International AI governance under incomplete information: A game-theoretic analysis

This paper develops a game-theoretic model of international AI governance, analyzing how prior trust, signal precision, and payoff structure jointly determine whether intergovernmental cooperation can be sustained. Existing game-theoretic literature focuses largely on firm-level dynamics and treats incomplete information as secondary. We model a dynamic Bayesian stag-hunt game: two governments choose between safety and competitive speed in AGI development; one may be industry-captured with positive probability, while the other observes imperfect signals of its counterpart's regulatory choice. The model yields three types of perfect Bayesian equilibria: Regime Erosion (self-confirming, exists for all parameters); Tit for Tat (requires high signal precision); and Strategic Tolerance (welfare-superior, requires high prior trust but noisy signals). A central, counterintuitive result is non-monotonicity: precise signals can destabilize cooperation in high-trust environments yet remain essential to escape mutual defection when trust is low. These findings support a governance framework organized around three policy levers: (1) behavioral track records to strengthen prior trust, (2) calibrated transparency measures adaptive to the strategic context, and (3) structural instruments that reduce the benefit-cost ratio of reckless AGI development by internalizing systemic risk.