Artificial Intelligence, algorithmic pricing, and predatory behavior

This paper investigates whether reinforcement learning algorithms can autonomously develop predatory pricing strategies in digital markets. We model a duopoly where two Q-learning agents repeatedly interact in a differentiated goods market with multino-mial logit demand, incorporating cash reserve dynamics and endogenous bankruptcy conditions. The theoretical framework establishes a Markov perfect equilibrium in which predation dominates collusion when one firm holds sufficiently larger cash re-serves and the discount factor satisfies a recoupment threshold. Simulations across 400 parameterized experiments — split between symmetric and asymmetric cash set-tings — reveal that predatory behavior emerges spontaneously in 43–45% of converged runs, without any explicit coordination or communication between agents. Cash flow asymmetry is the primary mechanism: as divergence in reserves grows during the com-petitive phase, algorithms increasingly converge toward predatory rather than collusive strategies. Intermediate learning rates amplify this tendency, contradicting prior find-ings by Calvano et al. 2020, who found monotonically pro-collusive effects of higher learning rates. Higher discount factors, conversely, favor collusion. These results carry direct policy implications: algorithmic predation is structurally undetectable through traditional conduct-based tools, calling for ex-ante monitoring frameworks, capital-flow audits, and modernized antitrust guidelines capable of addressing spontaneous exclusionary dynamics in AI-driven markets.