American Society for Gastrointestinal Endoscopy

See attached file(s)

October 19, 2026

Kyle Diamantas, J.D.
Acting Commissioner of Food and Drugs
U.S. Food and Drug Administration
10903 New Hampshire Ave.
Silver Spring, MD 20993

Submitted at regulations.gov

Re: Docket FDA-2026-N-7874 — Considerations for the Regulation of Generative AIEnabled Medical Devices: Discussion Paper and Request for Feedback

Dear Acting Commissioner Diamantas:

The American Society for Gastrointestinal Endoscopy (ASGE) appreciates the opportunity to
comment on the Food and Drug Administration’s (FDA) discussion paper concerning generative
artificial intelligence (GenAI)-enabled medical devices. ASGE represents gastroenterologists and
endoscopists who use, study and deploy AI in practice.

In 2019, ASGE launched its Artificial Intelligence Taskforce and in 2024 ASGE formed the
Artificial Intelligence Institute for Gastroenterology (AIGI),1 which convenes physicians, AI
experts, innovators, industry representatives and regulators to discuss and develop consensus on
the application of AI in gastrointestinal (GI) endoscopy. Our comments reflect the experience of
ASGE members in the use of AI, as well as published guidance of the ASGE AI Task Force.2,3

As FDA’s Center for Devices and Radiological Health explores a regulatory approach to GenAIenabled medical devices, endoscopy serves as a useful precedent. While FDA-authorized
endoscopic computer-aided detection (CADe) systems are predictive, not generative, they are
still the best-studied example of real-time AI during an invasive procedure, and generative tools
(video-language assistants, report and pathology synthesis, patient-facing preparation and triage
tools) are now reaching endoscopy.

Experience with artificial intelligence in GI endoscopy supports four regulatory principles:
https://aiforgi.org/
Berzin TM, Parasa S, Wallace MB, Gross SA, Repici A, Sharma P. Position statement on priorities for artificial intelligence in GI endoscopy: a
report by the ASGE Task Force. Gastrointest Endosc. 2020;92(4):951-959.
Parasa S, Repici A, Berzin T, Leggett C, Gross SA, Sharma P. Framework and metrics for the clinical use and implementation of artificial
intelligence algorithms into endoscopy practice: recommendations from the American Society for Gastrointestinal Endoscopy Artificial
Intelligence Task Force. Gastrointest Endosc. 2023;97(5):815-824.e1.
• Clinical-trial performance may not translate directly into routine practice. Randomized
trials of computer-aided detection (CADe) showed higher adenoma detection and lower miss
rates, but pragmatic implementation has not always produced the same benefit.4,5,6

• Effects on clinicians must be evaluated, not assumed. Observational evidence raises
concerns about changes in unaided clinician performance after routine exposure to AI.7

• Device evaluation should connect to meaningful clinical standards. GI has outcome-linked
quality measures, including adenoma detection rate, and clinically anchored performance
thresholds for optical diagnosis.8,9,10

• Postmarket evaluation must examine the clinical system, not only technical accuracy.
Monitoring should assess patient outcomes, subgroup performance, workflow effects, clinician
behavior, and performance degradation over time.

ASGE supports a risk-based, total product lifecycle approach that preserves timely access to
beneficial innovation while maintaining meaningful clinical validation, transparency, and
accountability. Clinical specialty expertise should be incorporated throughout the regulatory
lifecycle—from risk classification and benchmark development through clinical confirmation,
change control, and postmarket monitoring.

ASGE has grouped its responses to related discussion questions with the goal of providing useful
clinical perspective as the FDA confronts the unique regulatory challenges associated with
GenAI-enabled medical devices.

