American Gastroenterological Association (Byron Cryer)

The American Gastroenterological Association (AGA) appreciates the opportunity to comment on the discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices. The AGA is the trusted voice of the GI community, advancing the science and practice of gastroenterology. Our organization includes nearly 14,000 members from around the globe, including over 11,000 U.S. members, engaged in all aspects of the science, clinical practice, and advancement of gastroenterology. Our comment letter is attached.

October 19, 2026

Rick Abramson, MD, MHCDS, FACR
Director, Digital Health Center of Excellence
U.S. Food and Drug Administration
Center for Devices and Radiological Health
10903 New Hampshire Avenue, Bldg. 66, Rm. G609
Silver Spring, Maryland 20993-0002

Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices - Discussion
Paper and Request for Feedback

Dear Dr. Abramson,

The American Gastroenterological Association (AGA) appreciates the opportunity to comment
on the discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical
Devices. The AGA is the trusted voice of the GI community, advancing the science and practice
of gastroenterology. Our organization includes nearly 14,000 members from around the globe,
including over 11,000 U.S. members, engaged in all aspects of the science, clinical practice, and
advancement of gastroenterology.

Our letter addresses the questions most relevant to AGA members, their patients, and our
organization’s expertise, including risk assessment, patient- and healthcare professional-facing
functions, evaluation of open-ended outputs, postmarket monitoring, accountability, and
reliance on third-party foundation models. Across these issues, AGA recommends a riskproportionate, total product life cycle approach that preserves manufacturer accountability,
evaluates safeguards rather than presuming their effectiveness, and considers the care likely to
occur in the absence of the device.

FDA question 1
AGA agrees that the proposed two-axis framework, organized around device activity and the
consequences of relying on an incorrect output, provides a useful baseline for assessing the risk
of a generative artificial intelligence-enabled medical device software function. However, these
two dimensions alone do not fully capture the factors that determine whether a function will be
safe and effective in clinical practice.

AGA recommends that FDA expand the framework to account for contextual safeguards, user
characteristics, pre-answer behaviors, output variability, real-world access to care, and
postmarket performance.

Contextual safeguards and intended-user expertise
The risk associated with a GenAI-enabled function depends partly on the safeguards
surrounding its use and the intended user’s ability to evaluate the output. Relevant
considerations include whether a qualified healthcare professional reviews the output, whether
the user possesses the clinical expertise needed to identify an inaccurate or incomplete
recommendation, and whether the user has meaningful authority to reject or override the
device’s output.
The presence of a physician or specialist, however, should not automatically be treated as an
adequate safeguard or an infallible reference standard. Clinician judgment varies, and a
physician reviewing a device output may identify an error but may also introduce one.

The role of generalist and specialist reviewers is discussed under Question 4 and recommended
comparators and reviewer methods under Questions 7 and 14.

Usability and acceptability
FDA should include usability and acceptability among end-users as distinct risk considerations.
A device may demonstrate acceptable technical performance but still present risk if its interface,
communication style, or workflow integration causes users to misunderstand an output,
overlook important limitations, or place inappropriate confidence in a recommendation.

The framework should consider whether intended users can understand the output, recognize
uncertainty, identify when the device is operating outside its intended scope, and act
appropriately on warnings or escalation instructions. Human factors testing should evaluate
these issues in the types of clinical settings and under the conditions in which the device is
intended to be used.

Time pressure and action reversibility
Time pressure and the reversibility of a resulting action should also modify the assigned risk. An
incorrect output that can be reviewed and reversed before affecting patient care generally
presents a different risk from an output that prompts an immediate or difficult-to-reverse
intervention.

FDA should consider:
• The amount of time available to identify and correct an error;
• Whether a decision is made in a routine or time-sensitive setting;
• Whether the user can obtain additional information or consultation before acting;
• Whether the resulting action can be reversed; and
• The severity and duration of harm if correction is delayed.

Pre-answer safety behaviors
FDA should also evaluate the steps a GenAI-enabled function takes, or fails to take, before
producing that output.

