Hazing Period Required for AI Cooperation in Prisoner's Dilemma

Researchers Benedict Russell, Chin-wing Leung, and Paolo Turrini have published a theoretical analysis of cooperation in the repeated Prisoner's Dilemma under a 'trigger-restart' mechanism. In this setup, self-interested agents play a sequence of symmetric games and restart their interaction if their actions ever disagree. Using replicator dynamics in a well-mixed population, they model how cooperative strategies can evolve despite individual incentives to exploit. Their key insight: cooperation only stabilizes when agents initially defect for a period—a 'hazing' phase—before switching to indefinite cooperation. The length of this hazing period is critical; longer hazing leads to larger basins of attraction, making those strategies more evolutionarily stable even if they are less optimal in raw payoff terms.

The paper provides exact convergence guarantees for restricted strategy lengths and derives the parametric conditions necessary for stability in the general case. By computing the number of stable strategy sequences, the authors reveal structural properties: agents must 'earn trust' through initial defection. Surprisingly, populations converge more readily to suboptimal sequences with longer hazing periods than to theoretically optimal cooperative sequences. This has implications for multi-agent AI systems where agents must negotiate or collaborate autonomously—suggesting that imposing a brief period of non-cooperation may actually foster long-term trust and stability. The work bridges game theory and multi-agent reinforcement learning, offering a mathematical foundation for designing robust cooperative AI behaviors.