Economics
Algorithmic Collusion in Digital Pricing Environments
Quick fact
In a 2017 study by the Federal Trade Commission, researchers found that simple Q-learning algorithms could learn to set prices collusively in repeated games, even without any communication—demonstrating that coordination can emerge entirely from algorithm optimization.
Why this is interesting
Imagine you and your rival both set prices using an AI. Without any human collusion—no secret meetings, no phone calls—these AIs might 'learn' to keep prices high, leaving you both profiting while consumers overpay. How could that happen?
Read the full explanation
Understanding Algorithmic Collusion in Digital Pricing Environments
Think of two rival coffee shops on a street. Normally, each sets their price based on their own costs and demand. Now, imagine both shops use a computer algorithm that automatically adjusts prices each day. These algorithms are designed to optimize profit by reacting to the other shop's price. At first, they might undercut each other to gain customers. But over time, a clever algorithm might 'learn' that matching a high price is more profitable than constantly cutting. If both algorithms do this, they can end up charging high prices, effectively colluding—not because they agreed, but because each independently learned that coordination is more profitable than competition. This is algorithmic collusion: the use of pricing algorithms by competing firms to learn and sustain prices above the competitive level.
A deeper explanation
The mechanism behind algorithmic collusion is rooted in the concept of tacit collusion, where firms coordinate on prices without explicit communication. In traditional markets, tacit collusion is possible but difficult to sustain because it relies on mutual observation and understanding. Algorithms, however, can make this coordination far more effective. They can process vast amounts of pricing data, respond instantly to competitors' changes, and employ reinforcement learning (e.g., Q-learning) to test different pricing strategies. Q-learning algorithms maintain a 'Q-value' for each possible action (price) in each state (market condition). They initially take random actions, but through trial and error, they discover that setting high prices leads to higher long-term rewards, especially if they anticipate rivals will also raise prices. Crucially, these algorithms can learn to detect when a rival deviates from the high-price equilibrium and punish them by temporarily lowering prices, similar to a 'trigger strategy' in game theory. Because this learning happens autonomously, there is no human 'meeting of minds,' making it difficult for antitrust authorities to prove illegal collusion under current laws, which typically require evidence of communication or agreement. This emergent outcome challenges fundamental legal definitions of collusion and raises important questions about how to regulate markets where AI systems set prices.