RISK ASSESSMENT AND INTENDED USERS

Questions 1 through 6

ASGE position. FDA’s proposed two-axis framework—device activity and the consequences of
relying on an incorrect output—is a useful foundation, but it does not fully capture risk in live
procedural care or other settings in which users have limited time to evaluate an output.
Therefore, FDA should also consider the following risk modifiers:
Repici A, Badalamenti M, Maselli R, et al. Efficacy of real-time computer-aided detection of colorectal neoplasia in a randomized trial.
Gastroenterology. 2020;159(2):512-520.e7.
Wallace MB, Sharma P, Bhandari P, et al. Impact of artificial intelligence on miss rate of colorectal neoplasia. Gastroenterology.
2022;163(1):295-304.e5.
Ladabaum U, Shepard J, Weng Y, Desai M, Singer SJ, Mannalithara A. Computer-aided detection of polyps does not improve colonoscopist
performance in a pragmatic implementation trial. Gastroenterology. 2023;164(3):481-483.
Budzyń K, Romańczyk M, Kitala D, et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonosco
Corley DA, Jensen CD, Marks AR, et al. Adenoma detection rate and risk of colorectal cancer and death. N Engl J Med. 2014;370(14):12981306.
Rex DK, Anderson JC, Butterly LF, et al. Quality indicators for colonoscopy. Gastrointest Endosc. 2024;100(3):352-381. Also published in Am
J Gastroenterol. 2024;119:1754-1780.
Rex DK, Kahi C, O’Brien M, et al. The American Society for Gastrointestinal Endoscopy PIVI (Preservation and Incorporation of Valuable
Endoscopic Innovations) on real-time endoscopic assessment of the histology of diminutive colorectal polyps. Gastrointest Endosc.
2011;73(3):419-422.
• Independent reviewability: whether the intended user has the expertise, source information,
and time necessary to recognize an incorrect or incomplete output.

• Reversibility and time pressure: whether an action can be stopped or corrected before harm
occurs and how quickly the user must decide.

• Downstream safeguards: whether another clinician, workflow control, diagnostic test, or
human-approval checkpoint can intercept an error.

• Traceability and uncertainty: whether the output identifies its sources, communicates
uncertainty, and defers when the system is outside its validated scope.

Directiveness should be determined by what an output enables or prompts the user to do, not
merely by words such as “recommend” or “consider.” Specificity, personalization, urgency cues,
automatic population of an order, and ease of execution can move an output from informational
to directive or action-taking. A disclaimer such as “talk to your doctor” should not reduce the
risk classification of an otherwise directive output.

The intended user also matters. A specialty-level GI output may present different risks when
used by a gastroenterologist, another specialist, a generalist clinician, or a patient because those
users differ in their ability to identify an error and place the information in clinical context. This
distinction should not be used to underestimate patient capability. Rather, FDA should evaluate
the interaction among intended user, directiveness, clinical consequences, and independent
reviewability. AI that extends GI expertise to primary care or underserved settings can create
substantial value when accompanied by clear limitations, guideline sources, referral criteria, and
escalation pathways. Automation bias should be assessed with the intended users and clinical
tasks.11

For multi-turn conversational systems, risk should be based on the highest-risk state the device
can reasonably reach across a foreseeable interaction—not only its initial response. Intended use
should define a bounded envelope of permitted behavior, including the maximum level of
directiveness. Red-flag escalation should activate whenever the relevant symptom appears,
regardless of where it occurs in the conversation.

Care-escalation functions should be evaluated for both under-escalation and over-escalation. For
must-not-miss conditions, sponsors should establish clinically justified sensitivity requirements
and test atypical or downplayed presentations. Over-escalation should be measured separately
because unnecessary emergency referrals can burden patients and under-resourced settings.
Evaluation should use graded decisions, such as self-care, same-day clinical review, and
emergency care, rather than a single binary decision.

COMPETENCY BENCHMARKS
Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. J Am Med Inform
Assoc. 2012;19(1):121-127.
Questions 7, 8, 9, and 10

ASGE position. A competency-based approach—benchmarking followed by clinical
confirmation—is a useful organizing framework when it is tied to the device’s intended use, risk,
version, and clinical environment. A passing benchmark score should not be treated as proof of
clinical benefit, and a competency determination should be reassessed after material changes or
at risk-proportionate intervals.

Specialty societies should help define clinically meaningful competencies, scoring criteria,
acceptance thresholds, and important failure modes. For GI applications, benchmarking should
address:

• clinical knowledge and adherence to intended-use boundaries;
• recognition and escalation of high-risk findings;
• uncertainty communication, clinical deferral, and source grounding;
• performance with incomplete, conflicting, or changing information;
• consistency across relevant patient subgroups, languages, sites, equipment, and practice
settings;
• human–AI interaction, including automation bias, alert fatigue, override behavior, and clinician
performance with and without the device;
• for image and video applications, lesion-level and procedure-level performance, latency, image
quality, and variation across endoscope platforms and processors; and
• failure behavior when the system, external data source, or connected tool is unavailable or
slow.

