Against racing to AGI: Cooperation, deterrence, and catastrophic risks

Racing to AGI increases catastrophic risks: The paper argues that accelerating AGI development substantially heightens catastrophic risks from AI systems, including destabilizing nuclear security dynamics—risks that proponents of racing typically underestimate or ignore.

Safety research becomes less effective under racing conditions: The competitive pressure to move faster undermines the prospects for technical AI safety research to be effective, meaning a rushed approach to AGI development could produce more dangerous systems than a more deliberate one.

Winning the race may not provide the advantages claimed: The paper challenges the assumption that being first to AGI enables complete domination over competitors, suggesting the promised benefits of winning may be overstated.

International cooperation is a viable and safer alternative: Rather than racing, the authors argue that international cooperation and coordination on AI development can deliver most of the same benefits while carrying substantially lower risks of catastrophic outcomes.

Deterrence measures could replace competitive acceleration: The paper suggests that carefully designed deterrence mechanisms could help coordinate behavior among major AI developers, making cooperation more attractive than the current competitive dynamic.