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.
Note (2026-11-11): checked against the full text of the paper (Keppo, Li, Tsoukalas and Yuan, arXiv 2603.20281). The figures the essay gives (price lift of 22 percent falling to 10 percent with unequal patience and to 7 percent with unequal data; Q-learning agents against LLM agents; model-size differences stabilizing collusion) are as in the paper. Four points differ. (1) The essay ends by saying that research on AI collusion tests only the setting where the risk is greatest and so overstates it. The paper's own evidence is also from a stylized model and laboratory experiments, with one open-source model family (a distilled 32-billion-parameter DeepSeek-R1) and, in its main comparisons, two agents over 1,000 periods; the paper says field data are not yet available. (2) The essay says the agents collude 'only when deployed identically'. In the paper, collusion is reduced, not removed: lifts of 10 percent and 7 percent remain, and a pair of 32B and 14B agents still shows a 21 percent lift. (3) The essay says more competing algorithms disrupt coordination. The paper's table shows the effect is not monotonic: three patient LLM agents give a 36 percent lift, above the two-agent 22 percent, and four and five give 11 and 3 percent. (4) The comparison to federated learning security work is the essay's own; the paper does not mention it.
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.