Do language-model pricing agents learn to collude?
Price competition is rising steeply in the operations research journals and AI governance leads the AI journals, while collusion remains a small topic that is rarely tied to algorithms. When sellers let language-model agents set prices, do prices drift up without any agreement?
What prompted this#
On the trends page, "price competition" rises more steeply than any other growing keyword in the operations research journals (11 papers in July–September 2026 against about 2 expected), and "AI governance" tops the growing keywords of the AI journals. Since October 2025, 46 papers carry a collusion keyword (28 in economics); only 7 also carry an algorithm, reinforcement-learning or AI keyword, and "algorithmic collusion" appears on 5. Q-learning, the algorithm behind the best-known simulations of algorithmic collusion, shows up on the equation ranking almost only in AI journals.
Questions#
- In repeated pricing games, do language-model agents reach prices above the competitive level, as Q-learning agents do in simulations?
- Does it depend on the prompt, on memory, or on seeing rivals' prices?
- What tests could a competition authority run on such agents before they are deployed?
Possible approach#
Simulate repeated Bertrand competition with differentiated products, with language-model agents facing each other and facing Q-learning agents, and compare their prices with the Nash equilibrium and the joint-profit benchmark.
Next step#
Run a two-firm simulation with off-the-shelf models and record average prices over many rounds.