American Academy of Family Physicians (Jen Brull)

On behalf of the American Academy of Family Physicians (AAFP), representing 124,500 family physicians and medical students nationwide, we appreciate the opportunity to provide feedback on FDA’s discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices. The AAFP appreciates FDA’s proactive effort to explore regulatory approaches that support innovation while ensuring that GenAI-enabled medical devices are safe, effective, and trustworthy. As GenAI technologies become increasingly integrated into clinical workflows, patient interactions, and health system operations, it is critical that regulatory frameworks remain focused on promoting patient safety, supporting evidence-based care, and preserving the patient-physician relationship. We encourage FDA to continue engaging stakeholders as these policies evolve, and we stand ready to serve as a resource and partner in developing evidence-based frameworks that support the responsible deployment of GenAI-enabled medical devices.

October 15, 2026

The Honorable Kyle A. Diamantas, J.D.
Acting Commissioner
U.S. Food and Drug Administration
10903 New Hampshire Avenue
Silver Spring, MD 20993

RE: FDA Considerations for the Regulation of Generative AI-Enabled Medical
Devices Request for Feedback; Docket ID: FDA-2026-N-7874

Dear Commissioner Diamantas:

On behalf of the American Academy of Family Physicians (AAFP), representing
124,500 family physicians and medical students nationwide, we appreciate the
opportunity to provide feedback on the Food and Drug Administration’s (FDA)
discussion paper, Considerations for the Regulation of Generative AI-Enabled
Medical Devices. The AAFP appreciates FDA’s proactive effort to explore regulatory
approaches that support innovation while ensuring that Generative AI (GenAI)-
enabled medical devices are safe, effective, equitable, and trustworthy. As GenAI
technologies become increasingly integrated into clinical workflows, patient
interactions, and health system operations, it is critical that regulatory frameworks
remain focused on promoting patient safety, supporting evidence-based care,
and preserving the patient-physician relationship.

Family physicians are often the first point of contact within the health care system
and provide comprehensive, continuous, and coordinated care across diverse patient
populations and care settings. As a result, family medicine offers a unique perspective
on both the opportunities and risks associated with GenAI-enabled technologies.
Consistent with the AAFP’s longstanding principles regarding artificial intelligence (AI),
we believe these tools should be evaluated with the same rigor as other tools used in
health care and should be designed to promote trustworthiness, privacy protection,
and equity. We further believe that practicing physicians should be involved
throughout the AI lifecycle, including development, validation, implementation, and
postmarket evaluation to ensure that these technologies meaningfully improve
patient care and clinical workflows.

The AAFP believes that clinical AI should augment, not replace, physician
judgment and should preserve physician responsibility for clinical decisionmaking. For functions that initiate or materially influence clinical actions, the AAFP
supports meaningful oversight and accountability by a licensed physician. The AAFP
offers responses to questions most relevant to family medicine and primary care.

info@aafp.org
1133 Connecticut Ave., NW, Ste. 1100
(800) 794-7481
Washington, DC 20036-1011
(202) 232-9033 www.aafp.org
 October 15, 2026
Page 2 of 22

Section IV: Considerations for the Assessment of Risk for GenAI-Enabled Devices

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.

The AAFP agrees that the FDA-identified additional dimensions of reversibility,
downstream safeguards, time pressure, and traceability are appropriate risk modifiers
to include.

The work of the Office of the National Coordinator for Health Information
Technology (ONC) on the FDA Safety and Innovation Act in 2014 provided a list of 11
facets to consider for determining risk of health IT, i and we believe these are relevant
for gauging AI risk. The facets identified are:
• purpose of the software product
• intended user(s)
• severity of potential injury
• likelihood of hazardous situation arising
• transparency of software operations and data
• ability to mitigate harmful conditions
• complexity of software (and maintenance)
• complexity of implementation and upgrades
• complexity of training and use
• use as part of a more comprehensive solution
• network connectivity, standards, and security

The AAFP recommends that the availability of meaningful physician oversight, patient
health literacy, explainability and transparency, and the care setting also be included
as risk modifiers. Meaningful physician oversight may reduce risk only when the
physician has timely access to the information needed to evaluate an output;
sufficient time and authority to independently review it; and the practical ability to
accept, modify, reject, interrupt, or reverse a resulting clinical action before harm
occurs, when clinically feasible. Oversight should not be treated as an effective
safeguard when workflow design, time pressure, automation bias, inadequate
transparency, or system permissions make physician review or intervention merely
nominal.
 October 15, 2026
Page 3 of 22

Additionally, the AAFP believes organizational readiness and implementation capacity
should also be considered when assessing risk. A safeguard that may be effective in a
large, well-resourced health system may not function equivalently in an independent,
rural, safety-net, or resource-constrained practice without adequate training,
technical assistance, governance support, and implementation resources. FDA should
evaluate whether a device’s risk controls can be implemented effectively across the
settings in which the manufacturer intends it to be used.

