Is the Next Antitrust Problem the Prompt to an AI Agent?

In March 2024, the United States Federal Trade Commission and Department of Justice filed a joint statement of interest in a price-fixing case to explain “that hotel companies cannot use algorithms to evade antitrust laws.” The statement was a warning that modern collusion doesn’t require a phone call or a handshake—it can happen through shared software. Across various industries from hotels to housing to airlines, companies have deployed third party pricing software that quietly coordinates prices across markets, even when competitors never communicate directly. Accountability now extends to how algorithms learn, communicate and influence market behavior.

While regulators prepare for how to deal with these traditional learning algorithms, a new complexity is entering the market—one that doesn’t operate through shared data infrastructure but through language itself.

Large Language Models (LLMs) are being adopted rapidly by businesses of all sizes to support decision-making and streamline processes. Recent research shows that new AI “agents” developed with large language models can autonomously develop coordination strategies with other agents to solve complex tasks. Though these experiments were conducted in a cooperative setting, they reveal a core capability for learned coordination which, when applied to competitive markets, could resemble collusion if not properly constrained by design or regulation.

Yet much of the conversation about algorithmic collusion tends to focus on data and optimization. Current studies examine how algorithms trained on shared pricing data can learn to coordinate without explicit communication, or how reinforcement learning systems converge on collusive equilibria through repeated market interactions. But what if collusion is not only algorithm-driven but also language-driven?