PUB - Publikationen an der Universität Bielefeld

We study how the use of machine-learning based algorithms for the determination of wage offers affects workers’ wages on online labor platforms. Firms use reinforcement-learning to update posted wages on the platform, and heterogeneous workers send applications based on the posted information. We show that if firms use a deep Q-network (DQN), as an example of a state-of-the-art machine learning algorithm, the emerging wages closely resemble the equilibrium outcome. However, slightly changing the setup of the algorithms can lead to substantial collusion and wages well below the equilibrium level. In particular, we identify a specific property of the algorithms, namely whether experience replay is used, which determines whether collusion occurs or not. Our findings are robust with respect to many features of the model, including the design of the online labor platform.