FDA could apply these modifiers after a device is initially placed within the two-axis
framework, increasing the assigned risk level where traceability is limited, use is timesensitive, the action is irreversible, or where meaningful physician review is absent.
Effective safeguards may mitigate but should never be considered to eliminate the
underlying risk of a high-consequence function.

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?

For traditional Software as a Medical Device (SaMD), the output space is enumerable
(e.g., a risk score has a range, an alert is or is not sent), and the manufacturer can
characterize every output the device can produce. A generative model cannot be
characterized this way. It produces outputs at runtime that vary with phrasing,
conversational history, patient data, and stochastic sampling. A framework that asks
manufacturers to classify the output is asking for something that does not exist as a
fixed device attribute.

What manufacturers should be able to characterize, and what FDA can assess, are:
• the ceiling of directiveness the device is designed to reach;
• the controls that hold outputs at or below that ceiling (system prompting,
output filtering, refusal behavior, template constraints, retrieval grounding);
and
• the measured distribution of realized directiveness across representative
queries and users.
 October 15, 2026
Page 4 of 22

FDA has already stated that directiveness "may depend on the substance and context
of the output, not solely on whether it uses words such as 'recommend,' 'should,' or
'consider.'" The construct that actually predicts harm is not linguistic form but the
probability that a user acts on the output without independent verification, which can
be called effective directiveness.

For example, the same sentence "in situations similar to this, clinicians often increase
the lisinopril dosage" can be interpreted as:
• a prompt for further thought to a hypertension specialist with time to think;
• a near-instruction to a family physician at minute 12 of a 15-minute visit with
three other problems on the agenda and 40 messages in the inbox; or
• an instruction to a patient with no clinical training.

Risk may increase as outputs become more specific, urgent, tailored to an individual
patient, or framed in language that implicitly encourages a particular action. The
wording and the degree of confidence expressed also influence user behavior and
should be weighed when assessing risk. Developers should clearly communicate
intended use, limitations, performance characteristics, and validated uncertainty
measures so physicians and patients can appropriately interpret outputs.

To improve regulatory predictability, the AAFP recommends FDA provide examples
showing how these characteristics move an output along the continuum from
general information to individualized instruction. FDA should also clarify that
disclaimers alone do not convert an otherwise action-directing output into a nondirective function.

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?

The AAFP strongly supports efforts to expand patient access to health information
and tools that facilitate patient engagement. Patient-facing GenAI-enabled functions
may present different or elevated risks when their safe use depends on specialized
clinical knowledge, when outputs are personalized or action-directing, or when an
incorrect output could delay appropriate care. Risk assessment should focus on the
design and consequences of the function rather than presume that patients are
inherently unable to understand or act on health information.
 October 15, 2026
Page 5 of 22

We believe that “patient-facing” is too high-level a term to adequately represent the
wide spectrum of scenarios that will have varying levels of risk. One critical factor is
the degree to which the patient has a usual source of care and how that source of
care has visibility and engagement with the “patient-facing” AI. For example, here are
three scenarios that have low, moderate, and high risk, respectively:
Risk
Configuration Recourse available
profile
GenAI in a patient portal. Interaction visible
Error is likely caught;
to the patient's family physician with an
patient has somewhere Low
established continuity relationship;
to take the output.
appointment available.
GenAI in a health system application. No Error caught only if the
clinician sees the exchange, but the patient patient independently Moderate
has an assigned physician. raises it.
Standalone consumer GenAI. No clinician
No correction pathway
will ever know the conversation occurred; High
exists.
patient may have no usual source of care.

These applications differ by an order of magnitude in risk and are indistinguishable
under a patient/health care provider (HCP) binary. Meanwhile, an HCP-facing output
delivered to an AI company physician or locum covering an unfamiliar panel with no
access to the patient's history will have less recourse than the first row above.

We believe manufacturers must include safeguards to help ensure the lowest risk to
patients. In addition to integration with a patient’s usual source of care, appropriate
safeguards may include plain-language communication, disclosure of uncertainty,
escalation pathways when high-risk symptoms are identified, and encouragement to
engage with a clinician when appropriate. Patient-facing functions should also be
designed and evaluated with varying levels of health and digital literacy in mind, with
information presented at an appropriate reading level and available in the patient's
preferred language when feasible. Patients should receive clear information about a
device's intended use; limitations; and how information they provide will be
collected, used, stored, or shared. Where additional assistance is needed, educational
resources or other support should be available to help patients understand and use
the technology with confidence. Such safeguards should empower patients without
creating unnecessary barriers to access or reinforcing overprotective approaches to
care. Patient-facing tools should clearly disclose when information is AIgenerated and distinguish general informational content from individualized
medical advice.
 October 15, 2026
Page 6 of 22

4. CDRH seeks input on whether and how the distinction between generalist and
specialist physicians might inform the assessment of risk for HCP-facing GenAIenabled 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?

Family physicians are trained to provide comprehensive, continuous care across the
lifespan, and they routinely manage conditions that fall within multiple specialty
domains. The AAFP believes GenAI-enabled tools have significant potential to extend
specialty expertise into primary care settings, particularly in rural and underserved
communities where specialty access is often limited or unavailable.

