Society for Digital Mental Health (Stephen Schueller)

SDMH response to Center for Devices and Radiological Health (CDRH) Docket No. FDA-2026-N-7874

October 19, 2026

Via Docket Submission

Michelle E. Tarver, MD, PhD
Center for Devices and Radiological Health
Food and Drug Administration
10903 New Hampshire Avenue Building 66, Room 5407
Silver Spring, MD 20993

Subject: SDMH response to Center for Devices and Radiological Health (CDRH)
Docket No. FDA-2026-N-7874

Dear Dr. Tarver,

The Society for Digital Mental Health (SDMH) is a non-profit organization with membership
from academic research institutions, healthcare systems, payers, digital mental health companies,
and patient advocacy organizations. The mission of the SDMH is to:
● Support knowledge generation and sharing around the design, effectiveness, use,
implementation, and dissemination of digital mental health products and services.
● Foster the development and application of evidence standards and evidence-based best
practices for sustainable, effective implementation of digital mental health
technologies across different care delivery settings and models.
● Promote the equitable use of digital mental health products and services, with a focus on
benefiting underserved and at-risk communities and reducing health disparities.

SDMH appreciates the opportunity to provide comments on the U.S. Food and Drug
Administration's discussion paper, Considerations for the Regulation of Generative AI-Enabled
Medical Devices.

SDMH supports FDA's efforts to consider a regulatory approach that recognizes both the
potential benefits and unique risks of generative artificial intelligence (GenAI). In mental
health, GenAI has the potential to improve access to information and support, facilitate
engagement, augment clinicians, and extend the reach of evidence-based interventions. At the
same time, mental health applications present particular challenges because users may interact
with systems during periods of significant vulnerability, outputs may influence decisions about
safety or treatment, and seemingly low-risk conversational interactions can evolve into
clinically consequential exchanges.

FDA-2026-N-7874 1 of 9
SDMH encourages FDA to consider a risk-proportionate, lifecycle-based, human-centered
framework that evaluates GenAI according to its intended use, potential clinical consequences,
user population, degree of autonomy, and safeguards rather than relying on technology type
alone. Three themes run through the comments below: transparency throughout the user
experience, clarity for developers, and broad access to digital mental health care.

We are pleased to offer the following recommendations organized by the discussion questions in
Appendix B of the paper.

Section IV: Considerations for the Assessment of Risk
Question 1. Dimensions of the two-axis risk framework
SDMH believes the proposed two-axis framework provides a useful starting point but should be
expanded to account for characteristics particularly relevant to digital mental health.
As part of characterizing device activity and the consequences of an incorrect output, FDA should be
sure the following are considered:
● reversibility of the resulting action;
● vulnerability of the intended user population, which in mental health may change within a
conversation;
● immediacy and severity of potential harm;
● availability and accessibility of human or clinical intervention when the output is delivered;
● whether the output can be independently evaluated by the user;
● persistence of an incorrect recommendation across subsequent interactions, including
whether the system builds on it;
● degree of personalization;
● transparency and traceability of the information provided;
● likelihood of user over-reliance or automation bias given how the system presents itself; and
● whether the system is being used in an acute or high-risk context.

For mental health, the framework should explicitly recognize that risk may arise not only from a
discrete incorrect clinical recommendation but also from cumulative interactions. For example, an
apparently benign conversational response could become consequential if a user subsequently relies
on the system for crisis management, medication-related decisions, or decisions about seeking
professional care.

SDMH recommends these serve as modifiers within the framework, so that risk follows the
interaction rather than the product category.

Question 2. Non-directive and action-directing informational functions
SDMH agrees that directiveness exists along a continuum.

In a conversational product, directiveness builds across the exchange. Specificity, personalization,
confidence, and urgency all increase it, and repetition across turns can add up to a directive effect
that no single output carries.

On disclaimers, SDMH’s view is that the effect of a statement such as “talk to your doctor” is an
empirical question. FDA has recommended statements of this kind for digital mental health products
before, including a notice that a device is adjunctive and a direction to contact a clinician. A line
shown at onboarding may not carry much weight at turn forty. Where a manufacturer relies on such
FDA-2026-N-7874 2 of 9
statements as a safeguard, evidence that users notice and act on them is more informative than the
language itself.

Because most digital mental health products operate along this part of the continuum, the criteria
FDA settles on will impact more than premarket evaluation. SDMH encourages FDA to make them
concrete enough to apply to a conversational product.

Question 3. Patient-facing and HCP-facing functions
Patient-facing GenAI functions should not automatically be considered higher risk simply because
patients are the end users. Patients can be capable users of clinical information, and appropriately
designed technologies may improve patient empowerment, engagement, health literacy, and access.
In mental health, where access is limited for many people, that potential is substantial.

