Tacit Collusion in Pricing Agents: Robustness, Detection, and a Mechanistic Probe of a Pricing LLM

Tacit collusion, AI agents settling on prices higher than a competitive market would produce without ever communicating, is a recognized risk as pricing increasingly relies on AI agents, and a hard one to detect from the outside. We gave AI agents the job of setting prices and let them run a market on their own, then asked two questions. First, can you break a collusive habit once it forms, without more oversight? Yes: you do not need more competitors, just one agent that flatly refuses to go along with high prices, and that alone drags the market back down, for a simple rule-based learner and, more strongly, a large language model pricing on its own. Second, when a pricing LLM reacts to a rival, is that real reasoning, or just going through the motions? We opened the model up and found a real, physical trace of that reasoning inside it: nudge it, and the model's own price moves in response. But when we tested specifically whether it punishes a rival for undercutting it, we found nothing: it reacted just as strongly when a rival raised prices as when it lowered them. This is not a model plotting against a competitor, it is a model stuck on charging a lot by default. That is the core message of this report: watching AI agents settle on high prices is not the same as proving they are colluding. We leave the full statistics to the project README.