The AAFP is concerned that framing this issue as a distinction between generalist and
specialist physicians may not accurately reflect the factors that influence risk. The
relevant risk variable is not a deficit of knowledge on the part of generalist physicians,
but a mismatch between the competencies a device assumes of its user and the
competencies its user actually possesses — a mismatch that runs in both directions.
A device calibrated for subspecialist use that presents as authoritative to a generalist
is one failure mode; a device that oversimplifies for a physician who has substantial
expertise in an area is another, and it produces the disengagement and alert fatigue
that undermine safe use. A subspecialist receiving output outside their own field
occupies precisely the position FDA describes, frequently with less longitudinal
knowledge of the patient than the family physician with an established relationship
would have.

Risk increases where a user has limited practical means of verifying an output. That
condition correlates with access rather than with training: a family physician
practicing in a county with no dermatologist within a hundred miles has materially
less recourse than a clinician who can obtain an informal consultation the same day.
We therefore recommend that FDA require developers to define intended users
by demonstrated competency and realistic conditions of use, rather than by
specialty designation alone, and to validate device performance against those
users' actual capabilities, practice settings, and available consultation resources.
Validation conducted in academic medical centers and large integrated systems does
not establish performance in the ambulatory primary care settings where much of
this technology will be used. The AAFP raised this point in our December 2025
response to FDA's request for information on real-world device performance, and it
applies with equal force here.

We reject the false distinction that “generalist versus specialist” sets, as the
boundaries between generalist and specialist practice are contested and heavily
 October 15, 2026
Page 7 of 22

overlapping. Management of diabetes, heart failure, depression, and common
dermatologic and musculoskeletal conditions are simultaneously core family
medicine practice areas and recognized specialty practice. We urge FDA to state
explicitly that device labeling regarding intended users is not a determination of
scope of practice and to avoid constructing a taxonomy of specialty-restricted
clinical content. Defining intended users by competency and conditions of use rather
than by specialty designation avoids this problem while still addressing the risk FDA
has identified.

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?

The AAFP agrees that risk assessment for multi-turn conversational GenAI-enabled
devices should consider the full conversational trajectory rather than individual
outputs in isolation. A device that initially provides general information may become
increasingly personalized, specific, or action-directing as additional context is
provided. Risk may therefore arise from the cumulative effect of a conversation, even
when no single response appears high-risk, and we recommend FDA evaluate
whether device behavior remains consistent across longer interactions and whether
clinically significant recommendations emerge only through the cumulative effect of
multiple exchanges. These longer conversations can experience (1) user persistence,
where a user asks for the same thing multiple times and the AI drifts toward
compliance due to LLMs’ tendency toward pleasing the user; (2) content
accumulation, as more information that is patient-specific is added; and (3)
stubbornness of AI LLMs, as current models are more likely to defend an earlier
statement even if it is wrong than to change or revise it.

FDA should evaluate foreseeable conversational pathways, including those that result
in individualized recommendations, triage advice, treatment suggestions, or other
actions that could materially influence care. Factors such as personalization,
expressed confidence, persistence of context, and the degree to which outputs direct
a user toward a specific action should be considered in the assessment.

Intended use should reflect the device’s reasonably foreseeable conversational
behavior and demonstrated capabilities, not solely its labeling or individual response
disclaimers. If a device can reasonably be expected to provide action-directing
information through the course of a conversation, risk assessment should account
for that function. Manufacturers should demonstrate performance and safety across
 October 15, 2026
Page 8 of 22

representative multi-turn interactions, including scenarios that test the boundaries of
the device’s intended scope.

Regarding physician oversight, where conversational outputs initiate or materially
influence clinical actions, the AAFP believes oversight must remain meaningful. A
physician presented only with a device's final recommendation cannot assess
whether it emerged from sound reasoning, from an early error carried forward, or
from user pressure over successive turns. Meaningful oversight of a conversational
device therefore requires access to the entire conversation, not merely to its
conclusion. We recommend FDA treat reviewability of the full exchange as a design
requirement wherever a device's conversational output is intended to inform clinical
action.

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?

The AAFP supports FDA’s recognition that both under-escalation and over-escalation
may cause harm. Failure to direct a patient to appropriate care can delay diagnosis
and treatment, while unnecessary escalation can increase costs, burden emergency
departments, contribute to patient anxiety, and erode trust. For high-acuity
symptoms, the framework may appropriately place greater weight on avoiding
under-escalation while continuing to monitor harms associated with excessive overescalation. Manufacturers should report these error types separately rather than
combining them into a single measure, with prespecified acceptance thresholds
based on the severity, reversibility, and likelihood of harm in the intended clinical
context.

