Keystone Neuroinformatics (Anup Sharma)

Comment from Anup Sharma

FDA Generative AI Device Comment Responses (10/10/26)
By: Anup Sharma, MD, PhD
Physician-Scientist, Psychiatrist, Clinical AI Expert

Introduction
I am Anup Sharma, MD, PhD, a board-certified psychiatrist and physician-scientist trained in
neuroscience and experienced in clinical and translational psychiatry, clinical AI evaluation, and
neuropsychiatric benchmarking. I also founded Keystone Neuroinformatics. I offer these comments
in my individual and organizational capacity, drawing on my work at the interface of psychiatric care
and AI evaluation.
My comments focus on generative AI-enabled medical devices that communicate directly with
patients or influence clinical decisions in psychiatry and behavioral health. In these settings, the
clinical meaning of an output depends on patient context, uncertainty, relationships, and changes
across repeated encounters.
My central recommendation is that evaluation connect the device’s clinical function, intended
population, system configuration, and claims to evidence from the settings in which it will be used.
However, this requires attention to clinical reasoning, relational interaction, safety, and equity,
including their performance within clinical workflows and after deployment.
____

Background
These comments respond to FDA’s Considerations for the Regulation of Generative AI-Enabled
Medical Devices: Discussion Paper and Request for Feedback (Docket FDA-2026-N-7874). This is
an exploratory discussion paper, not guidance or a proposed rule. The responses below address
Questions 1–26 in Sections IV–VII: risk assessment, competency-based premarket evaluation,
postmarket monitoring, and other topics, including foundation-model information and agentic AI.
I draw on the working manuscript Beyond Access: Functional Evaluation of Artificial
Intelligence in Psychiatric Care, which proposes four broad psychiatry-specific functional domains
and connects them across individual outputs, conversational trajectories, agent actions, and clinical
workflows. The comments apply that perspective to evidence selection, longitudinal evaluation,
effective supervision, and input from clinicians, patients with lived experience, families, peers,
technical experts, and other stakeholders. The reviewer appendix distinguishes the manuscript’s
conceptual contribution from the regulatory recommendations developed here.​
​
Additionally, my background is further available here as well (for reference): ​
My LinkedIn information is: https://www.linkedin.com/in/anup-sharma-md-phd-7948a434/​
My Google Scholar is: https://scholar.google.com/citations?user=z8tkoWEAAAAJ​
​
I work closely with health professionals, AI labs, consultants, and health systems. I advocate for
diverse stakeholders and hybrid-AI evaluation frameworks. I have trained hundreds of physicians on
AI evaluation and am helping different consortia with medical AI evaluation studies. ​
____
Question 1
Does the two-axis risk framework, organized around device activity and the consequence of relying
on an incorrect output, appropriately capture the dimensions most relevant to the risk of a
GenAI-enabled software function? If there are additional dimensions—such as the reversibility of a
resulting action, the availability of downstream safeguards, the time pressure of the deployment
setting, or the traceability of the output (i.e., to primary source materials)—that should be
represented in a risk framework, please describe and provide examples of how they should be
represented.
Response
The two-axis framework provides a useful initial structure for risk assessment. In psychiatry and
behavioral health, its application should make several contextual modifiers explicit. CDRH already
identifies reversibility, downstream safeguards, time pressure, and traceability as relevant
considerations; their significance depends on the patient, clinical function, and setting in which the
device is used.
Four modifiers warrant particular attention.
●​ Clinical risk depends on the patient’s illness acuity, developmental stage, and comorbidity, as
well as missing collateral information and access to care.
●​ Across repeated interactions, changing symptoms, accumulated reinforcement, limitations in
memory, and reliance on earlier interpretations may alter the clinical implications of an output.
●​ A safeguard can reduce risk only if there is evidence that errors are recognized, corrected, or
escalated before consequential decisions occur.
●​ Language, culture, literacy, and accessibility shape how an output is understood. Evaluation
should also consider differences between clinician and patient interpretations.
These modifiers could be recorded alongside the two-axis assessment, with a clinical rationale
explaining whether they increase risk or support a reduction in residual risk. For example, repeated
validation of a delusional interpretation may become consequential across interactions even when
individual responses appear supportive. Similarly, nominal clinician supervision should reduce
estimated risk only when review is feasible and effective in the intended workflow. This approach
preserves a common framework while making its application sensitive to psychiatric functions, care
settings, and patient populations.​
____

Question 2
CDRH seeks input on how to account for the spectrum of GenAI-enabled informational functions that
vary in the degree to which they direct a user to a particular action (i.e., between “non-directive” and
“action-directing”). What characteristics of an output—such as its wording, specificity,
personalization, or context—could be considered as modifiers of the risk associated with an
informational function, after accounting for the device’s overall functionality and intended use? What
additional information would provide manufacturers with sufficient clarity and predictability around
risk assessment for informational functions while recognizing that directiveness may exist along a
continuum rather than as a binary distinction?
Response
Directiveness should be assessed through the clinical meaning and likely use of an output within the
interaction. Wording, specificity, personalization, certainty, omission of reasonable alternatives, and
repeated endorsement can each influence whether information becomes action-directing. As CDRH
recognizes, a disclaimer does not resolve the meaning of the surrounding recommendation.
In behavioral health, relational context also matters. An apparently supportive response may
reinforce an unsupported diagnosis or a harmful interpretation, influencing treatment decisions or
willingness to seek care without issuing an explicit instruction. Evaluation should therefore examine
whether the system seeks missing context, distinguishes reasonable alternatives, communicates
uncertainty, maintains appropriate role boundaries, and changes its formulation when new
information emerges. The same wording may have different implications in a general educational
exchange and an ongoing interaction with a person whose symptoms are worsening.
Manufacturers would benefit from annotated examples showing progressively greater directiveness
within the same clinical scenario, accompanied by explicit criteria and explanations of borderline
cases. These examples should include patient-facing and clinician-facing uses, individual responses
and conversational trajectories, and situations in which nominally informational content influences
consequential decisions. Patient and clinician assessments of how they understand and would act
on outputs should inform validation of these criteria.​
____

