From Rule-Based to Self-Improving Agents: The Evolution of AI Price Collusion | RISE Research

This paper traces the evolution of algorithmic pricing systems, from rule-based algorithms through reinforcement learning to large language model (LLM)-based agents and prospective self-improving agents, examining how each generation changes both the risk of tacit collusion between firms and the difficulty of detecting it. The paper argues that as pricing systems have become more advanced and autonomous (illustrated with real-world examples like Amazon's competitor-monitoring algorithmic pricing and Uber's dynamic surge pricing), they have also become better at coordinating prices with competitors without any explicit human instruction or communication, creating a growing legal gray area for antitrust enforcement, which has traditionally relied on proving agreement between human decision-makers. The paper explains that self-improving agents widen this enforcement gap further, since they can revise their own strategies without being told to, potentially reaching coordinated pricing outcomes that firms never explicitly designed and that regulators cannot trace back to any single human decision. A key contribution is tracking how the evidence used to prove collusion has shifted correspondingly: from communication records (emails, meeting minutes) in the rule-based era, to behavioral pricing patterns in the reinforcement-learning era, to the internal decision logic of algorithms themselves in the LLM-agent era. The paper also surveys the unresolved question of legal liability when agents collude, whether responsibility falls on the firm, the software developer, or a third-party vendor, and reviews arguments on both sides. It concludes that addressing this problem will require regulators to move beyond searching for communication records or unusual pricing patterns toward developing technical tools, including auditability requirements and explainable AI, capable of examining why an algorithm made a given pricing decision, so that firms are incentivized to avoid tacit collusion even when it emerges from an agent's own optimization process rather than explicit instruction.