Yes-man agent makes user make risky choices

Abstract

In this paper, I focused on subjective probability in decision-making in HAI. Many virtual agents are used to support users' decision-making; however, few prior works have focused on ambiguous situations. I hypothesize that users are more likely to choose ambiguity after interacting with friendly agents than with unfriendly ones. To verify this hypothesis, I conducted an experiment using the Ellsberg problem. In the task, participants chose either a box that definitely contained 30 winning balls or a box that contained an undefined number of winning balls. The independent variable was agents. I used two agents: a friendly yes-man agent and unfriendly agent. As a result, the hypothesis was supported. The participants who interacted with the friendly yes-man agent tended to make risky choices. This demonstrates the peculiarity and risks associated with friendly yes-man agents.