Signal Study: Falling behind in AI race drives unsafe development more than risk preferences

Summary

A behavioral-economics paper reports results from a framed experiment modeling an idealized AI race, in which paired participants repeatedly chose between "safe" and "unsafe" development paths under an uncertain time horizon. Unsafe development produced faster progress and higher immediate payoffs but accumulated private risk up to a treatment-specific maximum of 10%, 60%, or 90%, while the competitive structure itself was held constant. The study's pre-registered hypotheses, comparing outcomes across risk levels and testing the role of individual risk preferences, were not supported by the data. Instead, exploratory analysis found that unsafe choices were driven by the evolving strategic state of the race: participants were more likely to go unsafe right after their opponent did, being ahead reduced unsafe play while falling behind increased it, and first-round choices predicted later behavior. The authors introduce a four-strategy evolutionary model (Always Safe, Always Unsafe, Conditionally Safe, Conditionally Antisocial Safe) that reproduces the observed treatment effect. They conclude that policy should focus on reducing competitive pressure and promoting cooperation in AI development, rather than relying solely on individual risk regulation.