A response may be factually accurate in isolation but unsafe in context if the device fails to:
• Ask a question that could materially change the recommendation;
• Obtain information necessary to provide a clinically appropriate response;
• Recognize that the available information is insufficient;
• Detect personal identifiers or sensitive health information included in the input; or
• Defer or escalate when the request falls outside the device’s intended use.

AGA recommends that FDA recognize these pre-answer behaviors as a separate dimension of
risk. Evaluation should determine not only whether the final output is accurate, but also
whether the device gathers and interprets the information necessary to produce a safe response.
Variability across repeated responses
FDA should also account for variability across repeated outputs. Unlike conventional software, a
GenAI-enabled function may produce different responses to the same or substantively
equivalent inputs. A device that provides an appropriate response in one instance but an unsafe
response in another may present a risk that is not captured by evaluating a single output.
Therefore, manufacturers should assess the consistency of device behavior across repeated trials
and semantically equivalent inputs. The risk framework should consider the frequency, severity,
and clinical significance of inconsistent responses, particularly when variability affects
recommendations concerning diagnosis, treatment, medication use, or escalation of care.

The risk of delayed or absent care
Access to appropriate care should be a central component of the risk framework. The relevant
comparison is not always between a GenAI-enabled device and immediately available specialist
care. In many real-world settings, particularly in specialty care, the alternative may be delayed
access to an appropriate clinician, care delivered in an unsuitable setting, or no care at all.

FDA should consider the likely course of care in the absence of the device. A device may
introduce some risk while also reducing the potentially greater risk associated with delayed or
absent care. Conversely, a device that appears to improve access may create additional risk if
patients or clinicians rely on it in place of necessary specialty evaluation.

The appropriate counterfactual should be identified and justified for the device’s intended
population and deployment setting. Relevant considerations may include specialist availability,
expected wait times, geographic or socioeconomic barriers, and the clinical consequences of
delayed diagnosis or treatment.

Postmarket performance and change over time
Because GenAI-enabled device performance may change after deployment, FDA should apply
the framework across the total product life cycle. Risk assessment should consider whether the
manufacturer can detect model drift, performance degradation, changes in relevant data
distributions, and new failure modes under real-world conditions. AGA provides more detailed
recommendations concerning postmarket monitoring and stakeholder roles in its combined
response to Questions 18 and 21.

Recommendation:
AGA recommends that FDA retain device activity and the consequences of relying on an
incorrect output as the framework’s foundational axes. FDA should supplement those axes with
structured modifiers addressing intended-user expertise, the demonstrated effectiveness of
safeguards, usability and comprehension, time pressure and action reversibility, pre-answer
safety behaviors, repeated-response variability, the care likely to occur without the device, and
the device’s ability to maintain performance over time. FDA could use these factors to adjust a
function’s placement within the framework and determine the appropriate evidence, labeling,
safeguards, and postmarket controls.

FDA questions 2 and 3
AGA recommends that FDA assess patient-facing informational functions according to the
degree of personalization, the extent to which the output directs clinical action, the
consequences of relying on an incorrect output, and the patient’s ability to recognize the device’s
limitations. The inclusion of a disclaimer or general instruction to consult a healthcare
professional should not, by itself, convert an otherwise action-directing output into nondirective information.

AGA recommends distinguishing among three categories:
• General educational functions, which provide non-personalized information
without recommending a particular clinical action;
• Personalized but non-directive functions, which apply information to a patient’s
circumstances but stop short of endorsing a particular action; and
• Personalized action-directing functions, which recommend, endorse, or
discourage a diagnosis, treatment, medication, test, procedure, site of care, or timeframe.

Evidence requirements and safeguards should increase as an output becomes more personalized
and action-directing. For higher-risk functions, FDA should require either meaningful review by
a qualified healthcare professional who has the information, expertise, time, and authority to
independently evaluate the recommendation, or sufficient evidence that autonomous use is safe
and effective for the intended population and clinical context.

