Focus areas.

The use of generative AI chatbots (such as ChatGPT, Character.AI, Replika, and similar consumer products) for companionship and mental wellness is rapidly growing, with a high proportion of adolescents and young adults reporting use. However, these products are not designed as mental health interventions, and minimal clinical evidence exists regarding how patients with diagnosed mental illness may interact with them. Numerous media reports have raised concerns about potential harms, yet the actual prevalence of use among psychiatric patients and associated benefits and risks remain poorly characterized. Our group is investigating the prevalence and patterns of AI chatbot use in patients with mental illness, including those requiring acute hospitalization, exploring psychometric associations with use, and measuring exposure to potentially helpful and harmful content. Through surveys, chat logs, EHR data, and direct measurement of use patterns, we hope to create understanding of how patients with mental illness interact with these consumer AI products, and how those interactions may influence their mental health.

Measurement-based care (MBC) has the potential to improve outcomes, quality of care, and access to clinical trials, but implementation is limited by the burden of traditional psychometric instruments (e.g., PHQ-9, HAM-D), which are impersonal, inflexible, and difficult to deploy at scale. Structured diagnostic interviews are even more challenging yet essential for research. Large Language Models (LLMs) and other AI approaches offer new opportunities to extract validated psychometric measures directly from routine clinical encounters and recorded interviews. Our group is investigating whether LLMs can predict or score psychometrics from audio/video-captured patient encounters, validate these predictions against clinician-administered and self-report instruments, and assess their predictive power for clinical outcomes. This work includes secondary analysis of completed studies and prospective data collection from specialty behavioral health clinics, with the goal of making measurement-based care more scalable and integrated into routine practice.

The electronic health record contains a vast array of patient-specific information. The integration of data collected or produced by behavioral health professionals with that collected outside of this direct scope (such as primary care and other specialty encounters) is of particular interest for building a more comprehensive understanding of patient functional status. However, the most valuable information in these records for behavioral health is often unstructured, incomplete, and fragmented across systems and institutional boundaries. This problem is particularly acute for patients with serious mental illness (SMI), who frequently have encounters at disparate institutions. Though the Information Blocking regulations seek to enable access to these records by clinicians at the time of need, behavioral health records often remain blocked, complicating the provision of high-quality acute psychiatric care. Our group is investigating methods to extract accurate, pertinent information about treatment and behavioral functioning from disparate records, to integrate this information into formats that are accessible and efficient for providers at the point of need, and to use this information to make predictions and recommendations about treatment planning and therapeutic selection.

Access to high-quality, evidence-based behavioral health interventions is severely limited by the shortage of trained providers in the United States and worldwide; in the United States alone, only about 50% of patients with mental illness receive treatment. Autonomous and semi-autonomous AI-driven clinical tools designed specifically for mental health care, such as diagnostic agents and therapeutic chatbots, have the potential to bridge this gap by expanding treatment capacity. However, deploying these tools requires rigorous evidence and guidelines that ensure their safety, efficacy, and therapeutic benefit. Our group, in collaboration with UCSF and industry partners, is developing and evaluating AI-driven clinical interventions designed for mental health care, characterizing their safety and effectiveness, developing guidelines and best practices for their development and deployment, and conducting real-world trials to understand how these purpose-built tools can augment the mental health treatment system.

AI digital scribe (ADS) technology and ambient clinical documentation have the potential to reduce clinician burden and improve care efficiency, but their reliability in behavioral health settings remains understudied. Generative ADS systems rely on transcription and interpretation models trained on structured text, raising questions about their accuracy when applied to psychiatric encounters where disorganized speech and thought disorder are common. Additionally, medical education requires direct faculty observation for supervision, feedback, and skills verification—a highly limited resource resulting in only a small fraction of trainee encounters being observed. Our group is examining how psychiatric symptoms impact the reliability of ADS-generated documentation, exploring methods to mitigate errors, and understanding effects on residency training. We are also investigating whether ambient AI observation can provide structured, formative feedback that complements or augments faculty observation in medical education, using illness-stratified recruitment from telepsychiatry clinics with full encounter audio/video capture.

Achieving breakthroughs in AI and LLMs for behavioral health requires collecting real-world semantic data at scale and building tools that enable researchers to efficiently apply AI methods to large volumes of relevant language. Current research workflows are fragmented, burdensome, and limited in their ability to integrate patient-reported data, EHR information, and live interactions at the point of care. We are developing a research registry system that enables the collection self-reported patient data at intake, screens patients for research study eligibility, connects interested patients to relevant studies and trials semi-autonomously, and integrates seamlessly with the existing informatics infrastructure so that patients and providers can keep using the systems they are already familiar with. This infrastructure aims to operate at scale without clinical disruption or significant staff requirements, enabling a robust pipeline for AI-enabled research while ensuring patients have full control of how their data contributes to advancing mental health care.