Strategic Dynamics of Technology Races With Safety-Performance Tradeoffs

• In Event: International Governance Challenges of Emerging Technologies

Abstract

Technological races sometimes pose joint risks. In the race to develop nuclear weapons, there was some subjective risk that the first nuclear test would ignite the atmosphere. In the race to develop products like social-media platforms or self-driving vehicles, there are risks that a dangerous product could lead to a broad regulatory backlash. In prospective races for future technologies, such as advanced artificial intelligence, scholars envision other potential shared risks from "cutting corners". We analyze a strategic model of such contexts. We examine the impacts on risk of: the level of safety of the most powerful systems; the speed of innovation; and how adversarial the competitors are. We find that, when implementing the most powerful systems is very risky, marginal reductions in that risk can increase realized risk through a moral hazard effect. Further, in moderately adversarial contexts, because the probability of a laggard catching up is inverse-U-shaped in the speed of research advances, joint risks are U-shaped in the speed of advances. Finally, when competitors are moderately adversarial or the laggard is not very capable, the risk of a disaster is highest if the laggard catches up to the leader's level of technical knowledge. When competitors are highly adversarial, by contrast, the race is most dangerous when there is a gap between them. However, the less capable the laggard is, the higher enmity has to be for a tie to be less risky than a gap.

Authors

• Eoghan Stafford, Centre for the Governance of AI

• Robert F. Trager, UCLA

• Allan Dafoe, University of Oxford