ABSTRACT

Conversational Agents (CAs) are becoming part of our everyday lives. About 10 percent of users display aggressive behavior toward CAs, such as swearing at them when they produce errors. We conducted two online experiments to understand user aggression toward CAs better. In the first experiment, 175 participants used either a humanlike CA or a non-humanlike CA. Both CAs worked without errors, and we observed no increased frustration or user aggression. The second experiment (with 201 participants) was the focus of this study; in it, both CAs produce a series of errors. The results show that frustration with errors drives aggression, and users with higher impulsivity are more likely to become aggressive when frustrated. The results also suggest that there are three pathways by which perceived humanness influences users’ aggression to CAs. First, perceived humanness directly increases the frustration with the CA when it produces errors. Second, perceived humanness increases service satisfaction which in turn reduces frustration. Third, perceived humanness influences the nature of aggression when users become frustrated (i.e., users are less likely to use highly offensive words with a more humanlike CA). Our research contributes to our theoretical understanding of the role of anthropomorphism in the interaction with machines, showing that designing a CA to be more humanlike is a double-edged sword—both increasing and decreasing the frustration that leads to aggression—and also a means to reduce the most severe aggression.

摘要

会话代理 (CA) 正在成为我们日常生活的一部分。大约 10% 的用户对 CA 表现出攻击行为，例如当 CA 出现错误时辱骂他们。我们进行了两项在线实验，以更好地了解用户对 CA 的攻击行为。在第一个实验中，175 名参与者使用类人 CA 或非类人 CA。两个 CA 均正常工作，没有出现任何错误，并且我们没有观察到用户的挫败感或攻击性增加。第二个实验（有 201 名参与者）是本研究的重点；其中，两个 CA 都会产生一系列错误。结果表明，对错误的沮丧会导致攻击性，而冲动性较高的用户在沮丧时更有可能变得攻击性。结果还表明，感知人性通过三种途径影响用户对 CA 的攻击性。第一的，当 CA 产生错误时，感知到的人性会直接增加对 CA 的挫败感。其次，感知到的人性化提高了服务满意度，从而减少了挫败感。第三，人性感知影响当用户感到沮丧时，攻击性的本质（即，用户不太可能使用更具人性化的 CA 来使用高度攻击性的词语）。我们的研究有助于我们从理论上理解拟人化在与机器交互中的作用，表明将 CA 设计得更加人性化是一把双刃剑——既可以增加也可以减少导致攻击性的挫败感——同时也是一种手段减少最严重的攻击行为。