Breaking Algorithmic Collusion in Human-AI Ecosystems

The growing adoption of AI agents is leading to multi-agent ecosystems where these AI agents interact with each other and with humans at test-time. When humans deploy AI agents to make pricing decisions, these AI agents to implicitly compete with each other and with humans who opt to manually perform the task. In this work, we study multi-agent pricing ecosystems from a game-theoretic perspective, focusing on how the market price is affected by human-AI interactions. Specifically, we study a repeated game model where AI agents implement strategies that constitute Nash equilibria in the repeated game, and where humans manually performing the task implement a no-regret strategy. Motivated by how populations of AI agents can exhibit supracompetitive prices, we investigate whether high prices persist even when some humans manually override the algorithm with simpler heuristics. Our main finding is that a single human performing a no-regret override can already drive down the price. Although the price does not necessarily go down the competitive level in general, near-competitive prices are guaranteed when there are large number of AI agents or with multiple humans performing the task manually.