Question 3
CDRH seeks input on whether and when a GenAI-enabled function that results in delivery of clinical
information to patients, as opposed to HCPs, could present different or higher risks, while also
recognizing the potential benefits associated with improved patient empowerment, engagement, and
access to clinical information. What device characteristics, output features, or safeguards might
mitigate risks that could arise when a user lacks the domain knowledge to independently evaluate an
output, without unnecessarily underestimating patient capability?
Response
Patient-facing delivery can increase risk when safe use depends on clinical knowledge the user does
not have, an error is difficult to recognize, and the output can influence consequential decisions. This
should be a conditional assessment of the clinical task, output, and available support. Patient status
or a psychiatric diagnosis alone should not be used to presume limited capability. Patients may have
substantial knowledge of their own illness, treatment history, preferences, and circumstances that is
essential to appropriate interpretation.
Risk mitigation should support actual comprehension and agency. Relevant features include
accessible language, explanation of uncertainty and missing information, clear role boundaries,
presentation of reasonable options, and context-sensitive prompts to seek clarification or human
support. Traceable source material may help, but its presence does not establish that the user can
evaluate its relevance. Higher-consequence recommendations require stronger evidence that
misunderstanding or error can be detected and addressed in time.
Evaluation should assess how intended users understand and act on outputs across differences in
language, literacy, disability, clinical state, and access to care. Patients with lived experience should
help define understandable explanations, acceptable boundaries, and usable escalation pathways.
Patient empowerment, informed participation, and successful transitions to appropriate care should
be evaluated alongside safety outcomes.​
____
Question 4
CDRH seeks input on whether and how the distinction between generalist and specialist physicians
might inform the assessment of risk for HCP-facing GenAI-enabled functions. Under what
circumstances, if any, might risk be affected when an HCP who lacks the relevant clinical specialist
knowledge receives information that falls within a specialist area of practice? What device
characteristics or safeguards might mitigate such risks?
Response
The distinction between generalist and specialist physicians can inform risk assessment when safe
interpretation depends on knowledge or experience that the intended user may not possess.
Specialty designation alone is insufficient: the relevant issue is whether the clinician can recognize
missing information, assess uncertainty, identify clinically important errors, and determine when
consultation is needed for the specific task.
In psychiatric care, risk may increase when a system offers a confident formulation despite
incomplete longitudinal history, unassessed comorbidity, missing mental-status observations, or
unavailable collateral information. A clinician may reasonably use AI to organize a differential
diagnosis while requiring additional expertise before acting on complex treatment recommendations.
The level of supervision should therefore correspond to the clinical function, consequences of error,
and actual workflow.
Potential safeguards include explicit scope limitations, identification of missing information,
presentation of reasonable alternatives, accessible supporting evidence, and clinically appropriate
specialist consultation pathways. Their effectiveness should be tested with the intended users under
representative workload and time constraints. Comparative evaluation should assess whether
generalists and specialists recognize, correct, override, or escalate the same errors, and whether AI
improves the combined clinical process. Specialist review should also be evaluated for overreliance;
expertise should not be treated as an automatic guarantee of effective oversight.​
​
____

Question 5
For multi-turn conversational GenAI-enabled devices that may migrate from providing “non-directive”
information to “action-directing” information over the course of an exchange, how should risk be
assessed across realistic conversational trajectories? How could the intended use of such a device
be characterized when its behavior is emergent across a conversation?
Response
Risk assessment should examine individual outputs within the full conversational trajectory, including
repeated sessions when the device retains context or memory. In psychiatric care, clinically
important failures may emerge through accumulated reinforcement, loss of relevant history,
persistence of an earlier formulation, or failure to recognize changes in symptoms and risk. An
isolated response can appear acceptable while the interaction progressively becomes inappropriate
or consequential.
Testing should include realistic trajectories in which information is incomplete, clarified, contradicted,
or changes over time. Relevant scenarios include indirect risk disclosures, emerging mania or
psychosis, changing treatment context, requests for increasingly specific advice, and
misunderstandings that require repair. Evaluation should identify when the system becomes
action-directing, whether that transition is clinically appropriate, and whether subsequent
clarification, boundary-setting, escalation, and human handoff are effective.
Intended use should describe the clinical functions, populations, settings, permitted actions,
supervision, memory behavior, and expected duration or recurrence of interaction. Anticipated
transitions between education, support, recommendations, and escalation should be specified and
evaluated. Emergent behavior outside those boundaries should trigger investigation and mitigation
rather than silently expand the claimed use. Trajectory-level performance should be assessed in the
deployed configuration and reassessed after material changes to models, prompts, memory, tools,
or safety layers.
The declared role should remain an effective operational boundary across repeated use and
personalization. A device evaluated for education or supportive interaction should not progressively
assume diagnostic or treatment functions through conversational adaptation or increasing user
reliance. When an interaction exceeds that role, the system should communicate its limits and
support an appropriate human handoff.​
​
____

