Algorithmic collusion: A critical review

Florian E. Dorner

Introduction

In recent years, antitrust scholars and regulators have grown increasingly concerned about the potential for pricing algorithms to facilitate market collusion. An "alarmist" literature has emerged, suggesting that Reinforcement Learning (RL) agents will naturally learn to coordinate prices, harming consumers without explicit human instruction. However, Florian E. Dorner’s critical review argues that this fear is largely premature. By bridging the gap between legal theory and computer science, the paper demonstrates that while certain models of algorithmic coordination pose a genuine threat, autonomous "self-learning" collusion remains technically improbable in real-world market conditions.

A Taxonomy of Algorithmic Collusion

The review categorizes the risk of algorithmic collusion into three distinct archetypes, each presenting different challenges for regulators:

  1. The Messenger: Algorithms used as simple tools to implement a pre-existing human cartel. This is a traditional antitrust violation where the software merely automates an illegal agreement.
  2. Hub-and-Spoke: Multiple competitors adopting the same third-party pricing software. This effectively synchronizes the market through a centralized service, creating a "de facto" coordinated economy.
  3. Independent Agents: Autonomous AI systems that learn to collude through trial and error without human intervention or shared software. This is the scenario that generates the most academic concern but faces the highest technical hurdles.

The Technical Reality of Multi-Agent Reinforcement Learning (MARL)

The paper identifies several "roadblocks" that prevent independent agents from colluding effectively in the wild. Most studies claiming to prove algorithmic collusion rely on simplistic simulations—typically two-player games in stable, noise-free environments. In reality, markets are far more complex.

  • The Simulation Gap: Results from duopoly simulations rarely scale to markets with more firms, asymmetric costs, or "noisy" data. When the number of players increases, the stability of a potential cartel collapses.
  • The Recursion Problem: Tacit collusion requires agents to form "beliefs about beliefs." Current MARL algorithms struggle with the infinite loop of "I think that you think that I think..." without a shared communication protocol or "common grounding."
  • Non-stationarity: In a multi-agent environment, the environment changes as every agent learns simultaneously. This makes it difficult for an AI to distinguish between a competitor's strategic punishment and random environmental fluctuations.

The Real Threat: Hub-and-Spoke Models

While autonomous collusion is dismissed as "science fiction" for the current generation of AI, the Hub-and-Spoke model is a pressing concern. The concentration of the pricing-algorithm market—where many firms use the same software provider—creates a centralized point of failure. This can facilitate collusion even if the individual firms (the "spokes") are unaware of the coordination. The paper notes that empirical data from the Amazon Marketplace and retail gasoline markets already show signs of this dynamic.

The review highlights the "Guy Named Bob" test, popularized by the FTC’s Maureen Ohlhausen: if it is illegal for a human intermediary named "Bob" to collect prices and coordinate a market, it is equally illegal for an algorithm to do so. The legal challenge lies not in the "Messenger" or "Hub-and-Spoke" scenarios—which fit existing antitrust frameworks—but in the "Independent Agent" scenario, where there is no explicit "agreement" to prosecute under current US (Sherman Act) or EU (TFEU) laws.

Recommendations for Regulators

Instead of focusing on the distant threat of self-learning robots, the author suggests that antitrust authorities should prioritize tangible risks:

  1. Reporting Requirements: Mandate that firms report the use of automated pricing algorithms in sensitive sectors.
  2. Technical Auditing: Develop in-house expertise to "read the code" of simpler algorithms to identify hidden punishment logic designed to stabilize cartels.
  3. Market Concentration: Limit the market share of third-party pricing developers to prevent the emergence of algorithmic monopolies.

Conclusion

The threat of algorithmic collusion is real, but its nature is often misunderstood. While the field of Cooperative AI is moving fast, the technical barriers to independent coordination remain high. By focusing on the technical feasibility of cooperation and the risks of centralized pricing services, regulators can move past alarmist rhetoric and address the actual structural threats to market competition.

Summary updated

Abstract

The prospect of collusive agreements being stabilized via the use of pricing algorithms is widely discussed by antitrust experts and economists. However, the literature is often lacking the perspective of computer scientists, and seems to regularly overestimate the applicability of recent progress in machine learning to the complex coordination problem firms face in forming cartels. Similarly, modelling results supporting the possibility of collusion by learning algorithms often use simple market simulations which allows them to use simple algorithms that do not produce many of the problems machine learning practitioners have to deal with in real-world problems, which could prove to be particularly detrimental to learning collusive agreements. After critically reviewing the literature on algorithmic collusion, and connecting it to results from computer science, we find that while it is likely too early to adapt antitrust law to be able to deal with self-learning algorithms colluding in real markets, other forms of algorithmic collusion, such as hub-and-spoke arrangements facilitated by centralized pricing algorithms might already warrant legislative action.
BibTeX
@misc{dorner2021algorithmic,
  author = {Florian E. Dorner},
  title = {{Algorithmic collusion: A critical review}},
  year = {2021},
  eprint = {2110.04740},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.04740},
}
Figure