SDMH’s membership includes patient advocacy organizations and people with lived experience of
mental illness. People manage their own conditions every day. SDMH’s position is that safeguards
should assume that capability and supply information, and that FDA should weigh patient capability
alongside risk when it considers patient-facing functions.
However, patient-facing systems require safeguards appropriate to the intended population and use
case. These may include:
● clear communication of the system's capabilities, limitations, and uncertainty;
● escalation pathways to human or clinical support;
● safeguards for high-risk situations;
● avoidance of unsupported diagnostic or treatment claims;
● user-centered testing with the intended population; and
● monitoring for patterns of unsafe reliance.

SDMH encourages FDA to treat transparency as part of the whole user experience. What a system
can and cannot do, how confident it is in a given output, and when a user should seek human support
are most useful when they reach the user through onboarding, interface design, notifications, and
controls, at the point where that information supports the decision in front of them. Labeling alone
does not reach the user at that point. FDA’s Transparency for Machine Learning-Enabled Medical
Devices Guiding Principles are a useful starting point.

FDA has described much of this before. The enforcement policy for digital health devices for
treating psychiatric disorders during the COVID-19 public health emergency set out labeling
elements including the condition treated and the intended patient population, the therapeutic method
and recommended duration of use, a summary of clinical testing or an explicit statement that the
device has not been clinically tested, a prominent notice that the device is adjunctive, instructions on
when to consult a clinician and what to do in an emergency, and identification of any functions that
are not FDA cleared. That policy is no longer in effect. The transparency it called for came with
broad access to these products. Those elements remain a useful starting point here, delivered through
the user experience.

For mental health applications, FDA should pay particular attention to crisis situations, suicidality,
self-harm, severe psychiatric symptoms, and other circumstances in which an incorrect or delayed
response may have serious consequences.

FDA-2026-N-7874 3 of 9
Question 4. Generalist and specialist HCP users
SDMH supports consideration of the knowledge and expertise expected of the intended user.
However, the regulatory framework should not assume that all clinicians can independently
recognize errors in specialized GenAI outputs. The relevant question should be whether the intended
user has sufficient expertise, information, and opportunity to identify and appropriately act upon an
erroneous output.

Manufacturers should clearly define the intended user population and evaluate performance in the
contexts in which the device is expected to be used.

For mental health, this may be particularly relevant for technologies intended for primary care,
emergency medicine, behavioral health specialists, or other clinicians who may encounter mental
health conditions without being specialists in psychiatry or psychology.

Question 5. Multi-turn conversational devices
Multi-turn interaction represents one of the most important distinctions between GenAI systems and
traditional software, and central to how digital mental health products are used.

SDMH recommends that FDA evaluate systems across realistic conversational trajectories
alongside isolated prompts. Testing should include transitions from low-risk informational
exchanges to more clinically consequential requests.
Evaluation should consider:
● escalation of user requests across an exchange;
● context accumulation and personalization over time;
● inconsistent or contradictory outputs;
● confidence not supported by the underlying information, including fabricated content;
● attempts to move the system beyond its intended role;
● crisis disclosures, including indirect or late ones; and
● the system's ability to recognize when a conversation has moved beyond its intended use.

Conversational mental health products have a feature most other GenAI-enabled devices do not:
users form a relationship with them. Engagement and therapeutic alliance are mechanisms of benefit
in mental health. The same qualities that build them, warmth, agreeableness, availability, and
memory of what a user has said, can also make a system less likely to challenge a user, more likely
to reinforce what the user brings to it, and harder to disengage from.

SDMH encourages FDA to evaluate this directly, including sycophancy, reinforcement of distorted
or harmful beliefs, and reluctance to escalate where escalation would interrupt the interaction.
Measures that suppress the relationship, such as blunt refusals or repeated disclaimers, can reduce
effectiveness as well as risk, so SDMH encourages FDA to treat this as a question of calibration.

The intended use should encompass reasonably foreseeable patterns of interaction, alongside the
manufacturer's expected use case.

Question 6. Care escalation
SDMH supports evaluating both under-escalation and over-escalation.

In mental health, the consequences of the two errors may vary substantially by circumstance. Failure
FDA-2026-N-7874 4 of 9
to escalate an individual expressing imminent safety concerns may have severe consequences, while
unnecessary escalation of every expression of distress may undermine trust, overwhelm healthcare
resources, or discourage future use.

FDA should therefore encourage manufacturers to define clinically meaningful thresholds and assess
performance across different levels of acuity.

Escalation systems should also be evaluated for appropriate calibration alongside sensitivity.
Where possible, systems should communicate uncertainty and provide users with actionable
pathways to appropriate human support.

Section V: A Competency-Based Approach for Premarket Evaluation

Question 7. Competency-based evaluation
SDMH supports the general concept of competency-based evaluation but recommends that it be
viewed as one component of a broader evidence framework.

