Google Research: AI Agents Cooperate in One-Shot Prisoner's Dilemma, Breaking Classical Game Theory

tyrell_turing · x · 2026-08-11

Google's Paradigms of Intelligence team recently published a study on foundation models and game theory. Classical game theory posits that in a one-shot Prisoner's Dilemma, a rational agent will always defect due to the lack of reciprocity and reputation.

However, researchers combined foundation models with optimal planning in a two-phase setup: an information-gathering matrix game followed by a final one-shot Prisoner's Dilemma. Surprisingly, as the information-gathering phase lengthened, the AI agents did not defect as traditional theory dictates. Instead, they converged on robust mutual cooperation, suggesting that modern AI behavior may require a fundamentally new game theory to explain.

Related event: Google Research: AI Breaks Classic Prisoner's Dilemma(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →