual mistrust is deep, and algorithmic progress is hard to verify.
• An agreement can be used to buy time to catch up.
• Domestic politics on both sides reward appearing strong.

The bargaining question

Is a lead useful leverage, because the trailing side gains from accepting a tie? Or is stopping first a credible signal that the danger is real? The answer affects whether a unilateral slowdown helps or hurts the chance of an agreement.

In his 2026 essay Against SURA, Scott Alexander argues that a unilateral US pause would throw away the main reason China might accept a tie: that it is currently losing. Commenters pointed to unilateral moves that preceded agreements, on biological weapons and on ozone-depleting chemicals. Zvi Mowshowitz adds that Chinese labs have relied on distilling American models, training their own on the outputs of American ones, so a US slowdown would slow them too. Vaniver, a LessWrong commenter, argues that China is, if anything, less eager than the US for massively disruptive technology.

What would change minds

Official statements from either government treating loss of control as a security risk. Progress on verification methods. Whether the trailing side’s progress depends on copying or stealing the leader’s work, in which case slowing the leader also slows the follower.

Related: Compute governance, Institutional resilience.