May, 2026
Shilan Huang* & Jin Wan*
Both authors contributed equally to this work.
LLMs tend to generate affirming feedback, sometimes excessively or uncritically (Du et al.,2025). Such AI-generated compliments have been discussed in relation to issues such as AI manipulation and delusion (Bashkirova & Krpan,2024). While existing works are insightful, they are largely experimental, making it hard to evaluate users’ vulnerability to potential influence of AI compliments in settings with high ecological validity. Therefore, our project draws on recorded daily conversations and semi-structured interviews to explore how people perceive and respond to AI-generated compliments across different usage scenarios and conversational chatbots. Importantly, we investigate how users perceive such compliments’ outreach in their daily lives.
Methodologically, this project addresses the challenge of researching human-AI conversations in contexts with limited data portability, such as China. We aim to develop a pipeline for extracting, reconstructing, and analyzing conversational data from real-world user-recorded interactions. The project contributes empirical insights and a reusable toolkit for conversational data analysis.