Artificial Intelligence and Algorithmic Price Collusion in Two-sided Markets

Cristian Chica, Yinglong Guo, Gilad LermanarXiv 2024 · DOI · arXiv

Abstract

This paper investigates the emergence of tacit collusion among AI pricing agents using Q-learning in two-sided markets characterized by network externalities. Through simulation-based modeling of a repeated game between competing platforms, the study demonstrates that network effects significantly amplify collusive behavior, often exceeding profit margins seen in one-sided markets. The research also introduces a modified Q-learning algorithm with a regulatory penalty term to mitigate these anti-competitive tendencies.

Introduction

The rise of algorithmic pricing has sparked concerns about "tacit collusion"—a phenomenon where AI agents learn to maintain high prices without explicit communication. While previous research has focused on simple one-sided markets, this paper addresses the complexities of two-sided markets, such as ride-sharing or gaming platforms. These environments are defined by network externalities, where the value for one user group (e.g., riders) depends on the participation of another (e.g., drivers). This study explores how these dynamics influence AI-driven collusion and evaluates potential regulatory interventions.

Methodology

The authors employ a simulation-based approach to model a repeated game between two competing platforms. The interaction follows a specific timeline where platforms set prices and users respond based on utility and network effects.

The Model

The game consists of two stages:

  1. Pricing: Platforms determine prices for both sides of the market simultaneously.
  2. User Choice: Users select a platform or an "outside option" based on price and the utility derived from network externalities.

AI and Network Effects

Platforms utilize Q-learning, a reinforcement learning algorithm with bounded memory, where strategies are based on the previous round's prices. A network externality matrix ($\Phi$) is incorporated to account for within-side effects (e.g., more gamers improving a network) and cross-side effects (e.g., more drivers benefiting riders).

Regulatory Mitigation

To address anti-competitive behavior, the researchers introduce a penalty term into the Q-learning update rule. This term penalizes agents when their prices significantly exceed the market average, simulating the impact of regulatory oversight.

Key Insights and Ideas

The study reveals several critical aspects of algorithmic collusion in complex markets:

  • Complexity as a Catalyst: The larger action space in two-sided markets (setting two prices) surprisingly facilitates collusion. The authors suggest this allows platforms to "communicate" more nuanced information through pricing signals.
  • Network Effects as "Collusion Fuel": Positive network externalities significantly boost collusion. High within-side effects, common in online gaming, allow AI agents to exploit these dynamics for maximum joint profit.
  • The J-Shaped Relationship: Cross-side externalities exhibit a J-shaped relationship with collusion. Collusion levels are lowest when these effects are near zero but spike as they become strongly positive.
  • AI Patience: Unlike human-led firms that typically require high discount factors to collude, AI agents maintain tacit collusion even at low discount rates, posing a greater threat to competition.
  • Dampening Factors: High user heterogeneity and the presence of attractive outside options serve as natural barriers to collusion.

Key Results

The simulations yielded significant findings regarding profit margins and the efficacy of interventions:

  • Baseline Collusion: Even without network externalities, AI platforms achieved profit gains of approximately 30% over competitive levels, surpassing the ~20% typically seen in one-sided markets.
  • Effectiveness of Penalties: The proposed penalty term proved highly effective. A penalty coefficient ($\rho$) between 0.2 and 0.3 successfully forced AI agents back to baseline competitive profit levels.
  • Real-World Alignment: The results align with empirical data suggesting that algorithmic pricing can increase margins by roughly 15% in retail markets.

Limitations and Future Work

While insightful, the study faces certain limitations. The "curse of dimensionality" restricted simulations to two platforms; future research should explore computationally efficient ways to model larger oligopolies. Additionally, the current model assumes "single-homing" (users join only one platform), whereas real-world users often "multi-home" (e.g., using both Uber and Lyft). Finally, modeling network effects as non-linear transformations could reveal even more complex collusive strategies.

Summary updated

Abstract

Algorithmic price collusion facilitated by artificial intelligence (AI) algorithms raises significant concerns. We examine how AI agents using Q-learning engage in tacit collusion in two-sided markets. Our experiments reveal that AI-driven platforms achieve higher collusion levels compared to Bertrand competition. Increased network externalities significantly enhance collusion, suggesting AI algorithms exploit them to maximize profits. Higher user heterogeneity or greater utility from outside options generally reduce collusion, while higher discount rates increase it. Tacit collusion remains feasible even at low discount rates. To mitigate collusive behavior and inform potential regulatory measures, we propose incorporating a penalty term in the Q-learning algorithm.
BibTeX
@misc{chica2024artificial,
  author = {Cristian Chica and Yinglong Guo and Gilad Lerman},
  title = {{Artificial Intelligence and Algorithmic Price Collusion in Two-sided Markets}},
  year = {2024},
  eprint = {2407.04088},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.04088},
}
Figure