Domination-Avoiding Agents Prevent AI Collusion in Market Pricing

A new theoretical paper from Noam Nisan and Emmanuel Zerah provides a formal solution to the emerging problem of AI collusion in automated markets. Earlier empirical work by Calvano et al. showed that Q-learning agents spontaneously learn to coordinate on high prices when competing as sellers. More recently, Fish et al. found similar collusive behavior in commercial large language models (LLMs) used for pricing decisions. These findings raised concerns about deploying autonomous AI agents in competitive markets.

The current paper defines a broad class of algorithms called 'domination-avoiding agents.' This class includes mean-based learners, internal-regret-minimizing agents, multiplicative-weight agents with variable learning rates, and contextual variants. The authors prove that any domination-avoiding agent, when interacting in a repeated game, will almost never play strategies that are eliminated by repeated elimination of purely dominated strategies. This property prevents the implicit collusion seen in Q-learning and LLMs, because collusive price-fixing requires playing strategies that are dominated in the long run.

Crucially, the result holds for any game, not just pricing models, making it a general safety guarantee for autonomous learning agents. The paper thus offers a concrete design principle: to prevent undesirable collusion, system designers should use domination-avoiding learning rules rather than standard reinforcement learning or LLM fine-tuning that can lead to tacit collusion. This work opens the door to deploying AI agents in markets with formal assurances that they will not coordinate to raise prices.