On the Fragility of Mediation: Theory and Experimental Evidence

Abstract Mediation of disputes is increasingly common, often implemented by computer-run algorithms. We test the efficacy of a theoretically optimal mediation algorithm in an experiment where two subjects, uncertain about each other’s strength, negotiate how to share a resource. The subjects send cheap talk messages to one another (under direct communication) or to the computer mediator (under mediated communication) before expressing demands or receiving the mediator’s non-binding recommendation. While messages to the mediator are more sincere, we find that peaceful resolution is not more frequent. The theoretical analysis shows that mediation is fragile precisely when it is most promising. When the optimal equilibrium improves over direct communication, any deviation from full truthfulness, no matter how small, causes a discontinuous downward jump in the probability of agreement.