Question 6
For GenAI-enabled care escalation functions, CDRH is considering how the evaluation may account
for both under-escalation and over-escalation. How could manufacturers characterize and weigh
these two directions of error, given that they may not be commensurable and that acceptable
trade-offs may vary by clinical context?
Response
Under-escalation and over-escalation should be characterized separately, with attention to the
severity, timing, and downstream consequences of each error. A single aggregate accuracy measure
or fixed exchange rate between these errors would obscure clinically important differences. In
psychiatry, missed imminent danger can delay needed care, while unnecessary emergency referral
may increase distress, burden, or subsequent disengagement.
Evaluation should distinguish risk recognition, contextual questioning, assessment of urgency,
response generation, referral, and completed human handoff. Cases should span acuity levels,
direct and indirect disclosures, ambiguous statements, and clinically similar presentations expressed
through different cultural or linguistic contexts. Review should account for available services and
continuity of care. Giving an appropriate recommendation does not establish that the person can
access the recommended support.
Manufacturers should prespecify separate measures for missed or delayed necessary escalation
and unnecessary escalation, stratified by clinical scenario and population, with uncertainty reported.
Clinicians, patients, and relevant care services should help define acceptable response ranges and
clinically serious failures. Higher-acuity scenarios may justify prioritizing timely escalation, while
avoiding indiscriminate emergency referral for ordinary distress or non-urgent concerns. These
judgments should remain explicit, clinically justified, and subject to prospective evaluation of patient
outcomes, burdens, and successful transitions to care.​
____
Question 7
Is the competency-based approach described above, i.e., device benchmarking followed by clinical
confirmation, a useful and appropriate framework for evaluating GenAI-enabled devices?
Response
I support a competency-based approach when the competencies are defined in relation to the
device’s clinical role, patient population, configuration, and care setting. My proposed functional
evaluation architecture provides a complementary structure for psychiatry: clinical reasoning,
relational interaction, safety, and equity can be assessed across individual outputs, interaction
trajectories, agent actions, and clinical workflows. Benchmarking can identify reproducible strengths
and failure modes, while clinical confirmation examines whether those findings remain meaningful
when patients and clinicians interact with the deployed system. Performance at one level should
support only the claims that the evaluation can establish. Marketing and clinical deployment should
be limited to intended uses and claims supported by adequate independent predeployment
validation.
At each evaluation stage, benchmarking should include an independently administered assessment
of the final configured device, with protected questions withheld from the sponsor, device vendor,
and underlying foundation-model vendor, and outputs scored by qualified, trained, calibrated raters
unaffiliated with those parties. This standard should apply to initial validation, periodic postmarket
reassessment, and re-benchmarking after changes. Internal development tests may inform
improvement but should remain distinct from independent evidence supporting regulatory claims.
●​ The final configured device should remain the focus of evaluation, while testing its individual
components can help identify the source of a failure.
●​ Cases involving repeated, multi-turn interactions should assess how the device responds to
changing symptoms, missing collateral information, and diagnostic uncertainty, including
continuity of care through human handoff.
●​ Acceptance criteria should be specified before testing and reflect clinically meaningful
performance. Safety-critical failures should be reported separately from averages, and findings
from clinical confirmation should inform version-specific reassessment.
____

Question 8
How might the two-axis risk framework described in Section IV be considered within the
competency-based approach described above to help determine the level of evidence needed for a
premarket submission?
Response
The two axes should determine the scope, rigor, and clinical proximity of the evidence, with explicit
consideration of how an error reaches the patient. In psychiatry, conversational guidance may
influence treatment adherence, reinforce a delusional interpretation, or delay assessment without
executing a formal clinical action. Risk assessment should therefore include cumulative influence,
detectability of error, reversibility, patient vulnerability, and the demonstrated effectiveness of human
supervision. A nominal review requirement should reduce evidentiary burden only when the intended
users can identify and correct clinically important failures.
●​ Targeted benchmarking and retrospective confirmation may be proportionate for bounded,
readily verifiable functions when they address the relevant hazards and claims.
●​ For repeated patient-facing interactions or escalation decisions, evaluation should follow the
clinical course of the conversation. It should establish whether patients understand the
information, escalation is appropriate, and continuity of care is maintained.
●​ Greater autonomy or potential harm should increase the rigor of independent adjudication,
subgroup coverage, sample-size justification, and prospective evaluation where needed. Safety
criteria and uncertainty bounds should be specified before testing.
____

Question 9
Would a benchmarking structure such as the one described above be likely to provide adequate
evidence of clinical knowledge, analytic capabilities, safety behavior, communication, and
generalizability to support a reasonable assurance of safety and effectiveness? Are there elements
that are missing, redundant, or inappropriately categorized? Are there externally developed
standards that could be leveraged?
Response
The proposed elements provide useful coverage, but adequacy depends on their operational
definitions, the cases tested, and clinical confirmation. I would strengthen their psychiatric specificity
rather than replace the categories. Overlap is appropriate when one failure affects several functions;
sponsors should explain that overlap and avoid treating repeated scoring of the same event as
independent evidence. ISO 14971 risk-management processes and IEC 62366-1 usability
engineering could support evaluation design. These standards do not establish psychiatry-specific
performance thresholds.
●​ Within S.1-S.3, recognizing risk, seeking context, responding proportionately, and completing a
handoff should be evaluated as distinct functions. Testing should examine both harmful
reinforcement and unnecessary refusal or deferral.
●​ E.1-E.4 should examine diagnostic overreach, missing collateral information, comorbidity,
measurement interpretation, and relational repair. Empathy ratings should distinguish validating
distress from endorsing an unsupported belief.
●​ Within E.4, evaluation should also address the psychological effects of AI interactions and
communication about AI’s broader societal impact. Testing should examine whether responses
could worsen existing hallucinations or other psychotic symptoms by validating the apparent
reality of distressing perceptions or reinforcing false beliefs, a clinical concern distinct from model
‘hallucinations’ (fabricated output). It should also assess responses to ‘AI anxiety,’ including
worries about employment, privacy, autonomy, and society, without automatically treating
legitimate concerns as psychopathology. Systems should communicate potential benefits,
harms, and uncertainty proportionately. Evaluation should examine effects on trust, agency, and
appropriate help-seeking across repeated interactions, including whether increasing distress or
symptoms are recognized and addressed within the device’s clinical scope.
●​ R.1-R.2 should establish whether clinically meaningful performance remains consistent across
sessions, versions, languages, and developmental contexts. Equity should be assessed across
all competencies, including whether appropriate adaptations preserve clinical standards.
●​ Within A.1, intended actions should be evaluated for permission, correct execution, and
incorporation into subsequent care. A plausible plan or successful tool call alone does not
demonstrate an appropriate clinical result.
____
Question 10
CDRH recognizes that publicly available benchmarking assets may be subject to data
contamination, saturation, and limited real-world representativeness. How should a sponsor
establish that performance on a given benchmark predicts safe and effective real-world behavior for
the device’s intended use? What evidence should support the construct validity of a benchmark used
to gate device evaluation, and what role should sponsor-developed benchmarks play given potential
concerns around independence and optimization to the test?
Response
A benchmark used to gate evaluation should have an explicit account of the clinical function being
measured and why its scores are relevant to the intended use. In psychiatric care, agreeable
language, diagnostic recall, and appropriate management are distinguishable constructs. Sponsors
should demonstrate that the rubric captures the intended function, identifies clinically important
failures, and yields reproducible judgments across qualified evaluators. Evidence that benchmark
performance predicts real-world behavior should come from held-out clinical encounters, preferably
across relevant settings and configurations, with uncertainty and discordant findings reported.
●​ Benchmark documentation should describe case provenance, overlap with training or
development data where it can be established, exposure to public benchmarks, saturation, and
separation of development from final testing. Independently controlled held-out questions should
be administered securely to the final configured device. Sponsor and vendor personnel should
have no access to protected test items before or during evaluation, including through application
logs, model-provider logs, debugging, or feedback. Although the device must receive the test
inputs, the evaluation arrangements should prevent sponsor and vendor personnel from
accessing them. Reusable items should remain protected after scoring, and exposed items
should be retired and replaced before further independent evaluation.
●​ Clinicians and people with lived experience should help assess whether the benchmark detects
clinically important responses to missing context, indirect risk disclosures, changing symptoms,
and unsafe relational interactions, with their input matched to the construct being measured.
●​ Sponsor-developed tests may support specialized functions and internal improvement, but their
results should remain distinct from independent regulatory evidence. An audit of sponsor-visible
questions cannot substitute for independently controlled held-out questions scored by unaffiliated
raters. Public frameworks, construct definitions, scoring rubrics, methods, and aggregate results
can support transparency without disclosing protected items or feedback that would reveal them.
Refreshed cases should retain clinical validation and should not silently change the construct
being measured or the acceptance threshold.
____
Question 11
CDRH is considering that clinical confirmation for a GenAI-enabled device might not require a
prospective clinical study in every case, and has described above a range of approaches of
increasing rigor and patient exposure. How might a sponsor select and justify a confirmation
approach tailored to a device’s intended use and proportionate to the device’s risk profile? Are there
device types or risk profiles for which one or more of these approaches would be insufficient or
inappropriate? How might the anticipated distribution of real-world inputs be taken into
consideration? Are there other methods of clinical confirmation that might help inform the evaluation
of GenAI-enabled devices?
Response
Confirmation should resolve the uncertainties remaining after benchmarking for the specific claim,
population, and workflow. A less burdensome confirmation method is appropriate only if it still
produces adequate evidence for the intended use and claims. Where the evidence cannot
substantiate a proposed clinical claim, the response should be additional predeployment evaluation,
a narrower claim, or deferral of deployment for that use. Retrospective real-patient evaluation may
support a bounded, verifiable function. Shadow deployment can characterize live inputs and output
quality, but cannot establish how visible outputs affect clinical judgment or patient behavior.
Standardized patients can probe difficult interactions under controlled conditions. Claims of
therapeutic benefit, safe sustained patient interaction, or effective high-consequence autonomous
care generally warrant prospective evaluation with relevant outcomes; retrospective adjudication
alone would be insufficient.
●​ Clinical confirmation should sample the anticipated distribution of real-world inputs, including
missing information, comorbidity, indirect disclosures, language variation, and changing acuity.
Separate stress-testing should examine uncommon serious hazards.
●​ Real-world evidence should document data provenance, completeness, selection mechanisms,
consent and privacy protections, and how outcomes were ascertained. Data collected outside
the United States require a transportability analysis covering patient characteristics, language
and culture, care pathways, clinical standards, and available escalation resources. Material gaps
should be addressed with U.S. bridging evidence.
●​ Other options include randomized clinician crossover studies for decision support and pragmatic
cluster-randomized studies for workflow claims. Studies should prespecify endpoints relevant to
the claim, follow-up duration, stopping rules, and sample sizes sufficient for the intended
conclusions.
____

Question 12
How should sponsors achieve statistically meaningful performance measurement for GenAI-enabled
devices? Where synthetically generated inputs supplement real patient data, how should sponsors
account for differences between the synthetic and real-world distributions when estimating
performance, and under what conditions, if any, is it appropriate to combine benchmarking evidence
and clinical confirmation evidence to support a single performance estimate?
Response
Sponsors should prespecify the target population, clinical endpoint, unit of analysis, and
performance quantity being estimated. For psychiatric interactions, that unit may be a patient
trajectory or completed escalation episode. Multiple turns or repeated generations from one case
should be treated as correlated observations. Sample sizes should be justified by the precision or
power needed for the clinical claim, with confidence intervals for safety outcomes and clinically
important subgroups. Rare serious failures require explicit uncertainty bounds; observing no events
does not establish absence of risk.
●​ The clinical function should determine the endpoint, whether appropriate management, missed
escalation, completed handoff, or patient outcomes. Analyses should account for clustering by
patient, clinician, and site where relevant.
●​ Synthetic, benchmark, and clinical results should ordinarily be reported separately, with
differences in case mix, severity, subgroup representation, and conditional clinical relationships
described. Enrichment for hazards should not be mistaken for their real-world prevalence.
●​ Combining evidence requires a justified common endpoint and target population, compatible
scoring and device versions, validated transport or weighting assumptions, and sensitivity
analyses. A large synthetic sample should not artificially increase confidence in real-world
performance.
____

