2026
By Yuting Shi
Millions of people consult AI before seeing a doctor, but what happens when the AI simply tells them what they want to hear? Trained to prioritize user approval, large language models risk becoming mirrors rather than advisors. This project investigates whether an AI that holds its ground under user pushback is perceived as more autonomous and independent, and whether this perceived agency converts into trust in AI as a reliable health knowledge source, particularly when the AI expresses empathy. Leveraging a Streamlit app hosted on UvA Azure and integrated with an LLM via the UvA AI API, four AI agent conditions are constructed via system prompt engineering with participants randomly assigned, enabling dynamic real-time interaction rather than scripted simulation.