September 2024 - November 2024
Qinyi Zhou, Chang Liu, Xinjie Shen, Xingyu Bruce Liu, Sherry Tongshuang Wu, Xiang ‘Anthony’ Chen
Overreliance on AI can degrade the performance of human-AI collaboration. While various methods have been proposed to mitigate this overreliance, their indiscriminate application may impose additional cognitive burdens on users. Therefore, it is essential to develop a method for detecting overreliance. Currently, most researches on this topic are conducted in laboratory settings, which circumvents the challenges of identifying overreliance from user results in real-world scenarios. We propose a method that utilizes behavioral data during interactions with AI to detect the degree of overreliance in real-world use. To achieve this, we collect behavioral data during use and conduct regression on the relationship between the overreliance metric and behavioral features. We consider three representative AI application scenarios: quiz solving, article summarization, and trip planning. The quantitative results indicate that overreliance is correlated with two types of action features: (1) copy and paste, and (2) interaction frequency on task page and Conversational LLM page. Further qualitative
analysis indicates that: (1) overreliance exhibited through copy and paste can be urther understood in terms of user’s questioning and acceptance of AI suggestions, and (2) overreliance caused by frequent interaction with Conversational LLM may stem from users gradually being persuaded by the AI during multi-turn conversations.
As the second author, I am in charge of one (out of three) experiment design, two primary algorithm design, and paper writing (sections about experiment and result evaluation).
Publication:
Submitting to CSCW 2025.