Comparing the audience effect of anthropomorphic robots and humans in economic games

This is the authors' abstract. We don't add key points for this paper.

Research in human-agent interaction increasingly examines how advancements in AI systems, often designed to mimic human-like traits, impact human behavior. This study examines specifically whether the audience effect , which refers to changes in behavior when people know they are being observed, could be used to measure the influence of AI agents on human decisions related to generosity, fairness and honesty. A between-group study (N=289) was designed, where economic games (Dictator, Ultimatum and Mind Games) were played in front of one of three agents: a Computer (minimum expected audience effect), an anthropomorphic Robot, or a Human (maximum expected audience effect). Our first objective was to determine if these games can be used to observe and study the audience effect, by comparing interactions with Computer and Human agents. Our second aim was to assess whether an anthropomorphic social robot could impact decision-making by comparing interactions with Robot and Computer agents. Results showed no differences in the Ultimatum and Mind Games when comparing the Human and Computer conditions. In contrast, donations to charities in the Dictator Game were more generous in the presence of a Human than of a Computer, suggesting that this game can detect the audience effect of an agent. Importantly, however, no difference emerged between the Robot and Computer conditions, indicating that anthropomorphic design features alone are insufficient to trigger an audience effect. These findings identify the Dictator Game with charities as a method for studying social facilitation effects of an agent, while highlighting the limits of anthropomorphism. • The Dictator Game with charities can be used to observe and study the audience effect of any agent, under the specific conditions used in this paper. • The Ultimatum Game and the Mind Game have not been found to be suitable methodologies for observing an agent’s audience effect. • Humans tend to be more generous toward charities when making decisions in front of another person rather than in front of a computer or a robot. • Despite its human-like appearance, the Furhat robot–and likely similar robots–does not produce an audience effect on the social dimensions studied in this paper (generosity, fairness, and honesty). Making an agent anthropomorphic does not guarantee a human-like impact on social decision-making. • Participants did not subjectively report being affected by the human experimenter, although their donations in the Dictator Game were affected. Thus, behavioral measures, such as economic games, are important for studying social influence.