American Academy of Pediatrics (Andrew D. Racine)

See attached file(s)

AAP Headquarters October 19, 2026
345 Park Blvd
Itasca, IL 60143
Phone: 630/626-6000
Fax: 847/434-8000
Michelle Tarver, MD, PhD
E-mail: kidsdocs@aap.org Director
www.aap.org
Center for Devices and Radiological Health
Reply to Food and Drug Administration
AAP Washington Office
601 13th St NW, Suite 400N 10903 New Hampshire Ave
Washington, DC 20005 Silver Spring, MD 20993
Phone: 202/347-8600
E-mail: kids1st@aap.org

Executive Committee RE: Docket No. FDA-2026-N-7874
President
Andrew D. Racine, MD, PhD, FAAP Dear Dr Tarver:
President-Elect
Terri D. McFadden, MD, MPH, FAAP
On behalf of the American Academy of Pediatrics (AAP), a non-profit professional organization
Immediate Past President
Susan J. Kressly, MD, FAAP
of more than 67,000 primary care pediatricians, pediatric medical subspecialists, and pediatric
surgical specialists dedicated to the health, safety, and well-being of all infants, children,
Secretary/Treasurer
Joelle N. Simpson, MD, FAAP
adolescents, and young adults, I am writing to share feedback from the AAP on the Food and
Drug Administration (FDA) discussion paper and request for feedback on the regulation of
CEO/Executive Vice President
Mark Del Monte, JD generative artificial intelligence (GenAI)-enabled medical devices.
Board of Directors
The AAP appreciates FDA’s recognition that GenAI-enabled medical devices may present
District I
Patricia Flanagan, MD, FAAP different risks across populations and use cases. Although these technologies have significant
District II potential to support pediatric care, evidence regarding their performance in real-world
Jeffrey Kaczorowski, MD, FAAP pediatric settings remains limited. Pediatric considerations should therefore be incorporated
District III consistently across the regulatory framework rather than addressed only through individual
Lenore R. Jarvis, MD, MEd, FAAP
references to subgroup performance. Children are not simply small adults; age-specific
District IV physiology, developmental stage, body size, communication ability, dependence on caregivers,
Patricia Purcell, MD, MBA, FAAP
and access to pediatric expertise can affect how these devices perform and how their outputs
District V
Christopher B. Peltier, MD, FAAP are understood and acted upon. FDA should therefore treat population- and developmentalDistrict VI stage-specific considerations as risk modifiers across risk assessment, premarket evaluation,
Claudia Preuschoff, MD, FAAP and postmarket monitoring. Where pediatric evidence is limited and performance is
District VII extrapolated from adult populations, those limitations should be transparent to clinicians,
Susan Buttross, MD, FAAP
patients, and families. i
District VIII
Greg Blaschke, MD, MPH, FAAP
The AAP also encourages FDA to consider how several issues raised separately throughout the
District IX
Eric H. Ball, MD, FAAP discussion paper intersect across the device lifecycle. Assumptions about the effectiveness of
District X human oversight should be supported by evidence that users can identify and correct clinically
Madeline M. Joseph, MD, FAAP significant errors. Similarly, risks that emerge over extended, multi-turn interactions should be
At Large addressed consistently in both risk assessment and performance evaluation, and expectations
Angela M. Ellison, MD, MSc, FAAP for the use of synthetic data should apply consistently wherever those data are used to
At Large supplement real-world evidence. Across these areas, FDA should expect transparency
Kristina W. Rosbe, MD, FAAP regarding the populations and circumstances in which a device has been evaluated, including
At Large relevant age ranges and limitations, and should consider how differences in data quality,
Joelle N. Simpson, MD, FAAP language access, digital access, and availability of specialty care may contribute to disparities.
The AAP offers the following feedback in response to several of the questions outlined by FDA.

IV. Considerations for the Assessment of Risk for GenAI-Enabled Devices

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?

