Mediated talk: An experiment

We experimentally compare mediated (cheap) talk with direct (cheap) talk. Theory, guided by a characterization of equilibria in both environments, suggests that mediated talk has the potential to improve information sharing and welfare relative to direct talk. We sharpen the theory prediction by invoking Crawford's (2003) language-anchored level-k analysis. In the experiment, we find that mediated talk can indeed facilitate information transmission. We also find, however, that this requires that the language employed conforms with the mediation mechanism: mediation mechanisms improve information sharing for a variety of conforming languages, but fail to do so with a nonconforming language. These experimental findings match the predictions from the language-anchored level-k analysis. Strikingly, this is the case even when a whole array of alternative selection criteria (including iterative deletion of dominated strategies, strict equilibrium, Pareto efficiency, etc.) make a unique common prediction that sharply disagrees with the language-anchored level-k prediction.