Race dynamics describe a coordination problem. Each developer may believe the risks justify slowing down, and each also believes that if it slows alone, a rival it trusts less will take the lead. Racing then looks better for every developer whatever the others do, even though everyone racing leaves the group worse off. Game theorists call that structure a prisoner’s dilemma.
Armstrong, Bostrom and Shulman modelled the problem for AI in 2016 and found that more rivals, and more enmity between them, raise the risk of a dangerous outcome. Anthropic gave the same reasoning when it dropped the pause commitment from its Responsible Scaling Policy in February 2026.
Coordination has worked in other fields where the harm could be measured, a substitute existed and non-members paid a price. Frontier AI has so far had voluntary commitments without enforcement.