The AAP recommends that patient-facing messages, and the framing of risk around them, be assessed based
on their function, degree of personalization, foreseeable reliance by the user, and the consequences of an
erroneous output, rather than wording alone. Patient-facing education, explanation of an existing care plan,
and individualized treatment recommendations may present different levels of risk. The proposed wordingbased tiers, for both health care professional- and patient-facing content, are concerning because they may be
easy to game while providing limited real-world safety margin. If a clinician, patient, adolescent, or caregiver
asks a tool how to treat a condition, they are likely primed to rely on the recommendation based on the nature
and common use of the tool itself, not on whether the output says “you should” versus “most would.” Systems
should clearly distinguish educational information from individualized clinical recommendations and should
communicate when information is uncertain or outside the device’s intended scope.

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?

Patient- and caregiver-facing functions may warrant greater scrutiny when outputs could influence urgent
triage, medication use, symptom management, or decisions to seek care. The ability of a user to independently
evaluate an output cannot be assumed, particularly for children, adolescents, and caregivers. Where age
materially affects a user’s ability to evaluate or act on an output, sponsors should consider appropriate ageassurance mechanisms and age-responsive safeguards. Safeguards should account for whether the user is a
patient or parent/caregiver, particularly when information entered by a caregiver informs the device’s output,
and should include appropriate stopping points that direct users to a pediatrician or other qualified pediatric
expert when the available information is insufficient to support a reliable recommendation. Risk assessments
should also account for differences across neonatal, childhood, and adolescent populations. These distinctions
are particularly relevant to symptom interpretation, communication, dosing, and recognition of clinical
deterioration.

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?

The distinction between generalist and specialist knowledge is particularly relevant in pediatrics, where
clinical considerations can vary substantially by age and developmental stage. Clinicians working outside their
pediatric area of expertise may not be well positioned to independently identify an erroneous or inappropriate
output. Risk assessment and safeguards should therefore account for both the clinical specialty involved and
the intended pediatric population.

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?

Multi-turn conversations present risks that should be addressed consistently across risk assessment and
premarket evaluation. Although the discussion paper addresses multi-turn interactions in the risk-assessment
framework and long-conversation degradation as a performance-testing consideration, these reflect the same
underlying concern: safety behaviors and guardrails may degrade over the course of an extended exchange.
The AAP recommends that sponsors validate safety behaviors at depth, declare the maximum conversation
length tested, and consider appropriate length limits. These expectations should be applied consistently across
both risk assessment and performance evaluation. There may be no practical way to scale long-conversation
testing without developing LLM-as-judge frameworks, which remain immature but may be critical for future
evaluation because human review of long conversations is difficult to conduct meaningfully at scale.

We note that some of the highest-profile safety incidents and negative mental health outcomes for teenagers
have occurred after extended conversations, and that our current testing infrastructure tends toward much
shorter single- or few-turn conversations. ii Multi-turn systems should specifically be tested using realistic
conversations involving caregiver-entered histories, adolescent-entered information, incomplete histories,
and shifts from general information to action-directing advice. Testing should assess how the system handles
missing age or weight information, conflicting child and caregiver accounts, and conversations that become
more clinically consequential over time.

V. A Competency-Based Approach for Premarket Evaluation of GenAI-Enabled Devices

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?

The AAP is supportive of the benchmarking structure described in the discussion paper, but there are several
considerations specific to pediatric populations with regard to subgroup performance. Benchmarking should
include specialty-relevant tasks such as weight-based dosing, age-specific reference ranges, growth and
developmental considerations, and recognition of emergencies specific to the relevant population. Evaluation
datasets should include common conditions as well as rare diseases, congenital conditions, medically complex
patients, and presentations that differ from the general population. Comparators and expert adjudicators
should include relevant subspecialists when the intended use involves specialty care. Subgroup performance
should be prespecified and reported where clinically relevant, including age, language, disability,
demographic factors, and other characteristics that may affect performance or access.

Premarket evidence should include meaningful representation across relevant populations, developmental
stages, and clinical complexity, and performance in one population should not be extrapolated to another
without justification and supporting evidence. Evaluation should focus on clinically meaningful differences
rather than requiring separately powered studies for every age group or subgroup. FDA should support
scientifically justified evidence borrowing accompanied, where appropriate, by focused pediatric
confirmation. FDA should consider what evidence is being applied and whether the evidence supports
applying performance derived from one population to a distinct function or population, including the clinical
task, developmental physiology, disease presentation, and how relevant users interact with the device. Where
extrapolation is scientifically justified, sponsors should identify what confirmation remains necessary; where it
is not justified, sponsors should identify what evidence would resolve the uncertainty. Any remaining
uncertainty should be transparently reported. Overall performance estimates may not reveal clinically
important differences among subgroups or complex cases.