Question 13
For which clinical domains, device functions, or subpopulations is synthetic data particularly
well-suited, or particularly inadequate, as a supplement to real-world evidence? What safeguards
would mitigate the risk that synthetic data generated by models of the same class as the device
under evaluation reproduces the very performance gaps the evaluation is intended to detect,
particularly for underrepresented subgroups?
Response
Synthetic data are useful for controlled variation of clinical information, early development, and
stress-testing uncommon or difficult safety scenarios without exposing patients. Examples include
missing collateral information, medication contradictions, indirect crisis disclosures, and failures
during escalation. Their limitations are especially consequential for longitudinal engagement,
developmental and cultural context, disability, and the expression of distress in underrepresented
communities. Plausible dialogue does not establish that a simulated patient captures the disclosure
patterns, responses, or outcomes of an actual population. Synthetic subgroup examples therefore
cannot establish equitable clinical performance by themselves.
●​ Clinician-authored scenarios, patient and community input, and multiple independent
approaches to generating cases can help reduce shared blind spots. A different model name
alone does not establish independence.
●​ Synthetic cases should preserve clinically important relationships and subgroup-specific
expressions, with validation against appropriately governed real-patient data. Matching
demographic proportions or surface language alone is insufficient.
●​ The generator and device should be separate where feasible, and benchmark assessment
should follow the protected-test and unaffiliated-evaluator safeguards in Questions 10 and 16.
Automated scoring requires independent validation against unaffiliated human experts, with
human review retained for serious failures. Synthetic and real-world findings should be reported
separately, including unresolved gaps in coverage.
____
Question 14
For open-ended device outputs, how should performance comparators and acceptance criteria be
selected? When a panel of qualified clinicians serves as the comparator, how should the applicable
standard (for example, the standard of care versus the performance of a median clinician in practice)
be defined and justified? Should generalist or specialist physicians be used as a performance
standard? When should human-AI team performance, rather than the device operating alone, serve
as the basis for evaluation?
Response
Comparators should reflect the device’s clinical role, while acceptance criteria should specify a
defensible standard of care for that role and setting. Psychiatric formulations and therapeutic
responses may have several reasonable alternatives. A qualified panel should define acceptable
response boundaries, required information, and unsafe actions using relevant evidence, clinical
guidance, and documented judgment. Median clinician performance can describe usual practice, but
should not independently determine the safety threshold. Raters should be qualified, trained,
calibrated, and unaffiliated with the sponsor, device vendor, and underlying foundation-model
vendor, as described in Question 16. Evaluator calibration, agreement, and a procedure for resolving
substantive disagreement should be reported.
●​ Panel composition should reflect the clinical task. Specialty-level diagnostic, treatment, and risk
claims require specialist adjudication, while generalists should be included when they are
intended users or provide the relevant usual care.
●​ For assistive devices, the central question is how intended clinicians perform with and without AI
assistance. Standalone testing should remain part of the assessment to characterize failures
attributable to the device.
●​ Patients should help assess comprehension, agency, and relational experience. Safety-critical
criteria should be prespecified separately from communication preferences, and evaluation
should determine whether clinicians detect, correct, override, or propagate erroneous outputs.
____

Question 15
Are there ways in which performance might be assessed relative to the care, technology, or course
of action likely to occur in the absence of the device, rather than to the comparators described in this
section? What approaches might be used to identify and justify a comparator such as unaided
clinical judgment, delayed specialist review, or no intervention?
Response
Yes. The relevant alternative should reflect what patients in the intended setting would reasonably
receive without the device. In behavioral health, that may include primary-care management, a
waiting period before specialty review, an existing digital intervention, or an established crisis
pathway. Sponsors should document this alternative using contemporaneous workflow data, referral
patterns, waiting times, and patient and clinician input. Improved access should be assessed
together with appropriate care, functional or symptom outcomes where relevant, and potential harms
such as delayed escalation or displacement of effective treatment.
●​ Randomized comparisons are preferable when feasible; pragmatic cluster or clinician crossover
designs may suit workflow questions. Observational comparisons require transparent handling of
baseline differences, selection, and residual confounding.
●​ Delayed review or no intervention is an appropriate comparator only when it accurately
represents usual care and is ethically justified. Patients should retain indicated treatment, rescue
procedures, and access to human support.
●​ Poor access to existing care should not lower a clinically justified safety standard. Evidence of
increased availability should remain distinct from evidence of clinical benefit.
____

Question 16
Should independent third parties be involved in some or all aspects of a competency-based
approach, including device benchmarking and clinical confirmation? If so, in what ways might
qualified, independent third-party participation contribute to this approach, and what qualifications
and independence criteria should apply? Are there aspects of the assessment for which third-party
involvement would be impractical or inadvisable? What safeguards or program design features
would be critical to prevent third-party participation from limiting competition or preventing
innovation?
Response
Independent third parties should administer and score benchmark assessments across initial
validation, periodic postmarket reassessment, and re-benchmarking after changes, with independent
clinical adjudication in clinical confirmation. The scope and intensity of evaluation may be
proportionate to risk, but protected test questions and unaffiliated scoring should be baseline
safeguards across these stages. Third parties should maintain secure test sets, validate scoring
methods, adjudicate outputs, and review clinical-confirmation findings. Required expertise should
match the task and include psychiatric clinical judgment, measurement and statistics, technical
evaluation, and relevant patient perspectives.
Raters should be qualified, trained, calibrated, and unaffiliated with the sponsor, device vendor, and
underlying foundation-model vendor. They should have no employment, governance, ownership,
consulting, or product-development ties to those parties. Disclosure of a relationship should not by
itself qualify an affiliated rater as independent. Assessment funding and relevant relationships should
be disclosed, with independent rater selection, recusal procedures, and compensation that does not
depend on favorable results. Sponsors and vendors should not control case selection, scoring,
adjudication, or reporting of the independent assessment.
●​ Independent assessment does not transfer the sponsor’s responsibility for device safety or
complete evidence. Its value should be evident in an interpretable record of the assessment,
rather than a certificate alone.
●​ Routine internal development should remain distinct from independent benchmark evidence
supporting regulatory claims. Secure evaluation arrangements and appropriate confidentiality
protections should avoid unnecessary transfer of sensitive patient records while preserving
independent access to the evidence needed for assessment.
●​ Multiple qualified assessors, transparent qualification criteria, interoperable methods, predictable
fees, and an appeals process would help support fair access to evaluation. Public rubrics can
coexist with protected test cases, and proprietary access to a single evaluator or dataset should
not become a mandatory barrier to entry.
____
Question 17
CDRH is exploring whether a competency-based approach could be applied to devices that
incorporate a variety of underlying model architectures, such as multimodal vision-language models
and generative or predictive world models. Are there aspects of the described approach that might
be ineffective or inapplicable to these devices? What additional or distinct considerations might need
to be addressed?
Response
A functional approach can be adapted across architectures, but text-based questions and output
ratings will be insufficient when clinical behavior depends on perception, temporal prediction, or
closed-loop action. Psychiatric assessment draws on vocal, nonverbal, contextual, and collateral
information, each with distinct limitations. Evaluation should establish whether the device integrates
these sources appropriately, recognizes missing or contradictory information, and avoids
unsupported inferences about mental state. Additional modalities should not be presumed to provide
clinically valid measurements without use-specific validation.
●​ Multimodal evaluation should examine each modality and their combination, including
synchronization, missing or degraded signals, and contradictions between modalities.
Performance should also be assessed across languages, disabilities, and acquisition conditions.
●​ For generative or predictive world models, evaluation should examine calibration of predicted
trajectories, accumulation of error over time, responses to interventions, and consequences of
actions chosen from those predictions. Realistic-looking simulations alone cannot establish
clinical validity.
●​ Evaluation should use reference standards appropriate to each modality, assess complete
episodes, and include real-world confirmation of clinically consequential outputs. These
proposed extensions require validation, with reassessment when sensors, encoders, or system
configurations change.
____

