Adding an Intermediate Price Does Not Boost AI Collusion Risk

Janusz Meylahn’s paper, published on 2026/10/19 in the Journal of Artificial Intelligence Research (JAIR), reports extensive simulation studies of reinforcement learning algorithms in pricing environments ¹. The simulations suggest that when agents include a memory of one period they may learn to collude; prior work had focused mainly on settings with two prices while the simulations ran up to fifteen prices.

Here we address this gap by analyzing the probability of observing collusion using decentralized Q‑learning in a duopoly where each agent can choose from three possible prices and remembers its opponent’s last move. Using a logit demand model for payoffs leads to a payoff structure that is a three-action generalization of the iterated Prisoner’s Dilemma. Our central question is whether adding an intermediate price between “defection” and “cooperation” increases the likelihood of observing cooperation or collusion.

Decentralized Q-learning is provably convergent in this environment, and the basins of attraction of its equilibria can be calculated explicitly using Individual Best-Response Graphs. We find that even though the additional action increases the number of collusive equilibria, the likelihood of collusion does not increase.

This work builds on earlier research demonstrating that reinforcement learning agents can learn collusive pricing strategies—a concern given the increasing use of algorithmic trading in markets. The methodical strength lies in being able to calculate basins of attraction precisely; it allows a detailed analysis of how adding intermediate actions affects collusion likelihoods. While the extra action raises the count of collusive equilibria, it does not raise the chance of cooperation or collusion, suggesting current models may overstate risks or be too simplistic.

The findings carry implications for market regulators and AI designers who worry about hidden collusion among autonomous traders. By showing that simply expanding the set of available prices does not automatically increase collusion risk, the study offers reassurance while also pointing out that further investigation into richer state representations might still reveal vulnerabilities ¹.