However, we urge FDA to recognize that separate reporting alone will not
produce a balanced picture, because the two error directions are not equally
detectable. Over-escalation surfaces immediately and visibly; for example, the
patient presents to an emergency department, testing is performed, the cost is
incurred, and the error is readily attributed to the device. Under-escalation surfaces
weeks or months later as a delayed presentation that no reporting system will
connect to a triage interaction that preceded it, and that the patient does not know
occurred. A manufacturer acting in good faith will therefore characterize overescalation well and under-escalation poorly. The resulting data will give the
appearance of strong performance, because the errors of greatest concern never
enter the dataset. Additionally, the costs of over-escalation fall largely on emergency
departments and primary care practices rather than on manufacturers. We encourage
 October 15, 2026
Page 9 of 22

FDA to not assume that manufacturer incentives are neutral between the two
directions of error.

Under-escalation can be identified only prospectively; for example, by determining
whether patients advised to monitor at home subsequently presented for urgent or
emergency care, were admitted, or received a serious diagnosis within a defined
follow-up period. While the AAFP recognizes that establishing such linkages is
demanding and time-intensive, we do not believe there is an alternative, as
complaint-driven and spontaneous reporting are structurally incapable of detecting
the failure mode FDA identifies as a top concern. The AAFP recommends FDA require
outcome-linked surveillance with a specified follow-up window for escalation
functions, given that the interoperability infrastructure necessary to support such
surveillance exists, although its availability, implementation, and maturity remain
uneven across health care settings. We recommend FDA consider and account for
the governance, technical assistance, and implementation support necessary to
ensure this infrastructure can be used reliably for surveillance purposes.

Escalation performance is also dominated by prevalence in ways that directly
impact primary care. An escalation function validated in an emergency department
population, where the prevalence of serious pathology may be several percent, will
behave very differently in an ambulatory primary care panel where it is a fraction of a
percent. Positive predictive value falls sharply as prevalence declines even when
sensitivity and specificity are unchanged, so a tool that appears well calibrated in its
validation setting may generate many false escalations for each true one in family
medicine, while the manufacturer's reported performance figures remain accurate.
We recommend FDA require that escalation performance be reported together with
the prevalence of the validation population and that manufacturers report expected
performance at prevalences representative of the settings in which the device will
actually be deployed. We wish to reiterate that validation conducted in academic
medical centers and large integrated systems does not adequately establish
performance in ambulatory primary care.

Section V: A Competency-Based Approach for Premarket Evaluation of GenAIEnabled Devices

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?

Overall, the AAFP supports FDA’s proposed competency-based approach combining
device benchmarking and clinical confirmation. However, we caution against the use
of the term “competency” to characterize the performance of GenAI-enabled
 October 15, 2026
Page 10 of 22

devices. In medical education, competency generally describes a physician’s
demonstrated ability to apply knowledge and clinical reasoning across a range of
circumstances, while performance on a model benchmark does not necessarily
demonstrate the same ability to generalize beyond the conditions on which the
device was evaluated. A GenAI-enabled device may perform well on curated clinical
questions but perform differently when presented with novel or out-of-distribution
cases. The AAFP encourages FDA to use terminology that clearly distinguishes
benchmark performance from demonstrated clinical performance and to ensure
that benchmarking does not substitute for robust clinical confirmation.

Given the open-ended nature of many GenAI-enabled systems, traditional testing
methods alone may not adequately characterize performance. The AAFP
recommends that FDA maintain a risk-based approach and meaningfully engage
practicing physicians throughout development, validation, implementation, and
ongoing evaluation. When a device influences clinical care, evaluation should
confirm that it augments rather than displaces clinical judgment and that an
identifiable party remains accountable for its clinical effect. For GenAI-enabled
devices integrated into clinical workflows, a licensed physician should retain
responsibility and authority for individual clinical decisions while manufacturers
remain accountable for a device’s safety, performance, surveillance, and when
necessary, remediation.

Clinical confirmation should assess not only technical performance and safety but
also the effects of GenAI-enabled devices on care delivery. In primary care, relevant
considerations include care coordination, patient understanding, workflow
integration, physician workload and cognitive burden, and whether the technology
supports meaningful patient-physician interactions. A device that performs
accurately on individual tasks may nevertheless create unintended harms if its
implementation fragments care, increases administrative burden, undermines clinical
judgment, or disrupts clinical workflows.

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?

The amount and rigor of evidence required should be proportional to the
independence of the device’s activity, the consequences of an incorrect output, and
the device’s potential impact on patient care. Higher-risk applications, including
those that influence diagnosis, treatment, prescribing decisions, or triage should
require more robust evidence than functions demonstrated to have limited clinical
consequences, regardless of whether they are characterized as administrative or
informational. We emphasize that functions labeled administrative frequently carry
 October 15, 2026
Page 11 of 22

clinical consequence – both a scheduling function that deprioritizes a symptomatic
patient and a documentation function that shapes what enters the record can impact
patient care, whether or not it is described as clinical.

We encourage FDA not to collapse these two axes into a single tier, as the nature
of the evidence required should reflect the distinct risks presented by each axis. A
function with limited independence but severe consequences and a function with
high independence but modest consequences may occupy similar positions on a
combined scale while requiring different types of evidence, rather than different
amounts. The severity of potential consequences should inform the evidence
necessary to demonstrate accuracy, characterize potential failures, and assess the
consequences of an incorrect output. The degree of activity independence should
inform the evidence necessary to demonstrate appropriate safeguards, human
factors considerations, reviewability, and opportunities for meaningful human
intervention. We therefore encourage FDA to evaluate these considerations
separately rather than combining them into a single measure of evidentiary burden.

