ABSTRACT

Contrary to the traditional concept of artificial intelligence, explainable artificial intelligence (XAI) aims to provide explanations for the prediction results and make users perceive the system as being reliable. However, despite its importance, only a few studies have investigated how the explanations of an XAI system should be designed. This study investigates how people attribute the perceived ability of XAI systems based on perceived attributional qualities and how the power status of the XAI and anthropomorphism affect the attribution process. In a laboratory experiment, participants (N = 500) read a scenarios of using an XAI system with either lower or higher power status and reported their perceptions of the system. Results indicated that an XAI system with a higher power status caused users to perceive the outputs of the XAI system to be more controllable by intention, and higher perceived stability and uncontrollability resulted in greater confidence in the system’s ability. The effect of perceived controllability on perceived ability was moderated by the extent to which participants anthropomorphised the system. Several design implications for XAI systems are suggested based on our findings.

摘要

与传统的人工智能概念相反，可解释人工智能（XAI）旨在为预测结果提供解释，让用户认为系统是可靠的。然而，尽管它很重要，但只有少数研究调查了如何设计 XAI 系统的解释。本研究调查了人们如何根据感知归因质量来归因 XAI 系统的感知能力，以及 XAI 的权力状态和拟人化如何影响归因过程。在实验室实验中，参与者（N = 500) 阅读了使用具有较低或较高功率状态的 XAI 系统的场景，并报告了他们对系统的看法。结果表明，具有较高功率状态的XAI系统使用户感知XAI系统的输出更容易被意图控制，更高的感知稳定性和不可控性导致对系统能力的更大信心。感知可控性对感知能力的影响受到参与者拟人化系统的程度的调节。根据我们的发现，提出了 XAI 系统的几个设计含义。