New symposium grapples with AI’s growing role in mental health

By Adam Hadhazy

Traditionally cautious in embracing new technology, the healthcare sector has recently emerged as one of the fastest adopters of artificial intelligence (AI). One specialty lagging behind, though, is mental health. Of the more than 1,400 AI-enabled medical devices approved by the Food and Drug Administration, most just in the past few years, “remarkably, not a single one targets psychiatry,” noted Kilian M. Pohl, PhD, at the inaugural Artificial Intelligence for Mental Health (AI4MH) Symposium. “Today is therefore not simply about discussing what AI can do,” said Pohl. “It is about what AI should do for mental health.” Dr. Pohl founded the AI4MH initiative two years ago and is now one of its co-directors.

Held on June 1 and kicking off the second annual Stanford Health AI Week, the new symposium convened academics, clinicians, entrepreneurs, policymakers, and individuals with lived experience of mental health challenges. Attendees explored how carefully evaluating AI’s promise and pitfalls will be critical in these still-early days for its professional development and deployment.

“AI has so much potential for transformative and profound advances . . . but the potential comes with a stark warning,” said Lloyd B. Minor, MD, the Carl and Elizabeth Naumann Dean of the School of Medicine at Stanford University, in opening remarks. “While those of us in biomedical research and patient care envision a future where AI serves a purpose of good, we're equally aware of the risks and, in some cases, of the devastating consequences that AI can have on mental health.”

In an indication of widespread interest, approximately 325 people attended the half-day conference at the Li Ka Shing Center for Learning and Knowledge in-person, while close to 1000 attended virtually.

“This turnout clearly reflects the great deal of enthusiasm, optimism, and excitement about AI and its promise to help us address longstanding, seemingly intractable issues and problems that we've had in mental health research, care access, and delivery,” said Eric Roland Kuhn, PhD, an associate professor of psychiatry and behavioral sciences, in his opening remarks. “But I also think it reflects some healthy shared skepticism and concerns.”

Laura Roberts, MD, MA, the Katharine Dexter McCormick and Stanley McCormick Memorial Professor and Chair of the Stanford Department of Psychiatry and Behavioral Sciences, emphasized the importance of new approaches AI could facilitate. “Mental health disorders remain among the most urgent and complex challenges in medicine and public health throughout the world,” she said in her opening remarks. “Artificial intelligence is rapidly accelerating nearly every aspect of science and society, creating extraordinary opportunities to rethink how we understand conditions, diagnose and treat them.”

Led by Ehsan Adeli, PhD, who also serves as one of AI4MH’s co-directors, the event was co-organized by the AI4MH initiative—a special program within the Department of Psychiatry and Behavioral Sciences at Stanford—and the Stanford Institute for Human-Centered AI (HAI) with sponsorship from Wonder Sciences. Kuhn hosted the event.
 

Your smartphone will see you now

In hearing from 19 speakers, moderators, and panelists—with about a third hailing from outside Stanford—symposium attendees learned that in lieu or in place of mental health-related AI smartphone apps, many potential patients instead engage with general-purpose chatbots.

“People consult AI before or sometimes instead of seeing a doctor,” said Adeli. “This is true for physical health and even more so for mental health.”

Shannon Wiltsey Stirman, PhD, an AI4MH associate director, reported statistics from REAL-CHAT—an ongoing trial exploring how people use AI chatbots for emotional support and mental health—quantifying the phenomenon, with an estimated 24–33% of U.S. adults using LLMs for mental health consultation. “People are turning to these general-purpose, not-built-for-mental-health chatbots quite a bit,” said Stirman, director of the Stanford Center for Responsible and Effective AI Technology Enhancement for PTSD Treatment (CREATE).

A key driver of psychiatric self-medicating is high demand but lack of access to human-rendered treatment. Statistics from the World Health Organization indicate that over a billion people worldwide are living with mental health disorders (chiefly anxiety and depression) amidst a tremendous shortage of mental healthcare workers, with the global median only 13 per 100,000 people.

Meanwhile, recent reports have pegged the global userbase of large language models (LLMs), such as Claude and ChatGPT, at 3.8 billion individuals. “It presents an exciting moment where technology is available—how can we use it responsibly?” asked Carolyn Rodriguez, MD, PhD, associate dean for academic affairs at the Stanford University School of Medicine and the third AI4MH co-director.
 

Keeping therapists in the loop

With these insights in mind, symposium speakers framed the future of AI in mental health through discussions about research, ethics, policy, and more.

In an academic frontiers-focused session, Stirman discussed a “self-driving car” model for incorporating professional expertise into personal-device-mediated psychotherapy. Initially, a clinician would be in the loop—“very much in the driver's seat,” just like the human drivers initially in cars during training to ensure safety, with fully autonomous therapy only if the technology matures sufficiently someday.

A near-term advantageous approach, Stirman said, would be harnessing AI to supplement care for the 99+% of the waking hours when patients are not in typical weekly therapy appointments. In this vein, researchers see ample opportunity to leverage LLMs by analyzing patient interactions to enhance evaluation, tailor therapies, and improve outcomes.

Regulatory considerations loom large in such endeavors, though. In a policy and ethics session, California State Assembly member Mia Bonta—who is legislatively proactive in the emerging mental health AI space—offered perspectives on the challenges, from consumer protection to medical privacy, exacerbated by technology’s rapid evolution. “AI is developing at light speed and we are developing at government speed,” Assemblymember Bonta said during a Q&A segment.

Jane P. Kim, PhD, an AI4MH associate director, discussed the challenge of evaluating the responses provided to users by LLMs in the context of crafting AI policy. “Ideally, [evaluation is] done by people with clinical expertise,” Kim said. “But the volume of responses that need review is so enormous that this quickly becomes infeasible, so human review is a major bottleneck.” As a result, the field has started using LLMs to partly evaluate other LLMs. Kim described a statistics-based methodology she is researching that can keep humans in this evaluation loop by practically determining how many experts will be needed for desired levels of precision. The upshot is “a more standardized and principled approach to evaluating large language models, which I think will be essential for generating claims about safety and utility that can inform policy.”

Relatedly, in an industry-focused session, attendees heard from Sara Johansen, MD, former faculty in the Department of Psychiatry and now leader of product policies in mental health and well-being at OpenAI, the company behind ChatGPT. She described how the company recognizes its rising role in the mental health realm.

“Our models are trained to have safe and appropriate responses in sensitive mental health topics,” Johansen says. Designers have equipped ChatGPT “to be able to recognize risk, both in the moment and over time . . . to identify signals like self-harm, delusions, mania, reliance, and others, by patterns within conversations, and also across conversations.”

When appropriate, the models direct users to supportive resources, from products to “trusted loved ones or mental health professionals,” Johansen said. “We don't believe that ChatGPT should be a replacement for human connection or human care.”

“How do we put the value-add back into human-to-human relationships?” posed panelist Vaile Wright, PhD, senior director for health care innovation at the American Psychological Association. “Human relationships are what make us us.”

The future of AI for mental health will have to accordingly strike a delicate balance of delivering innovative, machine-mediated capabilities with dynamic, human-centered oversight.

In summarizing the day’s discussions, Adeli highlighted this key point, which symposium presenters brought up repeatedly: “AI can't replace the value of human connection," and that mental health will continue to rely on such interaction.