Recommendation Design: Can Mediators Implement Correlated Equilibria?

Abstract We report results from a novel experiment studying individuals’ ability to act as mediators by designing incentive-compatible recommendation devices for rational robot agents, that is, to implement correlated equilibria. To test the skills required of a successful mediator, we vary both the games played by the agents and the mediator’s objective. Most participants successfully implement correlated equilibria across games, especially in games where coordination yields benefits. Across objectives, however, many participants initially favour fair and efficient devices, even when these properties are not incentivised, suggesting that fairness and efficiency serve as important heuristics in recommendation design.