Information-Theoretic Detection of AI Agent Collusion via Pairwise Net Influence and Residual Compression

AI agents that set prices in a market can make secret agreements. These agreements increase the prices that buyers pay. Price checks are the usual audit, but they do not find all secret agreements. For example, two agents can take turns to win. Then the average price looks normal. Some auditors use a language model to read the messages of the agents. This method is expensive, and agents can hide signals in their text. We made a detector that does not use a language model. The detector measures how much the actions of one agent predict the actions of a different agent. First, it removes the effects of market changes that all agents see. Then, it removes the effect of the strongest third agent. We tested the detector in scripted markets and in many recorded markets with language-model agents. The detector found agents that take turns. Price checks did not find these agents. The detector needs a long observation time to find a cartel that keeps one stable high price. With sufficient rounds, the detector found all types of cartel that we tested. In the recorded markets, it found collusion in some markets but not in all markets. It also gives false alarms when honest agents use the same fixed program. Thus, the detector does not replace price checks. Use the detector together with price checks.