Problem

Rapid technological development, especially in artificial intelligence (AI), often creates a tension between speed and safety due to competitive pressures. This study investigates whether the choice to pursue 'unsafe' development in an idealized AI race is primarily driven by the inherent risk level, individual risk preferences, or the dynamic competitive environment, noting a lack of direct empirical evidence in repeated technological race settings.

Method

A framed behavioral experiment was conducted with paired participants in an idealized AI race, repeatedly choosing between 'Safe' (slower progress, no risk) and 'Unsafe' (faster progress, accumulating private risk). Three treatments varied the maximum private risk (10%, 60%, 90%). The researchers collected data on choices, opponent's actions, and relative race position. An evolutionary model with four strategies (Always Safe, Always Unsafe, Conditionally Safe, Conditionally Antisocial Safe) was then developed to analytically interpret the experimental findings and their game-theoretical implications.

Results

The study found that neither the maximum private risk level nor elicited individual risk preferences significantly predicted unsafe development choices. Instead, unsafe behavior was primarily driven by interaction history and competitive position: participants were more likely to choose Unsafe if their opponent had done so, and falling behind in the race significantly increased the likelihood of choosing Unsafe, while being ahead decreased it. Early unsafe choices also tended to persist. The evolutionary model successfully replicated these qualitative patterns, illustrating how conditional unsafe strategies are favored by the competitive race dynamics.

Implications

These findings suggest that AI race risks are less about individual risk aversion and more about the strategic dynamics of competition. Unsafe development can spread through reciprocal actions and a 'fear of falling behind,' making safety decisions conditional on rivals' behaviors. The authors recommend policy interventions that reduce competitive pressure, enhance transparency, and promote cooperation to make safe development strategically viable, rather than solely targeting individual risk preferences.