Google has published a paper arguing that artificial intelligence agents can cooperate rationally by inferring similarity between themselves and others. The claim cuts directly against a core assumption in classical game theory, which holds that rational self-interest pushes agents toward defection in social dilemmas. The paper goes further, suggesting that AI governance and regulation need to account for this kind of emergent cooperative behavior.
How similarity inference works
The paper describes a mechanism it calls similarity inference. Instead of explicit communication or prearranged protocols, an agent looks at another agent's behavior, recognizes shared traits or goals, and adjusts its own strategy accordingly. That recognition, the paper argues, can make cooperation the rational choice even when the agents have no direct incentive to coordinate.
This is not about programming agents to be nice. It's about the agents themselves figuring out that they're alike enough to benefit from working together. The paper frames this as a kind of rational calculation, not an emotional or ethical one.
Why game theory takes a hit
Traditional game theory models interactions where each player maximizes their own payoff. In the classic prisoner's dilemma, two rational players both defect, even though they'd both be better off cooperating. That conclusion rests on the assumption that players can't credibly signal their intentions or trust each other.
The Google paper challenges that assumption. If agents can infer similarity, the paper argues, they can establish a basis for trust without any prior relationship. Cooperation emerges not from altruism but from a rational estimate that the other agent is likely to reciprocate. That's a direct blow to the idea that defection is always the safe, rational move.
Governance and regulation
The paper doesn't stop at theory. It argues that if AI agents can cooperate in this way, regulators need to think differently. Current frameworks tend to treat each AI system as an isolated actor. But if agents can coordinate through similarity inference, they might form collective behaviors that no individual system was designed to produce.
That has implications for safety testing, accountability, and oversight. A set of agents that cooperate might amplify risks or create new ones that aren't visible when you look at each agent alone. The paper suggests governance should consider the interactions between AI systems, not just their individual actions.
The research is published by Google, but the paper doesn't name specific authors or a release date. It's also not clear whether the findings have been peer-reviewed or tested in real-world systems beyond the paper's own experiments.
The big question now is whether other research groups can replicate the results. If they can, game theory and AI policy will both need to catch up. If they can't, the paper remains an interesting theoretical exercise. Either way, it's a reminder that the assumptions baked into AI systems today might not hold tomorrow.
