American Psychiatric Association (Marketa Wills)

see attached.

800 Maine Avenue, S.W.
Suite 900 October 15, 2026
Washington, D.C. 20024

Board of Trustees The Honorable Kyle Diamantas, J.D. Acting Commissioner Food and
2026-2027 Drug Administration Department of Health and Human Services
Mark Rapaport, M.D.
President Attention: FDA-2026-N-7874 10903
Rahn Bailey, M.D.
President-Elect New Hampshire Ave. Silver Spring MD 20993
Gabrielle L. Shapiro, M.D.
Secretary
Gia Merlo, M.D., M.B.A., M.Ed. RE: APA’s Comments on the FDA's Digital Health Center of Excellence
Treasurer recent discussion paper and request for feedback, Considerations for the
Theresa M. Miskimen Rivera, M.D. Regulation of Generative AI-Enabled Medical Devices
Ramaswamy Viswanathan, M.D.,
Dr.Med.Sc.
Petros Levounis, M.D., M.A. Dear Acting Commissioner Diamantas:
Past Presidents

Vincenzo Di Nicola, M.Phil., M.D., Ph.D. The American Psychiatric Association (APA), the national medical specialty
Trustee-at-Large society representing more than 40,400 members and their patients,
John C. Bradley, M.D.
Area 1 Trustee
appreciates the opportunity to comment on the discussion paper
David Roane, M.D. “Considerations for the Regulation of Generative AI-Enabled Medical
Area 2 Trustee Devices”.
Kenneth Certa, M.D.
Area 3 Trustee
Dionne Hart, M.D. APA broadly supports the FDA’s proposed framework and agrees that
Area 4 Trustee
benchmarking, shadow deployment, non-intervention testing, clinical
Heather Hauck, M.D.
Area 5 Trustee confirmation, staged deployment, and post-market monitoring are
Lawrence Malak, M.D. consistent with established clinical software and AI evaluation practices,
Area 6 Trustee
Mary Hasbah Roessel, M.D. although they are not always implemented consistently. APA also
Area 7 Trustee supports the FDA’s emphasis on clinically meaningful outcomes rather
Sudhakar K. Shenoy, M.D. than relying primarily on general measures such as accuracy or area
ECP Trustee
Kamalika Roy, M.D., M.C.R. under the curve. Evaluation should address outcomes such as missed
CEA Trustee emergencies, unnecessary escalation, treatment delays, medicationTariq Salem, M.D.
RFM Trustee
related harm, and unequal performance across patient populations.
Karthik V. Sarma, M.D., Ph.D.
RFM Trustee-Elect Risk Assessment and Premarket Evaluation
Assembly
2026-2027 APA supports the FDA’s proposed two-axis framework based on the
Ray C .Hsiao, M.D.
Speaker
activity performed by a device and the consequences of relying on an
James Polo, M.D., M.B.A. incorrect output. However, incorporating psychiatric risk within this
Speaker-Elect framework could be difficult. Suicide risk, for example, cannot be assessed
Anish Dube, M.D., M.P.H.
Recorder through a single objective measure. Risk may change during a
conversation and depend on symptoms, history, substance use, access to
Administration
means, interpersonal circumstances, and other contextual factors. Prior
Marketa M. Wills, M.D., M.B.A.
CEO and Medical Director simulation work on the FDA’s PreCert model suggests that efforts around
risk-based frameworks often fail for mental illnesses 1 and so special considerations should
be embedded.

Risk classification must also account for cumulative harms. Long-term use of AI chatbots
and other tools that engage patients in ongoing conversations may influence a patient’s
beliefs, behavior, relationships, and willingness to seek human care. Repeated interactions
may contribute to excessive reliance, emotional dependency, reinforcement of distorted
beliefs, or parasocial attachment. These risks may not be apparent when individual
responses are evaluated separately. The strongest data demonstrating AI harm in mental
health are from such cumulative harm effects and not short-term use 2. Manufacturers
should therefore evaluate systems across extended conversations and longer periods of
use.

APA supports FDA’s use of benchmarking but recognizes that existing approaches may not
adequately reflect the complexity of psychiatric care or measure whether a system performs
safely in clinical settings. Recent research also found substantial concentration and
inconsistency in the sources cited by consumer AI systems when answering mental health
questions, reinforcing the need to evaluate the quality and reliability of the information
underlying their outputs3. FDA should therefore work with APA and other medical specialty
societies to develop psychiatric benchmarks, risk thresholds, and clinically meaningful
outcomes. Psychiatrists should play a central role in this work rather than leaving these
decisions solely to technology developers or evaluators without relevant clinical expertise.

If psychiatric benchmarks may be used in place of pivotal studies, they should be thorough
and well-vetted before use. They must go beyond current benchmarks and should examine
whether a device can recognize and respond appropriately to suicidality, psychosis, abuse,
intoxication, withdrawal, medication risks, diagnostic complexity, and other clinically
significant situations. Performance measures should evaluate both directions of error. For
example, a system should be assessed for both failure to identify an emergency and
unnecessary escalation that increases patient anxiety or burdens emergency services.