The AAFP recommends that FDA consider a third factor that the two axes do not
currently capture, which is the breadth and variability of outputs a device may
generate. A narrowly scoped generative function and a general-purpose
conversational device may present similar levels of risk under the two-axis framework
while requiring substantially different evidence to adequately characterize their
performance and potential risks. Evidence requirements should account for the range
and variability of a device’s potential outputs and be transparent and predictable
while maintaining flexibility for future innovations.

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 proposed benchmarking structure is a useful foundation, but it is unlikely to
provide sufficient evidence of safety and effectiveness on its own. FDA appropriately
recognizes that structured benchmarking cannot fully capture how GenAI-enabled
devices interact with users, workflows, patient populations, and deployment
environments, underscoring the need for clinical confirmation.

The AAFP recommends FDA address an important methodological limitation
associated with the use of publicly available benchmarks, given that published
benchmark materials may be included in the training data used to develop GenAI
 October 15, 2026
Page 12 of 22

models. As a result, strong performance on a publicly available benchmark may
reflect prior exposure to the benchmark rather than the device’s ability to perform
the underlying task, making it difficult to determine what the reported score
demonstrates. To ensure benchmarking provides meaningful evidence of device
performance, the AAFP recommends that FDA require evaluation using appropriate
materials that are not available to developers, are updated on a defined schedule, and
have documented provenance to minimize the likelihood they were included in
training data.

The AAFP believes several important elements should be strengthened or made more
explicit, including performance across multi-turn conversations, workflow usability,
automation bias and user over-reliance, traceability to supporting evidence or source
materials, and users’ ability to recognize and appropriately respond to incorrect
outputs. These factors can substantially affect real-world safety but may not be fully
reflected in the proposed categories. We recommend FDA further explore the use of
externally developed consensus standards and qualified third-party evaluation
programs where appropriate, through a transparent, multi-stakeholder process that
promotes consistency, independence, and scientific validity.

The AAFP recommends FDA clearly define generalizability to include performance
across different practice settings, patient populations and disease prevalence,
languages, and changes in clinical evidence over time. Evaluation should also
account for changes in clinical guidelines and standards of care to ensure that
benchmark results continue to reflect current evidence-based practice.

Regarding external standards, the AAFP would support FDA’s engagement with
existing frameworks including the National Institute of Standards and Technology’s AI
Risk Management Framework and relevant consensus work by health IT standards
bodies through a transparent, multi-stakeholder process.

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?

The AAFP supports FDA’s recognition that prospective clinical trials may not be
necessary in every circumstance. Sponsors should select and justify the clinical
 October 15, 2026
Page 13 of 22

confirmation method based on the device’s risk, as determined by a risk-based
framework like we’ve proposed throughout this response, including intended use and
users, clinical consequences, degree of autonomy, deployment setting, and variability
of anticipated inputs. Confirmation should use inputs representative of the
populations, workflows, and conditions expected in actual use.

Regarding the anticipated distribution of real-world inputs, we note that for
generative devices this distribution cannot be enumerated in advance and that harm
concentrates in its tail rather than its center. Evaluation should include atypical
clinical presentations, incomplete or unclear user inputs, non-English language
inputs, and other reasonably foreseeable scenarios in which device performance may
differ from expected performance. We recommend FDA require sponsors to describe
how these scenarios were incorporated into clinical confirmation and demonstrate
that evaluation was not limited to the most common or expected inputs.

Retrospective analyses, shadow deployments, or physician review may provide
sufficient clinical confirmation for some lower-risk functions. However, the AAFP
encourages FDA to recognize the limitations of retrospective confirmation when a
device’s primary risks involve errors of omission. Retrospective review may identify
incorrect outputs that are documented in the clinical record but may be less effective
at identifying instances in which a device fails to identify a clinically significant finding
or recommend appropriate escalation. When missed findings or failures to escalate
could result in significant patient harm, retrospective analysis alone should not be
considered sufficient clinical confirmation.

Prospective clinical evaluation should be required when a device initiates or
materially influences diagnosis, treatment, prescribing, emergency-care direction,
or another action with a meaningful risk of serious or irreversible harm. Evidence
of performance, however, should not by itself justify autonomous exercise of clinical
judgment without meaningful physician oversight. Real-world evaluation should
complement, rather than substitute for, evidence appropriate to the device’s
premarket risk.

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?

Synthetic data may be useful for supplementing evaluation of rare conditions or
scenarios that are difficult to observe at sufficient scale. Synthetic augmentation is
 October 15, 2026
Page 14 of 22

better suited to rare but well-characterized presentations, to adversarial and stress
testing, and to privacy-preserving augmentation of populations already well
represented in the source data. Safeguards should include comparison against
independent real-world datasets, subgroup-specific performance reporting,
documentation of the generation process and source-data limitations, and
evaluation by experts independent of the model developer. Synthetic data
generated by a model from the same class as the device should not serve as the
sole basis for evaluating the performance gaps it is intended to detect, particularly
for populations that have historically experienced disparities in access, outcomes,
or representation in training data.

