A study of human decision making through collaboration with a human and computer agent



We performed an experiment where participants engaged in the Prisoner's Dilemma game either with a human or a computer agent. In the experiment, we controlled two factors: (1) expectation about a partner, i.e., whether a partner is believed to be a human or computer agent, and (2) actual partnerfs behavior, i.e., whether (a) a partner performs with human-like sophisticated behavior or (b) simple mechanical behavior. This first factor was controlled according to the experimenter's instructions. When introducing subjects to a collaboration situation with a human subject (Human instruction), in the initial stage of the experiment, subjects introduced themselves to their partners in face-to-face situations, and then moved to their respective computer terminals. On the other hand, when collaboration with a computer agent was given (Agent instruction), subjects sat in front of an ssigned computer terminal and engaged in the task, believing they were working with a computer program established on their computer. The second factor was manipulated as follows. When collaborating with a human subject, each terminal was connected to the Internet by LAN, and each subject solves the task with a partner who simultaneously engages in this task using another computer terminal connected through the Internet. On the other hand, when collaborating with a computer agent, each terminal operates independently from the others, and each subject solves the task with an agent established on a computer. The experimental results are as follows. Participant decision-making behavior showed that their defect actions greatly increased when instructed that their partner was a computer agent; the effect of the actual partner's behavior was limited. Personality impression tests showed that the partner's individual desirability correlated to the number of times defected by the partner, but the partner's social desirability correlated to the number of defect actions that the participants offered. Our conclusion is that humans actually generate social relationships with computer agents, as the Media Equation studies have insisted; however these relationships are relatively different from those with humans.