The US Food and Drug Administration (FDA) Digital Health Advisory Committee recently convened to examine a fast-evolving frontier in care: generative AI–enabled digital mental health devices. During the virtual meeting, chaired by Ami Bhatt, MD, experts discussed clinician needs, regulatory evolution, and the responsible integration of AI within mental health care.
The committee acknowledged the significant promise of generative AI tools for supporting patients with psychiatric conditions. At the same time, members emphasized that human susceptibility to AI-generated outputs, the challenges of monitoring risks such as suicidal ideation, and the potential impact of long-term AI engagement must not be underestimated. The FDA noted that generative AI and large language models (LLMs) continue to show vulnerabilities precisely in areas where human therapists excel, reinforcing the notion that AI should complement, not replace, human therapeutic relationships.
Critical risks
Key themes emerging from the discussion included usability, privacy safeguards, content governance, and the role of clinicians in overseeing AI-enabled care. Generative AI tools offer unparalleled accessibility, available 24/7 and capable of providing a sense of anonymity that many patients value. Yet the committee underscored critical risks: AI systems may confabulate, introduce bias, omit clinically relevant information, or exhibit declining accuracy over time. Devices that operate autonomously also raise new regulatory questions, extending beyond the FDA’s experience with physiologic closed-loop systems.
Digital mental health technologies span mobile health platforms, wearables, telehealth, health IT, digital diagnostics, and digital therapeutics. While the FDA has authorized more than 1,200 AI-enabled medical devices, none are currently approved for mental health indications. Fewer than 20 digital mental health devices without AI have received authorization, highlighting the early stage of regulation in this domain.
Clinical oversight
As generative AI tools proliferate, patient use is rapidly outpacing formal clinical oversight. The committee highlighted the emerging public health implications of patient-facing chatbots that may attempt to diagnose psychiatric conditions, deliver therapeutic content, or implicitly substitute for clinician involvement. Expert opinions varied: some clinicians emphasized heightened risks around vulnerable populations, while others pointed to AI’s potential to expand access to care amid workforce shortages.
The FDA reiterated its commitment to ensuring timely access to safe, effective digital health innovations. As generative AI reshapes mental health care, the agency aims to develop clear regulatory pathways that balance innovation with robust safeguards, acknowledging both the transformative potential of AI-assisted mental health technologies and the complex risks they introduce.
Gen-AI framework
Last month, researchers at the University of Illinois Urbana-Champaign created a generative AI framework to support more personalized and culturally informed mental health care. Led by professor Cortney VanHook, the team used AI to build a detailed, simulated case of a fictional client, a young Black man experiencing depression, to explore how personal context, cultural factors, and access barriers shape care pathways. The system then generated a personalized treatment plan using evidence-based models such as Andersen’s Behavioral Model and Measurement-Based Care, demonstrating how AI can combine patient data and clinical reasoning to produce realistic interventions.
The approach offers a safe, privacy-preserving environment for clinicians, students, and trainees to practice decision-making without using real patient data. The researchers emphasize AI’s potential to illuminate inequities in mental health access, while noting its limitations in capturing emotional nuance. Their framework aligns with Illinois’ new restrictions on AI in mental health, positioning it as a supervised educational tool rather than a clinical decision-maker.