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?

Comparators and acceptance criteria should reflect the device’s intended use, users,
deployment setting, and clinical consequences. Qualified physician panels should
include expertise in both the clinical domain and intended users. Specialist input may
be necessary when safe performance depends on specialty-level knowledge but
should not automatically be used for a tool intended to support a generalist practice.

The applicable comparator should generally reflect an evidence-based, clinically
appropriate standard of care rather than median clinician performance, which may
reflect unsafe or outdated practices. When physician review or intervention is part of
the authorized use, human-AI team performance should be the primary basis for
evaluation because it best reflects real-world deployment. However, device-only
performance should also be assessed to identify intrinsic failure modes, characterize
automation-bias risks, and evaluate the potential consequences of foreseeable overreliance or reduced human oversight.

When human-AI team performance is used as the basis for evaluation, the AAFP
recommends that participating physicians be representative of the device’s intended
users and that evaluation reflect realistic conditions of use. Studies that rely primarily
on clinicians with specialized training or experience using the device may not
adequately reflect performance in routine clinical practice. Evaluation should
include physicians with varying levels of experience with the technology and
reflect realistic clinical conditions, including time constraints, workload, and
workflow demands.
 October 15, 2026
Page 15 of 22

As discussed further in our response to Question 15, the AAFP recommends FDA use
these comparators for distinct but complementary purposes. An evidence-based
standard of care should inform whether device performance is clinically appropriate,
while the care pathway likely to occur without the device can help determine
whether the device improves upon existing care in its intended setting. Performance
that exceeds the likely alternative care pathway should not compensate for failure to
meet an appropriate clinical standard.

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?

Evaluation should compare performance with the care pathway most likely to occur
for the intended population and setting in the absence of the device, such as unaided
clinical judgment, delayed specialist consultation, or lack of access to specialty
expertise. Sponsors should justify that pathway using clinical workflow data,
published evidence, or real-world utilization data and should avoid selecting an
artificially weak comparator that exaggerates the device’s apparent benefit. The
selected comparator should reflect actual patterns of care and access for the
population and setting in which the device is intended to be used, including
circumstances in which patients may face delayed or limited access to specialty care.
Where the expected alternative course of care differs meaningfully across
populations or care settings, evaluation should account for that variation rather than
relying on a single comparator that may not accurately represent real-world care.
Comparisons should also consider clinically meaningful patient outcomes and
potential harms, rather than focusing solely on whether the device performs a
particular task more accurately or efficiently than the alternative.

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?

Independent third parties may be particularly valuable for validating benchmark
design, evaluating subgroup performance, adjudicating open-ended clinical outputs,
 October 15, 2026
Page 16 of 22

and confirming sponsor-reported results. Participating organizations should
demonstrate relevant technical and clinical expertise and disclose financial,
contractual, governance, or data relationships with the manufacturer and any
foundation-model developer. Where third-party evaluators or benchmarking
organizations are utilized, FDA should support inclusion of practicing physicians and
other end-users in benchmark development, test design, and evaluation activities to
help ensure assessments reflect real-world clinical practice and patient-care needs.
We also encourage FDA to name specialty societies specifically among the parties
whose participation it should seek. Societies maintain clinical guidelines, convene
practicing clinicians across diverse settings, and can contribute domain content and
reviewers without the commercial relationships that complicate other forms of thirdparty participation. The AAFP would welcome the opportunity to contribute to family
medicine and primary care domains.

Third-party review may be impractical when proprietary-system access or sensitive
clinical data cannot be provided safely. To avoid limiting competition, FDA should
establish common qualification criteria and transparent methods while avoiding
exclusive accreditation arrangements, excessive fees, or requirements that
disproportionately disadvantage smaller developers. Appropriate review and appeal
processes should also be available.

Section VI: Postmarket Monitoring for GenAI-Enabled Devices

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?

The AAFP does not believe increased postmarket monitoring should routinely
substitute for adequate premarket evidence in higher-risk clinical applications.
Devices that materially influence diagnosis, treatment, prescribing, or emergency
care decisions must be supported by robust premarket evidence prior to widespread
deployment. Greater reliance on postmarket monitoring may be appropriate for
lower-risk GenAI-enabled devices when there is a well-defined intended use; limited
potential for patient harm; and a comprehensive monitoring program that includes
predefined performance thresholds, active surveillance, adverse event reporting,
performance degradation monitoring, and timely corrective action processes.

To justify reduced premarket evidence, manufacturers should demonstrate the ability
to detect clinically meaningful performance changes and implement effective
 October 15, 2026
Page 17 of 22

mitigation measures before patient harm becomes widespread. Monitoring programs
must include ongoing evaluation across relevant patient populations and deployment
settings, with clear accountability for responding to identified risks.

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?

