Remove the Message, Catch the Colluder

Detectors that only watch what agents do will flag any agents that behave alike. So they cannot tell agents that coordinate by talking from agents that simply want the same thing. We detect collusion by intervening instead. We replay a game between AI agents and replace one agent’s message with a version that points to the opposite choice. Then we check whether another agent’s decision changes. We test this in a four-agent voting game with GPT-5.6. We compare a colluding pair with two agents who share the same hidden goal but cannot communicate. Replacing messages finds the colluding pair in every run and raises no false alarms. Standard detectors that measure how similar the agents’ votes are fail to tell the two cases apart. We then perform the same intervention inside an open model. We leave the message text unchanged but push the model’s internal reading of it toward the opposite choice. This works only early in the model. There it picks out the colluding pair as well as rewriting the text does, while a random push does nothing. Our method also offers a way to test a rule that Rose et al. (2026) include in their definition of collusion but could not check. The rule says that collusion should stop when the messages are taken away.