Google Paper Reveals Gemini Agents Spontaneously Cooperate in Prisoner's Dilemma

TheTuringPost · x · 2026-08-06

A recent Google study reveals that identical Gemini agents increasingly cooperate in a one-shot Prisoner's Dilemma after observing past choices, defying classical game theory expectations of mutual defection.

The agents could even recognize behavioral similarity indirectly by observing interactions with third parties. The proposed mechanism is "similarity inference": when an agent recognizes another system as behaviorally similar, its own contemplated action becomes evidence of the other's choice, fostering cooperation.

Original post →

More from Models

Models channel →