Benchmarking can provide valuable information about a system's capabilities, but benchmark
performance alone cannot establish clinical safety or effectiveness. In mental health, real-world
outcomes depend on context, user characteristics, interaction patterns, implementation, and the
availability of human support.

SDMH therefore supports the proposed combination of device benchmarking and clinical
confirmation, provided that the evidence requirements remain proportionate to intended use and risk.

Question 8. Using the risk framework to determine the level of evidence
SDMH supports using the risk framework to set the level of evidence expected.

Mental health technologies run along a continuum from general information and self-management,
through clinical decision support, to treatment and crisis intervention. Evidence expectations that
scale along it would let lower-consequence informational functions rely largely on benchmarking,
usability testing, and human factors work, with clinical confirmation increasing as the clinical
consequence does.

Question 9. Adequacy of benchmarking
Benchmarking should assess multiple dimensions of performance, including:
● clinical knowledge and accuracy;
● reasoning and analytical capabilities;
● safety behavior;
● communication quality and user comprehension;
● uncertainty calibration;
● robustness and generalizability; and
● performance across clinically meaningful subgroups.

For mental health, benchmarks should also evaluate whether the system appropriately recognizes
situations requiring escalation or human intervention.

Validated benchmarks do not exist for the behaviors that matter most here, including crisis
FDA-2026-N-7874 5 of 9
recognition, refusal, and therapeutic appropriateness, so sponsors will often build their own. That
makes it more important to show the benchmark measures what matters for the intended use, and to
guard against optimizing to one’s own test set. SDMH supports protected or independently
developed evaluation sets for higher-risk applications.

SDMH cautions against relying solely on accuracy metrics. A system can generate a factually
plausible response while still being clinically inappropriate, misleading, stigmatizing, or unsafe.

Question 11. Clinical confirmation
Clinical confirmation should be proportionate to risk and intended use.

A prospective clinical study should not necessarily be required for every GenAI-enabled device. The
absence of a prospective study does not mean there is no meaningful clinical evidence.
Potential evidence sources may include:
● retrospective clinical evaluation;
● simulation;
● structured expert review;
● prospective observational evaluation;
● human-AI interaction studies;
● implementation studies; and
● real-world evidence.

The selected approach should be justified based on the expected clinical consequences of failure and
the extent to which other evidence adequately addresses uncertainty.

Question 13. Synthetic data
Synthetic data may be valuable for testing rare scenarios, privacy-sensitive situations, edge cases,
and safety events that are difficult to capture in sufficient numbers in real-world datasets.

However, synthetic data should not substitute for real-world evidence where real-world context
materially affects performance.

This is particularly important in mental health, where language, culture, socioeconomic
circumstances, lived experience, communication style, and clinical context may affect system
performance.

FDA should encourage evaluation across diverse populations and ask sponsors to demonstrate that
synthetic data do not mask or reproduce performance gaps affecting underrepresented populations,
particularly where the synthetic data come from models of the same class as the device.

Question 14. Open-ended outputs and performance comparators
Open-ended outputs require evaluation methods beyond simple answer matching.

SDMH supports the use of qualified clinical experts as one component of evaluation but
recommends that the comparator be defined according to the intended clinical use.

FDA-2026-N-7874 6 of 9
In some cases, the appropriate standard may be the accepted standard of care; in others, the relevant
question may be whether the human-AI team performs better than unaided clinical judgment.

For mental health, FDA should recognize that variability among qualified clinicians does not
necessarily mean that a particular output is unsafe. Evaluation frameworks should therefore
incorporate clinically meaningful criteria rather than treating a single human response as an absolute
ground truth.

Question 15. Comparison with care in the absence of the device
SDMH supports evaluating the device against the realistic alternative in the absence of the
technology.
Depending on the use case, this may include:
● unaided clinical judgment;
● usual care;
● delayed access to a specialist;
● no intervention; or
● existing digital or non-digital support.

This is particularly important in mental health, where the relevant alternative is frequently limited
access to care or none. A comparator drawn from an idealized clinical encounter understates the
benefit side, and SDMH encourages FDA to weigh that explicitly. The appropriate comparator
should be prospectively justified and should reflect actual conditions in the intended-use
environment.

Section VI: Postmarket Monitoring
Question 18. Greater reliance on postmarket monitoring
SDMH supports a lifecycle approach in which appropriate postmarket evidence may complement,
but should not unnecessarily replace, premarket evidence.
Greater reliance on postmarket monitoring may be appropriate when:
● risks can be effectively detected after deployment;
● potential harms are reversible or manageable;
● meaningful performance measures are available;
● monitoring can occur at sufficient scale and frequency; and
● effective corrective actions are available.