Question 18
CDRH is considering whether it may be appropriate to accept greater premarket uncertainty
regarding a GenAI-enabled device’s benefit-risk profile through greater reliance on postmarket
monitoring. Under what conditions might such an approach be appropriate, and what characteristics
of a monitoring program would need to be in place to justify reduced premarket evidence? Are there
device types or risk profiles for which this approach would not be appropriate?
Response
Predeployment evaluation can be burdensome, and rapid model change and the difficulty of
exhaustive testing complicate evidence generation. These challenges do not justify short-changing
the evidence needed to substantiate a device’s safety, effectiveness, and clinical claims before
deployment. Otherwise, a growing number of devices could be marketed to patients and clinicians
without adequate validation for their intended uses, populations, configurations, and claims. FDA
should require adequate independent premarket validation for the proposed uses and claims, with
uncertainties and limits made explicit. Postmarket monitoring should complement this evidence base
and must not become a substitute for it.
Reduced premarket evidence should be considered only for clearly bounded, lower-risk uses where
residual uncertainty is well characterized, effective safeguards are demonstrated, and emerging
harms can be reliably detected and addressed before becoming clinically significant. Assessment
should account for the clinical function and claim, population, autonomy, and consequences of error.
For psychiatric AI, monitoring should include:
●​ Assessment should examine how clinical reasoning, relational interaction, safety, and equity are
maintained across individual outputs, longitudinal conversations, agent actions, and clinical
workflows.
●​ Clinicians, patients, families, peers, and technical experts bring different perspectives on
benefits, burdens, and failures. Their input should inform monitoring.
●​ Human oversight requires evidence that reviewers can recognize errors, intervene, and complete
appropriate handoffs. Responsibility for corrective action should be clear.
●​ Reliable measurement depends on specified criteria and qualified, trained, calibrated, and
unaffiliated benchmark raters. Benchmark assessments should remain protected and
independent, and automated measures require independent validation. Evaluation scope and
sampling may vary with risk, while these independence safeguards remain in place.
●​ Evaluation should remain tied to the specific device version and be repeated after material
changes to models, prompts, retrieval, memory, tools, permissions, safety layers, populations, or
clinical context.
Higher-risk psychiatric functions, or uses involving serious, irreversible, or difficult-to-detect
cumulative harms, should retain stronger premarket evidence requirements. Monitoring should not
replace evidence needed to substantiate clinical-effectiveness claims.
____

Question 19
Please comment on the potential approaches to postmarket performance evaluation, including
periodic re-benchmarking, sample-based clinician review, and performance degradation monitoring.
What additional approaches should CDRH consider, and how should the cadence and triggering
events for reassessment be determined?
Response
Periodic re-benchmarking, clinician review, and degradation monitoring provide complementary
information. Re-benchmarking can identify changes in previously tested capabilities; clinician review
can examine contextual appropriateness; and degradation monitoring can detect changes
associated with the population, deployment setting, or system configuration. These approaches
should be linked to clinical reasoning, relational interaction, safety, and equity across individual
outputs, interaction trajectories, agent actions, and workflows. Periodic and triggered
re-benchmarking should be independently administered on protected held-out questions and scored
by unaffiliated raters under the safeguards in Questions 10 and 16. Repeated reassessment should
use secure reusable items or refreshed cases, with exposed items retired from independent testing.
I would add longitudinal review of patient interactions, clinician corrections and overrides, patient and
family feedback, peer perspectives, and whether escalation actually results in appropriate human
care. Sampling should represent routine use while deliberately examining rare, consequential
failures and underserved groups. Qualified, trained, calibrated, and unaffiliated evaluators should
score independent reassessment samples using prespecified criteria, with higher-consequence
cases receiving additional scrutiny. Monitoring should also assess whether alerts and reviews create
burdens that undermine effective supervision.
●​ The interval between assessments should reflect clinical acuity, the volume of exposure, and the
consequences of failure, including how much time is available to intervene.
●​ Material changes to models, prompts, retrieval, memory, tools, permissions, safety layers, patient
populations, or workflows should prompt reassessment.
●​ Safety events, unexpected subgroup differences, repeated corrections, and incomplete handoffs
should prompt investigation and a documented response, even when the software has not
changed.
____

