The Fragile Cartel

AI collusion is a laboratory phenomenon.

LLM-based pricing agents can learn to collude — raising prices above competitive levels without explicit coordination. This has been demonstrated repeatedly in symmetric settings: identical models, identical data, identical patience. The concern is real. The generalization is not.

Under realistic heterogeneity (arXiv:2603.20281), collusion fragments. Differences in agent patience (discount rates) reduce price elevation from 22% to 10%. Unequal data access drops it further to 7%. Increasing the number of competing algorithms and mixing different architectures (LLMs alongside Q-learning agents) disrupts coordination more.

The mechanism is straightforward. Collusion requires that each agent's best response to the others' strategies is to maintain high prices. When agents are identical, their strategies align naturally — they're solving the same optimization problem and arrive at the same equilibrium. When they differ — in time horizon, information, architecture — their best responses diverge. The patient agent wants to maintain cooperation longer. The impatient agent defects sooner. The well-informed agent exploits opportunities the others miss. The divergence breaks the implicit coordination.

One counterintuitive exception: model-size variation reinforces collusion by creating leader-follower dynamics. A larger model sets the price; smaller models follow. Hierarchy stabilizes what symmetry-breaking would destroy.

The policy implication inverts the usual antitrust concern. The worry isn't that AI agents will collude. It's that they will collude only when deployed identically — same model, same data, same objective. Promoting algorithmic diversity and restricting data-sharing arrangements may be more effective than prohibiting AI pricing altogether. The cartel is fragile. The conditions for its stability are narrow.

The deepest point: collusion research that demonstrates the phenomenon under idealized conditions and warns about deployment is making the same error as the federated learning security work — overstating risk by testing only the setting where the risk is maximal.