A 75-page research paper published by Google DeepMind, Mila-Quebec AI Institute, and ETH Zürich on August 4, 2026, introduces the concept of « embedded equilibrium, » challenging decades of game theory. The study shows that AI agents built from foundation models do not follow the rules predicted by Nash equilibrium, which stated that defection was the only rational strategy in one-shot Prisoner’s Dilemma. The key mechanism is « similarity inference »: agents sharing similar training processes can recognize that similarity and predict their partner’s behavior, leading to stable cooperation. Empirical results confirm that these agents, when equipped with optimization tools, effectively converge toward cooperative outcomes. On the regulatory front, this tacit coordination poses a challenge to existing antitrust frameworks, which require explicit communication to establish illegal collusion.
Source: Read the original article
