Why an AI Pause Could Lose Its Deterrent Power Even Without Cheating

An agreement to pause frontier AI would need more than promises and inspections. It would need to give each participant enough time—and sufficient incentive—to respond if the other secretly resumed development. In a November 2026 preprint, Texas A&M University economist Li Gan argues that this response time can shrink even when neither country has broken the agreement.

The mechanism is growing computing capacity. In the model, both countries freeze capability advancement while continuing safety research, preserving their capability gap. But installed compute keeps expanding. A secret research program launched later can therefore reach the frozen rival frontier sooner. The visible gap stays unchanged while the opportunity to answer a breakout contracts.

A lead is not a response window

Gan distinguishes a follower’s capability lag from the time it would need to catch a halted leader. The follower advances through its own research and through diffusion of knowledge from the leader. Diffusion weakens as the gap closes, slowing the final stretch of catch-up.

In the illustrative calibration, a follower six months behind needs 12 months to catch a leader that stops. About 81% of its initial progress comes from diffusion. These are scenario inputs and model implications, not measurements of actual national AI programs.

During mutual restraint, both own capability research and capability-enhancing adoption stop. If the follower later breaks that restraint, growing research capacity accelerates its advance. Its modeled catch-up time falls from 12 months at the pause’s start to 8.61 months after a year and 5.74 months after two years.

Detection must be affordable—and worth answering

The paper connects that physical window to a game between two governments. Each prefers mutual restraint to a race, but each would prefer an unanswered secret advance to cooperation. Monitoring can deter that temptation only if detection arrives soon enough and the other country would actually restart development.

That response is credible when recovering strategic position outweighs the additional development cost and catastrophe risk. A threat that would not be carried out cannot support deterrence in this game.

Cooperation also requires a financial overlap: monitoring must be intensive enough to discourage cheating, yet cheap enough that neither country would rather reject the agreement and race. As response windows shorten, the required monitoring increases. With cooperation’s benefits and monitoring costs held fixed, those two bounds eventually stop overlapping.

This establishes that cooperation can be supported, not that governments necessarily choose it. Mutual advancement remains another equilibrium even when a cooperative outcome is available.

What the numerical clock means

In the main worked example, cooperation remains supportable through a review at month 18.9. That result assumes illustrative national payoffs and monitoring costs, no implementation delay, and a separately fixed 12-month window for responding to a leader’s breakout. It is not an estimated lifetime for a real treaty.

Halving monitoring costs keeps cooperation supportable through month 31.8. Crucially, that is where the example’s remaining 36-month pause becomes as short as catch-up time. Beyond that point, scheduled resumption comes before the modeled breakout succeeds, so the simple game no longer applies. The number is not a predicted collapse date.

Verified matched safety research provides another extension. In an illustrative calculation, making only the cooperative outcome safer increases the reward for compliance and reduces the attraction of breakout, moving the last supportable review from month 18.9 to month 25.5. This changes incentives without changing the physical catch-up path.

Deterrence should not be confused with reliable interception. At the baseline’s initial threshold, detection within the window has only about a 32% probability. That suffices to remove the expected advantage of cheating under the chosen payoffs; demanding a much lower probability of successful breakout requires more monitoring.

The assumptions behind the result

The strategic meaning of breakout matters. The main example treats passing the frozen frontier as decisive, although the leader can still reach the model’s final capability milestone first after restarting. If winning meant only reaching that milestone first, the same follower breakout would not be decisive.

Risk assumptions matter too. The main game uses chosen catastrophe probabilities rather than the physical simulation’s outcomes. Using the scenario’s own high risks initially makes restarting unattractive; safety learning later restores credibility, producing both an opening and a closing date for cooperation.

The model solves individual review blocks, not repeated treaty renewal or renegotiation. Compute growth is externally specified, monitoring findings are truthful, and simulated risk reductions depend strongly on whether the first country’s completion ends the other’s hazards.

Gan’s central contribution is therefore a conditional design test: an unchanged capability gap does not guarantee an unchanged ability to enforce restraint. Verification, response speed and the benefits of cooperation must be assessed against the capacity a secret program could mobilize later.