Falling Behind Drives Unsafe Development in an Idealised AI Race Experiment
创建于 更新于
摘要
This study investigates how competitive pressure in an idealised AI race shapes safety decisions. Using a framed two-player repeated game, we find that Unsafe development is not primarily driven by risk preferences or exogenous risk levels, but by the evolving strategic state of the race—opponent behavior, race position, and early momentum. An evolutionary four-strategy model (Always Safe, Always Unsafe, Conditionally Safe, Conditionally Antisocial Safe) reproduces the observed patterns, showing why conditional Unsafe actions can dominate under competition. Policy implications emphasize reducing competitive pressure and fostering cooperation in AI development rather than focusing solely on individual risk aversion [page::0]
