Interface cues in AI chatbots: Effects of response mode, display polarity, and visual avatars on user perceptions and satisfaction.

Abstract

Generative artificial intelligence (GenAI) has become a research focus in recent years, and has been widely applied in many fields. Due to its excellent natural language processing ability and its capability to simulate human conversation, the Chat Generative Pre-trained Transformer (ChatGPT) has attracted much attention from both media and academia. However, studies have mainly focused on the technical implementation of ChatGPT and its social impact, with few paying attention to the influence of its interface design on users. This study aimed to explore the effects of three research variables-response mode, polarity display, and visual avatar-on users' perceptions and experience of chatbot graphical interfaces. The study consisted of two stages. The first stage adopted a 2 (response mode) × 2 (polarity display) × 3 (visual avatar) mixed factorial design, which aimed to evaluate the effects of the relevant research variables on users' perceptions and experience. The second stage applied the Kano model to assess users' satisfaction with and importance ranking of the related research variables. The results indicate that Word-for-Word Responses received higher SUPR-Q and perceived Technological Sophistication ratings, whereas Full-Length Responses received higher perceived Human-Likeness and overall satisfaction ratings. Positive Polarity displays generally performed better than Negative Polarity, but in the human-avatar condition, Negative Polarity can enhance the sense of Technological Sophistication. The interaction patterns differed across outcomes. For user experience quality, Word-for-Word Responses received higher ratings under Positive Polarity, whereas Full-Length Responses received higher ratings under Negative Polarity. In contrast, for perceived Human-Likeness, Word-for-Word Responses received higher ratings under Negative Polarity, whereas Full-Length Responses received higher ratings under Positive Polarity. Polarity Display interacted with Visual Avatar in perceived Technological Sophistication, although the polarity comparisons within each avatar condition were not significant after Bonferroni adjustment. In the Human Avatar condition, Full-Length Responses received higher perceived Human-Likeness ratings under Positive Polarity than under Negative Polarity. The Kano analysis classified Negative Polarity as Attractive, Positive Polarity and Full-Length Responses as Indifferent, Word-for-Word Responses and the Iconic Avatar as Reverse, the Human Avatar as Indifferent, and the Robotic Avatar as One-Dimensional. Overall, among the evaluated interface configurations, the combination of Full-Length Responses, Positive Polarity, and a Robotic Avatar was the most frequently selected by participants. These research results provide theoretical support for the iterative development of AI chatbots, and have significance for the development of user experience in future GenAI application scenarios in daily life.