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?

Synthetic data may play a role across both premarket and postmarket evaluation, particularly by
supplementing clinical confirmation for hard-to-sample subgroups. However, concerns about whether
synthetic data accurately represent the intended population, including the risk that they reproduce omissions
or assumptions from the model class used to generate them, apply in both contexts. Therefore, the AAP
recommends that FDA establish consistent expectations for synthetic data provenance and representativeness
across premarket evaluation and postmarket re-benchmarking.

When synthetic data supplement real patient data, sponsors should report performance on real and synthetic
cases separately and assess whether meaningful differences exist between those distributions before
combining results. Repeated responses to the same case may help characterize output variability but should
not be treated as independent observations when estimating clinical performance. Benchmarking and clinical
confirmation evidence should only be combined when differences in the populations, case distributions, and
conditions of evaluation do not undermine the interpretability of the resulting performance estimate.

13. For which clinical domains, device functions, or subpopulations is synthetic data particularly well-suited, or
particularly inadequate, as a supplement to realworld 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?

Synthetic data may help address limited sample sizes in smaller populations, but it should supplement rather
than replace appropriate real-world evidence when safety-critical decisions are involved. Synthetic cases can
help explore uncommon scenarios, but their role in estimating clinical performance raises separate questions:
how closely they represent the intended population, and whether their generation may reproduce the same
omissions or assumptions present in the device being tested. Distinguishing results from real cases and
synthetic cases would make the evidentiary contribution of each clear. Where a device lacks sufficient pediatric
data to support a reliable output for the relevant population or clinical circumstance, it should recognize that
limitation and direct the user to an appropriate clinician or other source of pediatric expertise, providing any
needed information to facilitate a connection to care, rather than provide an unsupported recommendation.

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?

Performance measures should extend beyond whether an answer is factually correct. Evaluation should assess
whether the device recognizes when information is insufficient, asks an appropriate clarifying question,
identifies a time-sensitive concern, and communicates an actionable next step. Testing should include
examples such as weight-based dosing and unit errors, age-dependent interpretation of symptoms, and
adolescent presentations involving eating disorders, substance use, or self-harm. Under-escalation and overescalation have different consequences and should be reported separately so those trade-offs are visible rather
than combined into a single accuracy measure.

If sponsors rely on human oversight as a basis for reduced evidentiary burden, they should demonstrate that
humans in the loop can actually identify and correct errors. Available data on automation bias are concerning
and suggest clinicians may miss systematically wrong model output. iii The discussion paper appropriately
questions whether a disclaimer makes patient-facing information less directive; similarly, the presence of
clinician oversight should not itself be assumed to reduce risk. Human-AI team performance should therefore
be substantiated rather than assumed, particularly if it is treated as a safeguard or comparator.

VI. Postmarket Monitoring for GenAI-Enabled Devices

19. Please comment on the potential approaches to postmarket performance evaluation, including periodic rebenchmarking, 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?

Postmarket monitoring should assess performance in clinically relevant subgroups separately rather than
allowing aggregate data to mask safety or effectiveness concerns specific to those subgroups. A central
question is whether a proposed monitoring program can detect a clinically important problem early enough to
prevent further harm; low utilization in a particular subgroup may make aggregate monitoring less
informative. For low-volume pediatric uses, individual institutions or manufacturers may not accumulate
enough cases to reliably assess performance in every subgroup at frequent intervals. Monitoring approaches
could therefore include pooled multisite surveillance, targeted review of safety-critical events, and
reassessment triggered by material changes or emerging safety or performance signals. Monitoring intensity
should reflect clinical risk and exposure, with clear acknowledgment of what small sample sizes can and
cannot establish. Monitoring should be capable of identifying subgroup-specific drift, changes in escalation
behavior, dosing errors, communication failures, and performance degradation after model or prompt
updates. Relevant details include subgroup-specific exposure, version tracking, reporting pathways, and the
time between detecting a problem and taking corrective action.

