Human-AI Collaboration to Support Mental Health and Well-Being

Abstract

As mental health conditions surge worldwide, healthcare systems are struggling to provide accessible and high-quality mental health care for all. Although therapy can support people struggling with mental health challenges, barriers like clinician shortages and mental health stigma commonly limit people's access to therapy. In this thesis, I study how human-AI collaboration can improve access to and quality of mental health support. First, I study how human-AI collaboration can empower people who provide support to conduct effective and high-quality conversations. Specifically, I focus on peer supporters on online peer support platforms like Reddit and TalkLife. While peer supporters are motivated and well-intentioned to help support seekers, they are typically untrained and unaware of key psychotherapy skills, such as empathy, that foster effective support. Using a reinforcement learning-based method, evaluated through a randomized trial with 300 peer supporters from the largest peer support platform, I demonstrate that AI-based feedback helps peer supporters express empathy more effectively in their conversations. Second, I investigate how human-AI collaboration can empower people who seek support by making self-guided mental health interventions more accessible and easier to engage with. Self-guided interventions, such as "do-it-yourself" tools to learn and practice coping skills, are often cognitively demanding and emotionally triggering, creating accessibility barriers that limit their wide-scale implementation and adoption. Using cognitive restructuring of negative thoughts as a case study, evaluated through a randomized trial on a large mental health website with 15,531 participants, I show that human-AI collaboration supports people in overcoming negative thoughts and informs psychology theory about processes that lead to positive outcomes. Third, I systematically evaluate human-AI collaboration systems used for mental health support. While there is great interest in utilizing AI for mental health support, there is a significant lack of methods to evaluate their effectiveness, quality, equity, and safety. I study how clinical trials can be conducted to effectively evaluate short-term and long-term outcomes, equity, and safety of AI-based mental health interventions comparing them to traditional approaches. Moreover, I develop a computational framework to automatically assess the behavior of large language models (LLM) when employed as therapists. By analyzing 13 different psychotherapy techniques, I compare the behavior of LLM therapists against that of high- and low-quality human therapy. My analysis reveals that LLMs often resemble behaviors more commonly exhibited in low-quality therapy rather than high-quality therapy, such as offering a higher degree of problem-solving advice when clients share emotions, which is against typical recommendations. My thesis develops two human-AI collaboration systems to support mental health and well-being, along with an evaluation framework for such systems. My work opens opportunities to improve the learning and practice of mental health strategies and coping skills for both support seekers and support providers through human-AI collaboration interventions.