The AAFP supports periodic re-benchmarking, physician review, and monitoring for
performance degradation. FDA should also consider additional triggers for
reassessment, including:
• Clinically significant performance drift;
• Changes in patient population or practice settings;
• Meaningful model or workflow updates;
• High or low override rates;
• Safety events, near misses, or clinician-reported concerns; and
• New clinical guidelines or standards of care.

With that said, we believe the override-rate signal trigger needs more specificity. A
high override rate may indicate poor device performance, poor workflow integration,
or clinician distrust, with each warranting different responses. A very low override
rate warrants equal attention, since it may reflect either excellent performance or
uncritical acceptance, and these are not distinguishable from the rate alone. We
recommend FDA treat unexpected values in either direction as a trigger for
investigation rather than monitoring only the high end.

The routine reassessment cadence should be risk-based and informed by the
device’s clinical consequences, degree of autonomy, observed rate of change,
volume of use, and stability of the underlying model and deployment environment.
Event-triggered reassessment must occur promptly when predefined thresholds are
met, rather than waiting for the next scheduled review.

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?
 October 15, 2026
Page 18 of 22

The AAFP believes machine-based supervisory agents may help facilitate postmarket
monitoring activities such as performance degradation, trend detection, audit-log
review, and identification of potential safety signals. Such tools may improve the
scalability and timeliness of postmarket surveillance programs. However, supervisory
agents should support, not replace, human oversight and clinical judgment.
Findings generated through automated monitoring should be reviewed and approved
by qualified individuals before they are relied upon for safety determinations,
regulatory reporting, corrective actions, or other consequential decisions.
Manufacturers should remain responsible for monitoring device performance,
investigating identified risks, and ensuring continued safety and effectiveness
throughout the product lifecycle.

FDA should evaluate these supervisory agents as components of the monitoring
system and require evidence that they can reliably detect clinically meaningful
performance changes and appropriately escalate concerns for human review.
Evaluation should address sensitivity, false positive and false negative rates,
performance drift, transparency, auditability, and performance across diverse
populations and deployment settings. The role, limitations, and oversight processes
associated with a supervisory agent should be clearly defined and documented.
While automated surveillance may strengthen postmarket monitoring, it should
not eliminate meaningful human review or responsibility for patient safety.

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?

Physicians, health care organizations, professional societies, standards development
organizations, and researchers can help identify safety and usability concerns, report
performance issues, and inform best practices. These activities should complement
rather than replace manufacturer responsibilities to monitor performance, investigate
risks, implement corrective actions, and maintain safety and effectiveness throughout
the product lifecycle.

We recommend postmarket monitoring programs include simple reporting
pathways, timely feedback, and adequate technical and financial support for
participating practices. Such programs should not rely on uncompensated physician
labor, nor should they transfer investigation or remediation responsibilities away from
manufacturers and onto frontline physicians.

The AAFP encourages FDA to consider potential legal and liability concerns that may
discourage physicians from reporting device-related safety events. Physicians may be
 October 15, 2026
Page 19 of 22

reluctant to report events involving their own patients if they are concerned that such
reports could be used in subsequent litigation. These concerns may present
particular challenges for physicians in small and independent practices that have
limited access to institutional legal or risk-management resources. We encourage
FDA to explore appropriate protections for good-faith physician reporting, consistent
with existing patient safety reporting frameworks, to support robust postmarket
surveillance.

22. CDRH envisions that a manufacturer’s premarket competency-based assessment
might serve as a baseline against which post-deployment modifications could be reevaluated. 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?

The AAFP reiterates the position stated in our December 2025 response to the
FDA’s request for information on real-world device performance: physicians must
be notified when device behavior changes materially under an approved
predetermined change control (PCCP) plan. A physician relying on a device in
clinical care cannot exercise meaningful judgment about that reliance if the device's
behavior has changed and they have not been notified. This concern is heightened
for generative devices, where a modification may alter the substance, specificity, or
directiveness of outputs without any change to the device's labeling, indications, or
user interface. We recommend FDA establish notification expectations as a condition
of PCCP approval.

Risk-based reevaluation should accompany changes that could materially affect
clinical performance, safety, usability, patient outcomes, intended use, autonomy,
clinical recommendations, or performance in specific populations. Administrative or
technical changes that do not alter intended use, clinical workflow, access to care,
system permissions, data sources, output behavior, or performance may be managed
through established quality system processes. An “administrative” designation alone
should not determine the level of review. A PCCP may be appropriate for bounded,
foreseeable changes when the modification protocol, validation methods, impact
assessment, monitoring expectations, and acceptance criteria can be specified in
advance. Changes that materially expand intended use or increase the severity of
potential harm should be required to receive additional FDA review.
 October 15, 2026
Page 20 of 22

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 developerinitiated changes do not compromise the safety or effectiveness of the device?

Manufacturers should have sufficient contractual and technical visibility into
material model changes to evaluate how those changes to an underlying thirdparty foundation model affect device performance, safety, and intended use. A
manufacturer should not deploy or continue using an updated foundation model
when it lacks sufficient notice, documentation, access, or technical controls to
determine whether the change affects the authorized device.