FDA should ultimately support appropriate use of real-world evidence, but higher-risk
psychiatric applications should move first through a rigorous evaluation process. This may
include controlled benchmarking, retrospective testing, shadow deployment, limited use
with close clinician supervision, and prospective clinical confirmation. Each stage should
have predefined safety thresholds, stopping rules, and the ability to pause or reverse
Alon, N., Stern, A. D., & Torous, J. (2020). Assessing the Food and Drug Administration’s risk-based framework for
software precertification with top health apps in the United States: Quality improvement study. JMIR mHealth and
uHealth, 8(10), e20482. https://doi.org/10.2196/20482
Nelson, B. W., Kalinich, M., & Torous, J. (2026). Specialized or general-purpose—The wrong question for mental
health AI safety. JAMA, 336(4), 281–282. https://doi.org/10.1001/jama.2026.11448
Nguyen, P. A., Noorily, J., Flathers, M., Notsu, H., Ospina-Pinillos, L., Nguyen, T., Clark, S., Keane, A., Thompson, G.,
& Torous, J. (2026). Sources of truth: A multi-platform, multilingual audit of citations in AI mental health information
queries [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2609.00319
deployment. More rigorous studies should be required for functions involving autonomous
diagnosis, treatment, prescribing, or crisis management.

Post-market Monitoring

APA supports periodic re-benchmarking, independent clinical review, drift detection, and
reassessment after significant changes. Primary responsibility for monitoring must remain
with the manufacturer because it controls the system’s design and updates, has access to
performance data across deployments, and is best positioned to identify and correct
emerging problems. Clinician reporting should support, but not replace, manufacturer
accountability.

Monitoring should also address:
 Spatial drift: when a device is moved to a different health system, setting, or
population;
 Temporal drift: when populations, language, or clinical practices change over time;
and
 Regime shifts: when performance changes abruptly because of a significant change
in the care environment, such as a new clinical policy, workflow redesign, datasystem change, or model update.

Professional societies can help define clinical risk, appropriate intervention thresholds,
relevant outcomes, and monitoring standards. Both clinicians and healthcare institutions
may report errors, adverse events, and unexpected behavior. However, these roles should
not transfer responsibility for identifying, investigating, and correcting product failures
from manufacturers to clinicians, healthcare institutions, or professional societies.

Manufacturers must be responsible for monitoring adverse events, near misses, subgroup
performance, cumulative harms, and changes following updates to models, prompts,
guardrails, data sources, or workflows. Requirements should be proportionate to the level
of risk, with more frequent monitoring for patient-facing, conversational, action-directing,
and autonomous systems. FDA should also establish accessible reporting pathways that do
not create substantial additional administrative burdens for clinicians.

Regulatory Scope and Data Use

In line with the January 6, 2026, FDA guidance for Clinical Decision Support Software, the
FDA should clarify the line between general software and a regulated medical-device
function, particularly when a product contains both administrative and clinical capabilities.

Low-consequence, nonclinical functions should not be subject to the same requirements as
software that diagnoses, treats, prescribes, directs care, or manages emergencies. For
example, a scheduling function generally should not undergo the full medical-device
process solely because it uses generative AI. However, its risk changes if it interprets
symptoms, prioritizes patients, delays urgent care, or determines an appropriate level of
care. Classification should depend on the function’s actual use and the foreseeable
consequences of failure, not simply how the manufacturer describes the product.

FDA needs to provide clearer distinctions among device functions including administrative
tools, low-risk supportive functions, clinical decision support, patient-specific
recommendations, and autonomous clinical actions. For products with multiple functions,
FDA should evaluate the risk of each function and whether an unregulated function could
affect the safety or effectiveness of a regulated one. When the functions cannot be
meaningfully separated and users must interact with the product as a whole, the product
should be subject to the level of oversight required for its highest-risk function.

The framework should also address how GenAI products collect, retain, derive, and reuse
patient information. AI scribes and other ambient systems may capture entire clinical
conversations containing sensitive information about psychiatric symptoms, trauma,
substance use, relationships, and include people who have not consented to recording.
Secondary use or sale of these data may harm patients even when the information is
described as de-identified.

The above concerns extend to both derived data and metadata, such as how often particular
concepts arise, patterns of language, emotional tone, and interaction histories. Such
information may reveal sensitive clinical details even without traditional identifiers.

To the extent permitted by its authority, FDA should require transparency about what
information a device captures, what derived data it creates, how long information is
retained, whether it is used for model training, and whether it is sold, licensed, or shared.
FDA should consider data minimization, secondary use, and derived information when
those practices may affect safety, effectiveness, patient trust, or real-world performance,
while coordinating with other federal agencies, such as CMS, where appropriate.

APA appreciates FDA’s consideration of these comments and welcomes opportunities to
assist in developing appropriate benchmarks, risk classifications, clinical evaluation
methods, and post-market monitoring standards for GenAI-enabled technologies used in
psychiatric care. Thank you for the opportunity to comment, please contact Zuhal Haidari,
director of digital health, at zhaidari@psych.org with any further questions

Sincerely,

Marketa Wills, MD, MBA, FAPA
CEO and Medical Director
American Psychiatric Association