Conditional pause

A commitment by an AI developer, or a country, to stop or slow frontier AI development if its rivals do the same.

Why it matters before anyone stops

The usual reason given for not pausing is the race: whoever stops alone hands the lead to someone less careful. That is the trap described in Moloch. A conditional commitment answers the race argument on its own terms. If every leading developer says it would stop when the others do, then nobody is choosing to race; they are stuck because there is no way yet to stop together and check that everyone has. The problem becomes one of coordination and verification, which is easier to solve than a lack of will.

Public commitments do work even while they remain conditional:

• They make a shared preference visible. Once each leader has said it would prefer to stop alongside the others, everyone knows that everyone else prefers it too. Agreements are much easier to reach from that starting point.
• They give the public something to hold leaders to. A leader who has said “if they stop, I stop” has a harder time arguing later that stopping is impossible.
• They are ready when a treaty becomes possible. If governments build a way to verify a pause, commitments already on record turn into the first signatures.

Examples

• Lab leaders. At Davos in January 2026, Demis Hassabis said he would support a pause if every company and country joined, while doubting that it could be coordinated. In June 2026 Anthropic said it would slow down or pause if other frontier labs did so verifiably, and that a pause by one lab alone would shift the lead without making anyone safer.
• Campaigners. In March 2026 marchers went from Anthropic to OpenAI to xAI in San Francisco with one demand: that each chief executive commit in public to pausing if every other major lab credibly did the same.

Objections

• Cheap talk. A promise that applies only once everyone else has moved costs nothing to make, and can let a lab look responsible while it keeps racing. Anthropic’s offer depends on a verification system that does not yet exist, so in practice it defers the pause.
• The others include China. A commitment among Western labs does little if the strongest rival is a state that has made no such promise; see Can the US and China agree to limit AI.
• Companies may not be allowed to agree. In the US, competitors agreeing to hold back their products can break antitrust law. Lawyers have asked whether labs could lawfully coordinate a pause without a government behind it.

What would turn conditional commitments into a pause is a way to check compliance, which is why Compute governance matters. Whether a pause can work at all is Can AI progress be paused.