AI race isn't a prisoner's dilemma, argues game theory essay

A new viral essay, "The AI Race is Not a Prisoner's Dilemma," takes direct aim at the dominant arms-race framing of AGI development. The author, a philosophy professor, opens by citing polling data: 63% of Americans told Pew that AI is advancing too fast, and 53% believe AI will likely destroy humanity. Yet despite this public unease, the common discourse insists we cannot pause development because of competitive pressure. The essay names Leopold Aschenbrenner's 2024 'breakout' argument—where being months behind on AGI means permanent disadvantage—and Hunter Ash's claim that unlike nuclear weapons, building AGI and using it are the same act, making coordination impossible.

The author then uses game theory to dismantle this view. Rather than a prisoner's dilemma where defection is always rational, the essay models the AI race as an iterated game where trust and repeated interaction between companies like OpenAI, Anthropic, and Google change the payoff structure. It argues that the 'unilateral breakout' scenario is unrealistic because superintelligent AI cannot be secretly developed and deployed without detection, and because the economic costs of racing unsafely—accidents, regulatory backlash, public distrust—outweigh short-term gains. The essay leverages AI assistants (Fable 5 and Opus 5) to iterate through model variations, verify mathematics, and fact-check sources, while acknowledging the visualizations were 'vibecoded.' The central conclusion: the arms-race narrative is a self-fulfilling prophecy, and cooperative safety agreements may be far more feasible than Aschenbrenner and Ash claim.