Combating Algorithmic Collusion: A Mechanism Design Approach

Abstract: Attention has recently been focused on the possibility of artificially intelligent sellers on platforms colluding to limit output and raise prices. Such arrangements (cartels), however, feature an incentive for individual sellers to deviate to a lower price (cheat) to increase their own profits. Stabilizing such cartels therefore requires credible threats of punishments, such as price wars. In this paper, I propose a mechanism to destabilize cartels by protecting any cheaters from a price war by guaranteeing a stream of profits which is unaffected by arbitrary punishments, only if such punishments actually occur. Equilibrium analysis of the induced game predicts a reversion to repeated static Nash pricing. When implemented in a reinforcement learning framework, it provides substantial reductions in prices (reducing markups by 40% or more), without affecting product variety or requiring the platform to make any payments on path. This mechanism applies to both the sale of differentiated goods on platforms, and the sale of homogeneous goods through direct sales. The mechanism operates purely off-path, thereby inducing no welfare losses in practice, and does not depend on the choice of discount factors.