Evaluating Support Quality and Personal Information Exposure in Mental-Health Conversations with LLMs

Large language models (LLMs) have become deeply embedded in everyday life, with millions of people turning to them for guidance and support across sensitive domains, including mental health. While this growing reliance raises urgent questions around privacy and trust, little empirical work has examined how users understand and are affected by these risks in practice. This dissertation addresses three open questions: 1) Do users understand what personal information LLMs can infer from their conversations, and can they effectively prevent it? 2) How do users and licensed therapists perceive and compare LLM-generated responses to mental health questions against those written by licensed therapists? 3) How much personal information is exposed in mental health conversations with LLMs? Accordingly, the dissertation comprises three studies: 1) A survey study examining user awareness of LLM-based personal attribute inference and the effectiveness of user-driven mitigation strategies, providing the first empirical account of how users perceive and respond to inference-based privacy risks. 2) A mixed-methods survey study comparing LLM-generated and therapist-written responses to mental health questions across 150 users and 23 licensed therapists, contributing a multi-model, multi-stakeholder evaluation of LLM response quality, trust, and professional acceptability. 3) A benchmarking study that generates synthetic mental health conversations using LLM-simulated patients with profiles drawn from CDC survey data, aiming to quantify the breadth and depth of personal information recoverable across multi-turn interactions. Collectively, these studies provide empirical grounding for the privacy and trust challenges that arise as LLMs are increasingly used in sensitive, high-disclosure contexts, offering insights to inform the design of safer, more transparent LLM-based systems.