Can AI progress be paused?

Could the world stop or substantially slow frontier AI development long enough to matter, and keep it stopped?

Why it matters

If alignment is very hard, a pause is the most direct response: do not build what cannot yet be made safe. If a pause is infeasible, effort shifts toward making the race safer. Disagreement about this question is often mistaken for disagreement about How hard is alignment.

What a pause could mean

• An international agreement with verification, the strongest form.
• A coordinated pause among companies, possibly conditional on everyone else stopping.
• A unilateral pause by one country or company, possibly temporary and meant to prompt others.
• Partial slowdowns, such as limits on training compute, rather than a full stop.

These differ in feasibility and in what happens if others continue. Arguments about “a pause” often move between them without saying so. Geoffrey Irving describes a strong version of a unilateral pause: temporary, defined by the lab itself, redirecting staff to safety, security, and product work, and ending if others do not follow.

A formal treaty is not the only route. Oliver Habryka argues that most real slowdowns would come from messier pressures, such as states threatening escalation against rivals’ projects, social censure, and conflict between companies. The main alternative to a pause is sometimes called burning the lead: the race continues, and the most careful developer spends its few months of advantage on safety before less careful competitors catch up. See Plans for the AI transition.

The case that it is feasible

• Frontier training depends on scarce physical inputs: a few chip designers and fabrication plants, and large datacenters that are hard to hide. See Compute governance.
• Dangerous technologies have been restricted before, including biological weapons, ozone-depleting chemicals, and nuclear testing.
• Governments have strong reasons not to let rivals or companies build systems no one can control.
• Public concern about AI can grow quickly after visible incidents.
• Those with the most influence over AI have rarely put it behind a pause. Their claims that it cannot be stopped are mostly untested; see Musk has not seriously tried to stop AI.

The case that it is not

• The economic and military incentives are enormous and growing.
• Algorithmic progress is harder to verify than compute, and algorithms keep getting more efficient, so a fixed limit erodes over time.
• Lasting enforcement may require intrusive monitoring of computing.
• A party that pauses while others continue may lose influence over the outcome. See Can the US and China agree to limit AI.
• Even the most powerful people in the industry say they cannot stop it. “There’s nothing I can do to stop it,” Elon Musk said in 2026. If he cannot, it is hard to see who could.

Is the time worth it?

A pause buys time; the value depends on use. Time could go to alignment research, verification tools, institutions, or augmenting human intelligence. Some argue that time later, with better AI tools and more public attention, is worth more than time now. Others argue any delay before an irreversible step is valuable. In his 2026 essay Against SURA, Scott Alexander argues that a year of an earlier situation buys less than a year of a later one, with more attention and better tools. Zvi Mowshowitz replies that this cannot mean decades of delay would have bought nothing.

What would change minds

Progress on verifying compute use. Whether major governments treat AI as a security threat as well as a prize. Whether efficiency gains let frontier systems be trained without large, visible infrastructure.

Related: Should safety-focused labs keep building, Pivotal act.