Key points are not available for this paper at this time.
Natural language-based AI services like ChatGPT are gaining attention in writing. However, while utilizing LLMs, many users do not receive desired answers in one attempt and resort to multiple reprompts. Despite the frequent occurrence of these trial-and-error processes, research on this issue remains limited. Therefore, this study analyzes user dissatisfaction and response strategies during continuous conversations, focusing on reprompting in writing tasks using ChatGPT. Fifteen participants completed a writing task in three rounds using ChatGPT. Analysis of collected log data and in-depth interviews revealed most users were unsatisfied after a single prompting, leading to reprompting. Additionally, during continuous conversations, users experienced two dissatisfaction patterns: repeatedly facing the same dissatisfaction or encountering unexpected new dissatisfaction after resolving previous issues. In these cases, users tended to try various strategies without clear direction. Based on these findings, we discuss users' perceptions and suggest implications for future interface design.
Jung et al. (Mon,) studied this question.