Consistent with the accountability principle described above, a manufacturer that delivers
action-directing guidance directly to patients should remain responsible for the function’s
safety, effectiveness, monitoring, and corrective action. A manufacturer should not be permitted
to bypass clinician review while disclaiming responsibility for the resulting recommendation.
FDA should also require clear descriptions of the respective roles of the manufacturer, deploying
organization, reviewing clinician, and patient.

Recommendation: FDA should regulate patient-facing informational functions along a
continuum of personalization and directiveness, with progressively stronger evidence,
safeguards, and postmarket controls as the potential consequences of patient reliance increase.

FDA question 4:
AGA agrees that the distinction between generalist and specialist healthcare professionals may
inform risk, but professional designation alone should not determine the classification of an
HCP-facing GenAI-enabled function. FDA should assess whether the intended user has the
knowledge, information, time, and practical authority needed to critically evaluate the output for
the specific task.

A GenAI-enabled function may expand access to specialty knowledge and improve care when
timely specialist consultation is unavailable. Risk may increase, however, when a user receives
specialized guidance that the user cannot independently assess and is therefore likely to accept
without meaningful review. As the user’s ability to evaluate the output decreases, the evidence
supporting the device and the safeguards surrounding its use should become more rigorous.

FDA should not presume that labeling, training, or the nominal presence of a physician or
specialist provides an adequate safeguard. Manufacturers should demonstrate that intended
reviewers have appropriate expertise, access to relevant information, sufficient time, an
understanding of the device’s limitations, and authority to reject or modify the output before it
affects patient care.
For devices intended to augment a generalist’s decision-making, FDA should evaluate the
performance of the generalist-device team, including whether use of the device improves
decisions, introduces new errors, promotes inappropriate reliance, or changes referral patterns.
The comparator should also reflect the care likely to occur without the device, which could
include delayed specialist review or no specialty evaluation.

We address the selection of objective endpoints, clinician comparators, blinded reviewers, and
methods for reporting reviewer disagreement in our combined response to Questions 7 and 14.

Recommendation: FDA should assess healthcare professional-facing functions using a taskspecific approach that considers the expertise needed to evaluate the output, the intended user’s
ability to identify an error, the likely degree of reliance, the seriousness and reversibility of the
consequences, and the demonstrated effectiveness of available safeguards. Requirements should
reflect the device’s actual role in decision-making rather than treating either generalist or
specialist status as conclusive.

FDA questions 7 and 14
AGA supports a competency-based approach for evaluating GenAI-enabled medical devices,
particularly when devices accept open-ended inputs or produce variable, open-ended outputs.
Traditional testing methods developed for bounded inputs and predictable outputs may not
adequately characterize performance across the range of conditions these devices may
encounter.

AGA recommends that FDA implement a risk-proportionate, three-phase evaluation pathway:
1. Technical and algorithmic validation, including testing of accuracy, reliability,
safety boundaries, security, confabulation, repeated-response variability, and clinically
significant failure modes;
2. Human factors and usability testing, including evaluation of user comprehension,
workflow integration, cognitive load, automation bias, and the ability of representative
users to recognize erroneous or out-of-scope outputs; and
3. Prospective clinical validation, using real-world or clinically representative
conditions to assess the safety and effectiveness of the final user-facing device and,
where applicable, the human-device team.

Evaluation should address the complete deployed configuration, including the foundation
model, prompts, retrieval methods, safety controls, user interface, and workflow components
that may affect performance. Sponsors should prespecify the testing methods, scoring criteria,
failure definitions, and acceptance thresholds for each phase and explain how findings in one
phase inform testing in subsequent phases.

For open-ended outputs, FDA should prefer objective clinical outcomes or clearly defined,
guideline-based endpoints when available. Physician judgment should not automatically be
treated as ground truth because reasonable clinicians may differ and clinician involvement may
correct, overlook, or introduce errors.

When clinician evaluation is necessary, sponsors should use prespecified, item-specific criteria
established before grading, multiple qualified and blinded reviewers, and transparent reporting
of reviewer disagreement. The choice of generalist, specialist, or human-AI team as the
comparator should be justified by the intended use and the role the device is expected to
perform. Human-AI team performance should not be presumed superior to device-only
performance; the benefit of clinician involvement should be demonstrated.