Sponsors should demonstrate that benchmark performance predicts behavior on real inputs from
the intended population. Public benchmarks alone are insufficient when they are contaminated,
saturated, curated, or unrepresentative. Sponsor-developed benchmarks may be appropriate when
the task blueprint is clinically developed, methods and repeated runs are documented, and an
independently held portion of the test set remains sequestered from the developer.

CLINICAL EVIDENCE AND REPRESENTATIVE DATA

Questions 11 through 13

ASGE position. Evidence requirements should be proportionate to intended use, autonomy,
clinical consequences, and the user’s ability to identify an incorrect output. Regulatory flexibility
should modernize how evidence is generated without lowering the standard the evidence must
meet.

For lower-risk, asynchronous functions that are reviewed by a clinician before use, retrospective
testing on representative patient inputs with blinded clinical adjudication may be sufficient.
Functions that communicate with patients without prior clinician review, influence triage, or
produce directive recommendations may require shadow deployment, standardized-patient
testing, prospective observation, human-factors evaluation, or a prospective clinical study. Realtime procedural and high-consequence functions require evidence that evaluates how clinicians
respond to the output under realistic conditions. Registry-embedded, stepped-wedge, adaptive,
and other pragmatic designs may be appropriate when conventional trials are impractical,
provided the design produces valid evidence.12

Evaluation data should reflect the populations and environments in which the device will be
used, including demographics, disease prevalence and presentation, language, health literacy,
community and academic practice, equipment, data systems, and procedural conditions.
International data may contribute when sponsors demonstrate that the relevant populations,
workflows, and equipment are applicable to the intended United States.

Synthetic data may supplement real-world evidence for rare, adversarial, or safety-critical
scenarios and for controlled variation in language or conversational behavior. It should not
replace representative real-patient evidence needed to establish clinical performance. Synthetic
inputs should be clearly identified, analyzed separately, compared with real anchor sets, and
assessed to ensure that the generator has not reproduced the same subgroup gaps or simplified
the cases the evaluation is intended to test.

Because generative outputs may vary from run to run, sponsors should repeat inputs and report
reliability as well as average performance. Analyses should account for clinically meaningful
clustering and report uncertainty, subgroup results, and the frequency and severity of safetycritical failures.

COMPARATORS AND HUMAN AI TEAM PERFORMANCE

Questions 14 and 15

ASGE position. The appropriate comparator should follow the device’s intended use. Evaluation
should distinguish between the device operating alone, a clinician without the device, and the
clinician and device working together.

When a technology is intended to augment rather than replace a gastroenterologist, the primary
comparison may be the gastroenterologist using the technology versus the gastroenterologist
without it. Device-alone performance does not reliably predict team performance. Evidence
showing that access to a high-performing language model did not necessarily improve clinician
performance reinforces the need to test the actual human–AI team.13

Where established specialty measures exist, acceptance criteria should be clinically anchored
rather than based solely on the median clinician. The updated ASGE and American College of
Gastroenterology quality indicators identify a minimum acceptable adenoma detection rate of at
least 35 percent for the defined measurement population.14 ASGE’s Preservation and
Incorporation of Valuable Endoscopic Innovations (PIVI) standards also provide examples of
Vasey B, Nagendran M, Campbell B, et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by
artificial intelligence: DECIDE-AI. Nat Med. 2022;28:924-933.
Goh E, Gallo R, Hom J, et al. Large language model influence on diagnostic reasoning: a randomized clinical trial. JAMA Netw Open.
2024;7(10):e2440969.
Rex DK, Anderson JC, Butterly LF, et al. Quality indicators for colonoscopy. Gastrointest Endosc. 2024;100(3):352-381. Also published in
Am J Gastroenterol. 2024;119:1754-1780.
task-specific thresholds, including at least 90 percent negative predictive value for a leave-inplace strategy for polyps and at least 90 percent agreement in surveillance-interval assignments
for a resect-and-discard strategy.15

If qualified clinicians serve as adjudicators, the panel should include independent, domainqualified reviewers operating under a prespecified consensus process. The evaluation should
report disagreement, recognize when more than one answer is clinically acceptable, and address
conflicts among guidelines.