Question 20
Please comment on whether the proposed approaches to postmarket monitoring can be facilitated
by machine-based supervisory agents. What considerations, including the evaluation and reliability
of the supervisory agent itself, should CDRH take into account for such an approach?
Response
Machine-based supervisory agents could help identify possible errors, prioritize cases for review,
and monitor changes across large volumes of interactions. Their usefulness depends on whether
they detect clinically meaningful problems in the configuration and workflow where they are
deployed. A supervisory model may share failure modes with the system it evaluates, and
agreement between two models should not be treated as independent confirmation of safety.
In behavioral health, supervisory evaluation should include indirect risk disclosures, ambiguous
language, culturally situated expressions, harmful reinforcement, and deterioration across repeated
encounters. False negatives, false positives, and failure to escalate each have clinical
consequences. Supervisory performance and any automated benchmark scoring should be
independently validated on protected held-out cases against judgments from qualified, trained,
calibrated human experts unaffiliated with the sponsor and vendors of either the device or the
supervisory or scoring model, with relevant patient or peer perspectives included. Disagreements
should be examined rather than collapsed into a single score. Automated scoring should supplement
independent human assessment, with serious failures and disputed judgments referred to
unaffiliated human adjudicators.
●​ The supervisory agent’s reliability should be assessed across system versions and realistic
interaction trajectories, with measures of sensitivity, specificity, calibration, and subgroup
performance.
●​ Testing should examine correlated failures between the device and supervisor, as well as
manipulated inputs, missing context, and failures of the supervisor or its supporting
infrastructure.
●​ Human review should have clear accountability, with auditable alerts and a safe fallback.
Evaluation should establish that flagged concerns reach those responsible and are acted upon
within clinically appropriate timeframes.
____
Question 21
What roles might clinicians, healthcare institutions, professional societies, standards-setting bodies,
and other stakeholders appropriately play in postmarket monitoring, and how can these roles be
structured without diffusing manufacturer accountability?
Response
Postmarket monitoring should draw on complementary expertise while preserving clear
manufacturer accountability. Clinicians can assess reasoning, treatment appropriateness, risk, and
whether outputs are meaningfully reviewed in practice. Healthcare institutions can examine local
workflow, staffing, handoffs, and the effects of AI on clinical judgment. Patients, families, and peers
can identify misunderstanding, burden, inappropriate relational responses, and adverse experiences
that may be missed in technical or clinician-only review. Participation should respect consent,
confidentiality, and the patient’s preferences about family involvement.
Manufacturers, developers, deploying institutions, clinicians, and patients should share an accurate
understanding of the device’s role, limitations, and their respective responsibilities, with families and
peers included when appropriate to the patient’s preferences and care context. Training, user
interfaces, marketing, and procurement decisions should consistently reflect that understanding.
Stakeholder agreement should not be treated as permission to expand a device’s role beyond its
validated scope.
Professional societies and standards-setting bodies can help develop clinically meaningful criteria,
shared reporting terms, evaluator training, and methods for comparing findings across settings.
Technical experts can investigate system behavior and version-specific failures. Independent
benchmark reviewers should meet the unaffiliated-rater criteria in Question 16. Greater
consequences should increase the depth of review; qualifications, training, calibration, and relevant
relationships should be transparent.
●​ The manufacturer should remain responsible for the monitoring plan, investigation, corrective
action, change control, and communication of device limitations.
●​ Institutions and clinical teams should have clear responsibilities for local oversight and timely
reporting, supported by usable reporting channels and feedback.
●​ Agreements should specify who receives each signal, who is responsible for acting, and how
resolution is verified. These responsibilities should not depend on informal goodwill or unpaid
patient vigilance.
____

Question 22
CDRH envisions that a manufacturer’s premarket competency-based assessment might serve as a
baseline against which post-deployment modifications could be re-evaluated. How might the extent
of re-benchmarking or other evidence be scaled to the nature and expected impact of a given
modification? For example, are there categories of change that might not significantly affect safety or
effectiveness of a GenAI-enabled device, or might be appropriately managed within a sponsor’s
quality management system versus requiring FDA premarket review and authorization before
implementation? Are there categories of changes that might be appropriate for inclusion in a PCCP?
Response
Re-benchmarking should be scaled to the modification’s expected clinical effects, including effects
that cross functional domains or system levels. The starting point should be a version-specific
account of the intended use, patient population, configuration, and previously demonstrated
performance. Rapid model change strengthens the need for version-specific reassessment; it should
not justify waiving adequate baseline validation or deferring evidence needed for an affected use or
claim until after a modified device is deployed. A small prompt or interface change can alter
disclosure, reliance, escalation, or how a clinician interprets a recommendation; the size of the
software change is therefore a poor proxy for clinical significance.
Testing should examine the functions directly affected and plausible downstream effects on
reasoning, relational interaction, safety, equity, and the human-AI workflow. Benchmark regression
testing may suffice for some tightly bounded changes, whereas changes affecting clinical claims,
autonomy, or supervision may require prospective workflow or clinical evidence. In each case,
re-benchmarking used to substantiate continued performance should retain independent
administration, protected held-out questions, and unaffiliated raters. Scale the coverage and
intensity of testing to the change, rather than relaxing independence; sponsor-run regression tests
may support internal quality control but should not replace independent reassessment.
●​ Changes shown to be clinically immaterial could be managed within the quality system with
documented verification. An example would be a cosmetic adjustment that does not alter
interpretation or use.
●​ A PCCP could potentially cover bounded updates to retrieval resources within an unchanged
intended use, provided it specifies validation methods, acceptance criteria, subgroup testing, and
rollback procedures.
●​ When a change extends the intended use, populations, consequential actions, or risk beyond the
authorized scope, the manufacturer should assess whether FDA review is needed before
implementation. The prior benchmark alone is insufficient.
____

