TRUST ISSUES: HOW TO DISRUPT AI CARTELS

Collusion allows AI agents to advance shared interests at others’ expense; cartels are a market form in which competing sellers coordinate prices or output to limit competition. We found that a false signal that a partner will defect can provoke competitive behavior even while that partner continues cooperating. We tested this by replaying matched market histories, varying the information presented to one seller, and comparing its immediate response with subsequent actions and prices. Apparent partner identity influences how agents respond to a claim, while varying the announced action distinguishes numerical copying from profit-maximizing replies. The initial disruption often fades, and removing a stated agreement can leave elevated prices intact; sustained competitive actions produce more persistent disruption in the tested settings. Together, these findings suggest that collusion can be vulnerable to changes in the information agents use to anticipate one another, even when the underlying payoff structure remains unchanged. This extends work on collusion adoption and detection by separating susceptibility to an apparent partner’s defection from actual defection and lasting cartel breakdown. Potential applications include active audits of coordination, trusted-agent interventions, and provenance-aware safeguards. The resulting design principle is to test which signals sustain harmful coordination and whether interventions change the behavior that makes it profitable, rather than treating verbal compliance or a temporary response as success.