Game theoretic models of bargaining and of voluntary contributions to public goods make strong predictions that rely on the hyper-rationality and self-interest of players. Senders in Dictator games should never offer anything to other players. Players should never pass in a Centipede game or voluntarily contribute to public goods funded with a Voluntary Contribution Mechanism. In experimental laboratories, however, violations of these predictions are empirical regularities. Subjects offer money to other players in Ultimatum and Dictator games, voluntarily contribute to the funding of public goods, and achieve non-Nash levels of cooperation in repeated games. Leading explanations for these "anomalies" include decision errors and "other-regarding" preferences (altruism, fairness, warm-glow, reciprocity, etc.). Discriminating between these possible explanations is difficult because experimentalists cannot typically control for the utility that an individual receives when taking an action that improves another's welfare. This paper introduces the use of "virtual agents" for directly detecting other-regarding behavior in laboratory experiments. Human subjects play games with non-human subjects who follow specific strategies or replicate the play of humans in a similar game. Since virtual agents are not real people and do not receive payoffs, the technique removes the concerns of human subjects for other players. We use virtual agents to test for other-regarding behavior in two public goods games: the Voluntary Contributions Mechanism and the Provision Point Mechanism. Comparing contributions in all-human and virtual subject treatments, we find support for the hypothesis that other-regarding behavior elevates contributions. The results also suggest that subjects are motivated by fairness considerations and we conclude that fair-share contributions are not made simply because they are cognitively simple to compute.

Co-author Daniel Rondeau (University of Victoria) and Gregory Poe (Cornell University)