Introduction
Recent advances in the constellation of artificial intelligence (AI) technologies, including the revolution in neural network models over the past two decades, show great promise in multiple domains of modern life. At the same time, in clinical fields AI may disrupt interpersonal and patient-provider relationships, raising concerns about the ethical, psychological, and societal influences of these applications. Of particular note is the growing capability and broad accessibility of large language models (LLMs). Advances in “generalist” LLMs have produced computing systems with the ability to engage in human-like dialogue with end users. Because of the rapidity with which these technologies are being put into use, as well as the absence of a clearly defined regulatory framework guiding their development and use, little of the research and scholarly analysis that would ideally have preceded their widespread adoption has been conducted.
We also face important as-yet-unanswered questions about the long-term effects that the use of anthropomorphic LLM applications may have on the well-being of individuals and their interpersonal relationships. Yet the reality is that the use of these technologies has spread very rapidly. A recent survey finds that over one-half of Americans have already interacted with an AI LLM. Research on societal implications has not kept up.
Within this context, the current project focuses on the use of AI technologies in the delivery of mental health care.1 We first lay out the kinds of questions that must be answered through research and analysis to ensure a clear understanding of the current and future opportunities and limitations of AI, for use both alone and in conjunction with care providers, in addressing mental illness prevention, diagnosis, and therapy. We then present examples of the array of perspectives people hold on how best to answer those questions and guide future use of the technologies. This project considers both AI designed explicitly for the purpose of addressing mental health care (purpose-built) and the use of rapidly growing generic LLM technologies widely available as “chatbots,” or custom-tailored variants constructed to play particular roles and services.
Background
Digital mental health interventions (DMHIs) are not new. By the 1990s, computer-based cognitive behavioral therapy (CBT), psychoeducational materials, and structured self-help programs were already in widespread use. These early tools gained traction due to expanding Internet access, growing demand for services, and persistent barriers to traditional care.2 Extensive research on the efficacy of these computer-based mental health interventions, in multiple contexts and modalities, dates back to 1968.
Since 2019, the integration of large language models and machine learning systems into digital mental health care has allowed for the rapid development and deployment of AI-enabled tools, which are a major advancement over earlier digital interventions. Some are general-purpose models adapted by users (like ChatGPT and Replika); others, such as Woebot, Wysa, and Youper, are purpose-built for mental health. These tools are now used in clinical settings, educational institutions, workplace wellness programs, and direct-to-consumer platforms.
Despite this rapid spread, uptake is uneven, and basic usage data remain limited. For example, a 2023 survey by the Pew Research Center found that only 4 percent of U.S. adults reported using purpose-built AI-enabled mental health tools, with higher usage among younger adults and those with higher income and education levels.3 Some platforms have published user numbers, such as Replika’s claim of over 25 million accounts, but these figures are unaudited and offer little insight into duration, intensity, or purpose of use.4 Adoption appears to correlate with digital literacy, comfort with technology, and cultural attitudes toward mental health, highlighting much lower levels of utilization for older adults, low-income users, and those from underrepresented groups.5
The evidence base concerning the effectiveness of these interventions is still emerging. A growing number of randomized controlled trials (RCTs) and meta-analyses suggest that AI-driven tools may help reduce symptoms of anxiety and depression.6 But these findings are tentative, as trial design often lags behind commercial development. Some studies have been conducted by the developers themselves, which can introduce the possibility of selective reporting or conflicts of interest.7 Long-term effects, particularly for high-risk or severely ill populations, are largely unknown. Most research focuses on short-term symptom reduction in mild to moderate cases and relies on self-reported outcomes. Many studies do not include appropriate control groups. While some promising results of RCTs are emerging, without this being the standard, the generalizability and durability of results remain unclear.
Privacy, consent, and accountability remain serious concerns. Mental health data are uniquely sensitive, and even unintentional leaks, such as those involving pixel-tracking or behavioral metadata, can lead to harm.8 AI models trained on historical data risk reinforcing already existing systemic biases. Commercial incentives often discourage transparency, and few tools provide users with meaningful explanations of how decisions are made or how their data are used. Currently, no unified regulatory framework governs these systems. Most are entirely unregulated; others operate under consumer technology rules, not health-specific standards.
Also unknown is how the existence of these tools affects overall access to and use of clinical care. AI can extend access in settings with provider shortages or long wait times. These situations are increasingly prevalent, with nearly half of individuals with mental illness receiving no treatment and those who do having an average wait time of forty-eight days.9 But overreliance on automated tools may also displace human connection. The therapeutic alliance, defined here as the relationship between clinician and patient, is a core mechanism of many effective treatments. The consequences of replacing or supplementing that alliance with automated systems are poorly understood. More broadly, these technologies may shift how societies define emotional health and how individuals interpret their own experiences of suffering, resilience, and care.
Our Approach
This document outlines a scholarly agenda to help guide the interdisciplinary inquiry that we believe is lacking in current approaches to addressing this topic. It does not offer conclusions nor attempt consensus. Rather, it identifies key empirical, practical, and ethical questions, distinguishes many of the knowns and unknowns, and seeks to stimulate useful inquiry in a field made noisy from hype, black boxes, and cultural stigma. Our goal is to create a foundation for collaborative work by researchers, clinicians, technologists, and policymakers that will ultimately benefit both those suffering from mental illness and the providers who work to help them. If this document implies a consensus, it is simply this: More research efforts are urgently needed.
Throughout this agenda, contributors draw comparisons across various axes: between AI interventions and traditional talk therapy, between AI use and no intervention at all, and between standalone tools and those embedded in human-led clinical practice. Contributors also consider the differential impacts of AI across distinct populations, including children, individuals with severe mental illness, and users with mild or moderate symptoms. The intended outcomes of the tools considered by our authors likewise vary, from crisis mitigation and triage to long-term therapeutic engagement and ongoing symptom monitoring. We have sought to be explicit in clarifying these distinctions, as each presents unique challenges for research, regulation, and design. We have not attempted to capture the full complexity but rather to pose critical questions and illustrate the diversity of responses. Fully answering the questions will require granular, population-specific inquiry as well as broader systemic analysis.
One of the strengths of this project has been the opportunity to engage contributors from across disciplines, including psychiatry, computer science, ethics, sociology, and public policy. This range of perspectives has allowed for sharper articulation of key questions, identification of blind spots, and recognition of potential unintended consequences. Such integrative work is especially valuable in a domain that is evolving rapidly and unevenly.
While this report primarily examines AI-driven mental health care in clinical settings, we acknowledge these tools operate within a far broader ecosystem. Millions now use conversational LLMs for informal quasi-therapeutic experiences, treating bots as confidants, friends, or romantic partners. Focusing solely on clinical applications risks ignoring complex interactions and unintended consequences. Similar oversights occurred in domains like gaming, where isolating technology from its surrounding culture obscured significant harms. We recognize these interactions raise substantial ethical and societal questions that, although beyond the scope here, must inform both research and policy.
The stakes are immediate. These tools are already shaping real-world decisions by patients seeking care, clinicians allocating attention, and systems determining coverage or reimbursement. The decisions our society makes now will influence AI’s potential role in reducing disparities and improving care, as well as its potential to exacerbate societal anomie or replicate structural inequities. The task ahead is neither abstract nor optional. It requires shared frameworks, clear evidence, and sustained interdisciplinary engagement. We hope that the following questions and diverse responses are the first steps toward laying out a roadmap for future work.