Multiple investigators have seen model drift, or changes in model behavior even over brief periods of time
(weeks to months). As such, more frequent interval surveillance may be important to ensure continued high
quality outputs. iv,v

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?

Manufacturer accountability should remain clear even when hospitals, clinicians, professional societies, or
other stakeholders contribute to monitoring and reporting. Predetermined change control plans should define
when re-benchmarking or additional clinical confirmation is needed after changes to models, guardrails,
retrieval systems, or user interfaces. Evaluation should distinguish changes to the underlying model from
changes to prompts, retrieval sources, guardrails, or connected tools, since each can alter the final device’s
behavior.

VII. Other Topics
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?

For devices built on third-party foundation models, sponsors should understand and disclose available
information about the extent of relevant representation in training and evaluation data, known limitations,
and processes for notification of model changes. Where detailed information about the foundation model or
its training data is not available to the sponsor, those limitations should be disclosed and relevant uncertainty
should be addressed through independent evaluation of the final device in its intended population. FDA
should also consider what information and update notifications are available to the device manufacturer, and
what information is available to clinicians and institutions about the version in use and its limitations.
Foundation Model Device Master Files could be useful if they include relevant performance data, limitations,
update history, safety evaluations, and sufficient information to assess whether the underlying model is
appropriate for the intended use.

26. Are there additional considerations that inform the premarket and postmarket evaluation of agentic GenAI-enabled
devices, beyond those applicable to nonagentic 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?

Agentic systems that can take clinically significant actions should have strong, use-case-specific constraints,
auditable logs, fail-safe behavior, and human confirmation where warranted by the severity and reversibility
of potential harm, demonstrated system performance, and the system’s operating limits. Stronger controls
should apply to higher-risk actions, while appropriately validated automation may be suitable within bounded
workflows. FDA should consider who authorizes clinically consequential actions, how errors may propagate
through a sequence of steps, and whether an action can be interrupted or reversed.

Appendix A, E.4: Communication quality and user comprehension

Outputs intended for pediatric patients or caregivers should be evaluated for comprehension, health literacy,
uncertainty communication, and appropriate escalation, with particular attention to the evolving capacities of
youth across the pediatric age span. Adolescent use requires special attention to confidentiality, autonomy,
assent, caregiver involvement, and situations in which safety concerns may require escalation. Testing should
include family-centered workflows in which information is entered, interpreted, or acted upon by more than
one person.

Conclusion

The AAP appreciates FDA’s attention to the emerging regulatory questions raised by GenAI-enabled medical
devices. As FDA continues to refine its approach, pediatric-specific evaluation, transparency, and lifecycle
monitoring should be built into expectations from the outset so that innovation proceeds in a manner that is
safe, equitable, and clinically meaningful for pediatric populations, families, and the clinicians who care for
them.
The AAP appreciates the opportunity to comment on this discussion paper. If we may be of further assistance
in lending our expertise, please do not hesitate to contact the AAP Washington Office at 202-347-8600 or
kids1st@aap.org.

Sincerely,

Andrew D. Racine, MD, PhD, FAAP
President

ADR/ncp

i
Suresh S, Hunter RB, Periard B, et al; American Academy of Pediatrics, Council on Clinical Information
Technology; and Section on Innovation in Therapeutics and Technology. Recommendations for the Development and
Implementation of Generative Artificial Intelligence Tools in Pediatric Clinical Care: Policy Statement. Pediatrics. 2026;
158(5):e2026079037
ii
Tiku N, Schaul K. A teen’s final weeks with ChatGPT illustrate the AI suicide crisis. Washington Post. December 27, 2025.
iii
Jabbour S, Fouhey D, Shepard S, et al. Measuring the Impact of AI in the Diagnosis of Hospitalized Patients: A
Randomized Clinical Vignette Survey Study. JAMA. 2023;330(23):2275–2284. doi:10.1001/jama.2023.22295
iv
Gupta M, Virostko J, Kaufmann C. Large language models in radiology: fluctuating performance and decreasing
discordance over time. Eur J Radiol. 2025;182:111842. doi:10.1016/j.ejrad.2024.111842.
v
Chen L, Zaharia M, Zou J. How is ChatGPT’s behavior changing over time? Harv Data Sci Rev. 2024;6(2).
doi:10.1162/99608f92.5317da47.