Recommendation: FDA should formalize competency-based evaluation as a three-phase
pathway comprising technical validation, human factors testing, and prospective clinical
validation. For open-ended outputs, comparators and acceptance criteria should be tied to the
intended use, objective endpoints where available, and transparent, reproducible review
methods when expert judgment is required.

FDA questions 18 and 21
AGA supports greater reliance on postmarket monitoring for appropriate lower- and moderaterisk GenAI-enabled devices when real-world evidence is more informative than increasingly
burdensome premarket testing. This approach should not be used when delayed detection of a
failure could expose patients to serious or irreversible harm, when clinically significant failures
cannot be detected promptly, or when effective corrective action is not feasible.

Greater postmarket reliance should be conditioned on a monitoring plan capable of detecting
performance degradation, model drift, emerging failure modes, subgroup differences, and
material discrepancies between premarket testing and actual use. The plan should establish
appropriate monitoring intervals, triggering events, escalation procedures, and corrective
actions.

The manufacturer should retain primary regulatory, legal, and financial responsibility for
surveillance, adverse-event reporting, investigation, and corrective action. Clinicians, healthcare
institutions, academic researchers, professional societies, and standards-setting organizations
may contribute complementary evidence, but their participation should not transfer or diffuse
the manufacturer’s obligations.

Independent academic research and investigator-initiated studies may identify long-term
outcomes or safety concerns not fully captured through manufacturer surveillance. Professional
societies may also contribute specialty expertise through voluntary resources such as open
question banks, evaluation rubrics, and accessible incident-reporting mechanisms. These
resources should supplement, not replace, manufacturer monitoring.

FDA should avoid requiring professional societies, certifiers, or other third parties to serve as
mandatory gatekeepers. High fixed costs, exclusive testing arrangements, or reliance on a
limited number of intermediaries could restrict competition and make validated systems less
accessible. FDA should instead permit flexible, scientifically valid approaches consistent with
the risk and intended use of the device.

Recommendation: FDA should permit increased reliance on postmarket evidence for
appropriate lower- and moderate-risk devices when manufacturers maintain effective
monitoring and corrective-action systems. Outside stakeholders may contribute research,
shared evaluation resources, and incident reports, but manufacturers should remain fully
accountable for continuous surveillance and response.
FDA question 24
Manufacturers of devices built on third-party foundation models should maintain controls that
allow them to identify the model used, detect upstream changes, assess their effect on device
performance, and respond before patient safety is compromised. At a minimum, manufacturers
should:
• Use pinned model and version identifiers where technically available;
• Log the resolved model, version, and provider associated with each device interaction;
• Require contractual notice of material upstream changes, including changes to model
behavior, content policies, output formatting, safety controls, or availability;
• Reexecute a fixed, clinically relevant benchmark on a defined schedule and after relevant
triggering events; and
• Maintain documented procedures for investigating detected changes, suspending
affected functionality when necessary, and determining whether additional validation,
FDA review, or action under a predetermined change control plan is required.

These controls serve complementary purposes. Version identification and logging provide
traceability, contractual notice can provide advance warning, and repeated benchmarking can
detect performance changes even when an upstream modification is not disclosed. The device
manufacturer should remain responsible for determining whether the modified foundation
model continues to support the safety and effectiveness of the finished device.

Recommendation: FDA should expect manufacturers relying on third-party foundation
models to combine technical, contractual, and change-control mechanisms that support timely
detection, evaluation, and response to upstream changes.

We appreciate the opportunity to comment in response to the discussion paper. If you have any
questions or need assistance, please contact Kathleen Teixeira, Vice President of Public Policy
and Advocacy, at Kteixeira@gastro.org or Leslie Narramore, Director of Regulatory Affairs, at
Lnarramore@gastro.org.

Sincerely,

Byron Cryer, MD, AGAF
President, AGA Institute