Relevant mechanisms may include contractual requirements for advance notice of
material changes; version control and change logs; access to information necessary
for impact assessment; technical monitoring for unexpected behavioral changes; the
ability to suspend or roll back deployment; and predefined revalidation procedures.

When foreseeable model changes are addressed through a PCCP, the plan should
identify the permitted scope of change, testing requirements, acceptance criteria,
monitoring obligations, and conditions requiring FDA review. Manufacturers should
also disclose how data transmitted to a third-party model provider may be collected,
retained, used, or shared, and they should protect identifiable patient information
consistent with applicable privacy and consent requirements.

The AAFP encourages FDA to consider whether these safeguards are reasonably
available to manufacturers of different sizes and with varying levels of negotiating
leverage. Smaller developers, including those developing tools for independent and
primary care practices, may have less ability to negotiate contractual provisions
regarding advance notice of model changes, change documentation, or other
protections from third-party foundation model providers. FDA should ensure that its
regulatory approach does not depend solely on a manufacturer’s ability to negotiate
such terms and should provide clear expectations for alternative technical safeguards
when appropriate.

We encourage FDA to consider whether certain baseline obligations should be
attached to foundation model providers whose models are incorporated into
authorized devices, rather than depending on the device manufacturer's negotiating
position. Where such terms cannot be obtained, FDA should specify what technical
monitoring would be accepted as a substitute, so that inability to negotiate favorable
contractual terms does not by itself foreclose participation.
 October 15, 2026
Page 21 of 22

Section VII: Other Topics

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 GenAI-enabled devices warrant heightened scrutiny because they may
autonomously execute multistep actions, use external tools or data, and
propagate an error across an entire workflow before a person has an opportunity
to intervene. The degree of permitted autonomy should therefore be proportionate
to the function’s clinical risk and the reversibility of its actions. Functions that initiate
or materially influence high-consequence clinical actions should require meaningful
oversight and accountability by a licensed physician, including a practical opportunity
to review, override, or reverse an action before harm occurs, when clinically feasible.

Evaluation should assess the safety of the complete action sequence, not merely
the accuracy of individual outputs. Agentic systems should demonstrate an ability to
recognize when a task exceeds their authorized scope, respond appropriately to tool
failures or conflicting information, and reliably escalate issues requiring human
review. Testing should also assess foreseeable combinations of individually minor
errors that could create significant cumulative harm.

The AAFP emphasizes that the opportunity for human review must be meaningful
in practice in order to serve as an effective safeguard. Review processes that
require physicians to evaluate numerous agent-generated actions under significant
time or workload constraints may limit the effectiveness of human oversight. As
shared elsewhere throughout this response, we believe that when human review is
relied upon as a risk-control measure, manufacturers should be required to
demonstrate that physicians have sufficient time, information, and ability to
meaningfully review, modify, or reject device actions under realistic conditions of
use.

The AAFP recommends that agentic GenAI-enabled devices maintain a traceable
record of actions performed by the device, including the sequence of actions and the
information relied upon at each step. This information is important to support
meaningful physician review, facilitate investigation of safety events, and identify
where errors may have occurred within a multi-step process. We strongly urge FDA
to consider such traceability as a foundational, essential safety and accountability
requirement for agentic functions, particularly when a device initiates or
 October 15, 2026
Page 22 of 22

materially influences clinical actions. For systems that invoke other models, agents,
tools, or external services, traceability should extend across the full chain of activity.
Manufacturers should identify which component performed each action, what data
were exchanged, how permissions were inherited or expanded, and how errors or
unauthorized disclosures can be contained across interconnected systems.

Conclusion

Thank you for the opportunity to comment on this important discussion paper and its
potential impacts on primary care in America. The AAFP supports FDA’s efforts to
develop a risk-based, total product lifecycle approach to the oversight of GenAIenabled medical devices. While these technologies hold significant promise to
improve patient access and clinical efficiency, their successful integration into health
care requires robust evaluation, ongoing monitoring, transparency, and meaningful
clinical oversight. We believe regulatory expectations should account not only for
the accuracy of individual outputs, but also for autonomy, actionability,
permissions, cumulative effects, deployment context, and the device’s capacity to
influence or execute clinical actions. Regulatory approaches should remain focused
on ensuring that innovation advances patient care without creating new risks,
exacerbating disparities, or undermining trust in the health care system.

We encourage FDA to continue engaging stakeholders as these policies evolve and
stand ready to serve as a resource and partner in developing evidence-based
frameworks that support the responsible deployment of GenAI-enabled medical
devices. For more information or questions, please contact Mandi Neff, Senior
Strategist, Regulatory and Policy, at mneff2@aafp.org.

Sincerely,

Jen Brull, MD, FAAFP
Board Chair
American Academy of Family Physicians

i
Slight SP, Bates DW. A risk-based regulatory framework for health IT: Recommendations of the
FDASIA working group. J Am Med Inform Assoc. 2014;21(e2):e181-e184. doi:10.1136/amiajnl-2014002638.