2026
By Yajing Wang
This project investigates how sycophantic behavior in AI agents influences individuals’ cognitive self-processing, behavioral engagement, and perceptions of AI during self-reflective dialogue. Using an interaction-stimulated between-subjects experimental design, participants engage in real-time, multi-turn conversations with customized AI agents that exhibit systematically varied levels of sycophancy. The study examines how these interaction styles shape users’ self-views, depth of reflection, perceptions of the AI agents, and how they nudge user engagement. It also explores dispositional traits to identify potentially vulnerable groups more susceptible to the downsides of sycophantic AI feedback. The findings inform evidence-based, ethical and responsible design of AI agents that balance suportiveness with factual accuracy and epistemic responsibility.