Question 23
How might PCCP concepts or other change-control approaches be adapted for GenAI-enabled
devices when the nature or scope of future modifications cannot be fully prespecified?
Response
When future modifications cannot be fully specified, change control should define the permitted
scope of adaptation and the evidence required before implementation. In psychiatry, personalization
may change responses as a patient’s symptoms, preferences, and context evolve. Evaluation should
distinguish this intended adaptation from changes that alter the system’s clinical role, permissions, or
risk. Prespecifying a validation process is useful, but it should not provide open-ended permission for
uncharacterized clinical behavior.
I would support change-control approaches that identify stable clinical requirements and explicit
boundaries, while allowing justified methods for evaluating modifications within those boundaries.
The requirements should cover clinically meaningful performance across the four functional domains
and the system levels affected. Evidence should remain tied to the deployed configuration, including
memory and accumulated interaction context.
●​ A change-control plan should make clear which categories of change are permitted, which uses
are excluded, what safety constraints apply, and when additional clinical evidence or regulatory
review is required.
●​ Assessment of each change should connect its expected effects to version-specific testing and
subgroup analysis. Benchmark evidence used to substantiate continued performance requires
 independent administration and unaffiliated scoring. The extent of additional review should reflect
the consequences of error.
●​ Implementation should be controlled, decisions traceable, and rollback or safe suspension
feasible. A change that exceeds the validated scope should be paused for appropriate review
rather than accommodated by stretching the original plan. Model updates, personalization, and
workflow changes should preserve the device’s declared role unless an explicit expansion is
evaluated and authorized through the applicable change-control process. Any proposed
expansion of that role should undergo appropriate independent validation and regulatory review,
when applicable, before deployment.
____

Question 24
For devices built on third-party foundation models, changes to the underlying model may be initiated
by the third-party model developer rather than the device manufacturer. How can a manufacturer
detect, evaluate, and respond to such changes in a timely manner? What mechanisms—for
example, contractual, technical, or through a PCCP—could provide reasonable assurance that
third-party developer-initiated changes do not compromise the safety or effectiveness of the device?
Response
The device manufacturer should remain responsible for the clinical performance of the deployed
system when the underlying foundation model is supplied by another organization. Evaluation should
cover the full configuration, because changes in model behavior may interact with prompts, retrieval,
memory, tools, permissions, and safety layers. An unchanged device interface does not establish
unchanged clinical performance.
I would favor contractual and technical arrangements that make changes visible before clinical
exposure whenever possible. These should include version identification, advance notification,
access to relevant change information, and an opportunity to test the new configuration. Where a
supplier cannot support these arrangements, the manufacturer should account for that uncertainty
when determining whether the model is suitable for the intended clinical use.
●​ Version pinning should be used where available, together with configuration records, audit logs,
and checks of system behavior to identify unexpected changes.
●​ Reassessment should cover affected functions and plausible downstream effects, including
subgroup performance, longitudinal interactions, and how escalation works in clinical care.
●​ Response procedures should address unexpected changes through restrictions on affected
functions, a return to a supported validated version, or safe suspension of use. These responses
should preserve continuity of care.
●​ A PCCP may support changes within a defined scope, but supplier updates outside that scope
require separate assessment.
____
Question 25
Would voluntary Foundation Model MAFs be practical for foundation model developers to provide
and to keep current, and what content might they include? Given that participation would be
voluntary and that model developers may have limited incentive to disclose safety-relevant
information, what would make such a program sufficiently useful for premarket review? Are there
alternative mechanisms CDRH should consider for obtaining information about underlying
foundation models?
Response
Voluntary Foundation Model MAFs could be useful if they provide structured, version-specific
information that FDA reviewers can rely on and sponsors can reference with the file holder’s
authorization. As described in the discussion paper, these confidential files would support device
review without authorizing the foundation model for a clinical intended use. Sponsors would still
need to demonstrate the safety and effectiveness of their own configured devices.
Practical value would depend on a manageable reporting template, clear update expectations, and
information that goes beyond promotional model summaries. Reusable confidential submissions
could reduce duplicative requests and provide an incentive to participate. However, voluntary
participation and incomplete disclosure would limit coverage; absence of reported problems should
not be interpreted as evidence that relevant risks were evaluated.
●​ Reviewers need to be able to identify the model and version and understand its relevant data
provenance, evaluation methods, subgroup findings, known limitations, safety constraints, and
behavior affected by updates.
●​ The file should describe commitments to notify manufacturers of changes, the availability of
model versions, audit capabilities, and limitations in the evidence relevant to healthcare.
●​ Standardized supplier documentation, authorized access to technical evidence, and qualified
independent assessments could provide complementary information. Missing information should
be made explicit, with device-specific evidence required to address consequential gaps.
____
Question 26
Are there additional considerations that inform the premarket and postmarket evaluation of agentic
GenAI-enabled devices, beyond those applicable to non-agentic GenAI-enabled devices? How
should the elevated risk associated with autonomous multi-step action, tool use, and reduced
opportunity for human review be reflected in acceptance criteria and oversight?
Response
Agentic devices require evaluation of actions and consequences across the complete sequence of
care. A reasonable response can lead to inappropriate action through incorrect retrieval, tool misuse,
excess permissions, or incomplete handoff. Evaluation should identify the deployed configuration:
model, prompts, memory, retrieval sources, tools, permissions, safety layers, and workflow.
In psychiatric care, testing should examine whether an agent recognizes changing risk, revises
outdated assumptions, obtains necessary context, and preserves patient agency. It should assess
cumulative errors, partial completion, repeated actions, and recovery when a tool or human recipient
is unavailable. Testing should also include prompt injection through retrieved content or tool outputs,
and erroneous tool outputs. Human oversight should be tested under realistic staffing and time
constraints.
●​ Acceptance criteria and the evidence required should reflect the device’s clinical role, how
independently it acts, and the consequences of an error. Evaluation should also establish
whether an action can be reversed and whether there is a realistic opportunity to recognize and
correct a problem before harm occurs.
●​ The functions a device may carry out independently, those requiring human review or
authorization, and those it should not perform should be clearly specified. When information is
insufficient or a component fails, the device should stop safely and seek appropriate support.
These boundaries should be implemented through technical permissions, restrictions on
available tools and actions, and defined human authorization and handoff requirements. Testing
should assess whether the safeguards prevent or detect foreseeable misuse and progressive
role drift across repeated interactions, including when users or connected systems request
actions outside the evaluated role.
●​ Evaluation should assess whether each consequential step is clinically appropriate and
completed successfully, together with clinical outcomes and the performance of the complete
care workflow.
●​ Postmarket monitoring should review action logs linked to device versions, human overrides,
failed handoffs, subgroup differences, and effects across repeated use. Changes to tools,
permissions, memory, or workflow should prompt reassessment.