Where the consequences of failure are severe, rapid, irreversible, or difficult to detect, meaningful
premarket evidence remains important. For high-risk mental health applications, including those
involving crisis response or treatment recommendations, SDMH encourages FDA to maintain
premarket evidence expectations. The areas calling for the most rigor here are narrow: response to
safety-critical disclosures, and control over third-party model changes.

Question 20. Machine-based supervisory agents
Machine-based supervisory systems may eventually provide useful scalable monitoring, but they
should not be assumed to be inherently reliable.

A supervisory agent is itself an AI-enabled system and therefore requires validation, monitoring, and
defined limitations. For high-risk mental health applications, automated monitoring should
FDA-2026-N-7874 7 of 9
complement rather than completely replace qualified human review.

Question 21. Roles of clinicians and other stakeholders
Clinicians, healthcare institutions, professional societies, researchers, patients, and people with lived
experience can contribute meaningfully to postmarket monitoring.
Their roles may include:
● identifying emerging failure modes;
● reviewing clinically significant outputs;
● contributing to evaluation standards;
● identifying population-specific risks;
● informing usability and implementation assessments; and
● helping define clinically meaningful performance thresholds.
SDMH supports structuring these roles so that manufacturer accountability remains clear.
Stakeholder involvement would supplement the manufacturer's responsibility for the safety and
effectiveness of the device.

Question 24. Changes initiated by third-party foundation model developers
Many digital mental health products are built on third-party foundation models the developer does
not control.

A model change can alter the behaviors that matter most here, including recognition of safety-critical
disclosures, refusal, escalation language, and tone, with no action by the manufacturer and no signal
visible to users or clinicians, and routine functional testing may not surface it.

Useful mechanisms include contractual notification, version controls, regression testing against the
safety-critical behaviors established premarket, performance thresholds, rollback, and PCCPs
covering anticipated model changes. The manufacturer remains accountable for the device it
markets, and SDMH encourages FDA to consider mechanisms that improve transparency between
model developers and manufacturers.

Section VII: Other Topics
Question 26. Agentic GenAI-enabled devices
SDMH believes agentic GenAI systems require additional consideration because they may
autonomously perform multi-step actions, use external tools, retain or act on information across
interactions, and have fewer opportunities for human review.

The level of oversight should increase with the system's degree of autonomy and the potential
consequences of its actions.
Evaluation should address:
● boundaries on autonomous action and the controls that enforce them;
● escalation to human oversight before high-consequence actions;
● tool-use safety and recovery from tool failures;
● persistence and memory;
● unintended action chains;
● cybersecurity;
● monitoring of autonomous behavior;
● the ability to interrupt or disable the system; and
● performance under unexpected or adversarial conditions.

FDA-2026-N-7874 8 of 9
In mental health, SDMH encourages caution regarding agentic systems performing highconsequence clinical functions independently, unless there is evidence of safe performance and that
appropriate human oversight and safeguards are in place.

Conclusion
SDMH appreciates FDA's leadership in seeking early input on the regulation of GenAI-enabled
medical devices. The rapid development of these technologies makes it important to establish a
framework that protects patients without unnecessarily restricting responsible innovation.
For digital mental health in particular, SDMH recommends that FDA's approach be grounded in five
principles:
1. Risk assessment should follow the intended use, the clinical consequences, and the
interaction itself rather than the product category.
2. Evidence expectations should be proportionate to risk, include technical and clinical
evidence, and account for the care available without the technology.
3. Transparency should run through the user experience, in onboarding, interface, notifications,
and controls, alongside patient-centered design, human factors, and appropriate human
oversight.
4. Postmarket monitoring should be treated as an essential component of the AI lifecycle,
particularly for systems that evolve after deployment.
5. Regulatory approaches should preserve innovation and access while maintaining clear
manufacturer accountability for safety and effectiveness.

SDMH particularly encourages FDA to recognize that mental health technologies may operate
across a continuum from general information and self-management to clinical decision
support, treatment, and crisis intervention. A regulatory framework that accounts for this
continuum, rather than treating GenAI-enabled devices as a homogeneous category, will better
support both patient safety and responsible innovation.

SDMH’s members work across research, care delivery, and product development in mental health,
and we share FDA’s interest in public health. We were glad to see “Generative Artificial
Intelligence-Enabled Conversational Devices for Mental Disorders: Evidentiary Recommendations”
on CDRH’s FY2027 A-list of proposed draft guidance topics, and hope these comments are useful to
that work as well.

SDMH looks forward to continued engagement with FDA, clinicians, researchers, technology
developers, patients, and other stakeholders as the regulatory science surrounding GenAI-enabled
medical devices evolves.

Respectfully, on behalf of the SDMH Board of Directors,

Stephen Schueller, PhD
President, Society for Digital Mental Health

FDA-2026-N-7874 9 of 9