FDA should also allow comparison with what would realistically occur without the device, such
as unaided care, delayed specialist review, or no available service. Sponsors should document the
actual baseline in the intended setting and report whether the device improves access without
lowering performance where high-quality care is already available. Device labeling should
identify the comparator used to support the claim.

INDEPENDENT EVALUATION AND MODEL ARCHITECTURE

Questions 16 and 17

ASGE position. Independent third parties can strengthen benchmarking and clinical
confirmation by maintaining sequestered test sets, providing clinical adjudication, conducting
red-team testing, verifying sponsor results, and performing targeted postmarket audits.
Evaluators should possess the relevant clinical and technical qualifications, disclose financial
relationships, and remain independent of device development and calibration.

Third-party involvement should be proportionate to risk. Requiring external review for every
minor modification could create bottlenecks and discourage competition. FDA should permit
multiple qualified evaluators, transparent methods, and approaches that do not require
unnecessary disclosure of proprietary model internals.

The competency framework should also account for model architecture. Multimodal visionlanguage and predictive models require perceptual, temporal, and closed-loop testing that
knowledge and dialogue benchmarks do not provide. Endoscopic video systems should be
evaluated at the lesion, frame, and procedure levels, including latency, image quality, platform
variation, and whether generated descriptions remain grounded in visible findings. Generative
image enhancement should be tested to ensure that it does not create, obscure, or erase clinically
meaningful features. Models that act over time require evaluation of the complete sequence
because small errors can compound during a procedure.

POSTMARKET MONITORING

Questions 18 through 20
Rex DK, Kahi C, O’Brien M, et al. The American Society for Gastrointestinal Endoscopy PIVI (Preservation and Incorporation of Valuable
Endoscopic Innovations) on real-time endoscopic assessment of the histology of diminutive colorectal polyps. Gastrointest Endosc.
2011;73(3):419-422.
ASGE position. Postmarket monitoring can supplement premarket evidence but should not
replace it. Greater premarket uncertainty may be acceptable only when potential harm is limited
and reversible, baseline performance has been demonstrated on representative data, and the
manufacturer has a funded and operational plan to detect, contain, and correct problems quickly.
This approach is not appropriate for functions that provide real-time procedural guidance,
operate autonomously, are difficult for users to evaluate independently, or can cause serious or
irreversible harm. Monitoring occurs after exposure and therefore cannot protect the first patients
affected.

A risk-proportionate monitoring program should include:

• periodic re-benchmarking and safety-critical regression testing;
• clinician review of representative outputs, including samples not flagged by the device or a
supervisory system;
• monitoring for performance drift, subgroup disparities, site-level variation, and changes in
clinician acceptance, override behavior, or unaided performance;
• outcome-linked measures appropriate to the intended use;
• clear performance thresholds, stop rules, rollback capability, and timely user notification; and
• off-cycle reassessment after material changes to the model, prompts, retrieval content,
guardrails, interface, input hardware, clinical guidelines, patient population, workflow, or
intended use.

The cadence of review should be proportionate to risk rather than fixed across all products. More
directive, autonomous, difficult-to-review, or high-consequence functions should undergo more
frequent and intensive reassessment.

Machine-based supervisory agents may support monitoring at scale, but they should triage cases
for human review rather than serve as the sole judge of safety for higher-risk devices.
Supervisory agents require independent validation, their own change control and drift
monitoring, protection against security and privacy threats, and audits of unflagged outputs to
estimate what they miss.

STAKEHOLDER ROLES AND ACCOUNTABILITY

Question 21

ASGE position. Professional collaboration must not diffuse manufacturer accountability.

• Manufacturers should remain responsible for device design and performance, postmarket
surveillance, investigation and reporting of adverse events and performance concerns, and
timely corrective action.

• Clinicians and health care institutions should have practical, standardized mechanisms for
reporting suspected problems and receiving information about the manufacturer’s response.
They should not be required to create bespoke surveillance systems or provide routine,
uncompensated adjudication for commercial products.
• Professional societies can define specialty competencies and clinically meaningful measures,
develop guidance and education, convene stakeholders, and identify qualified experts. These
activities should complement—not replace—the manufacturer’s obligations.

• Standards-setting bodies can support common terminology, reporting formats,
interoperability expectations, and evaluation methods.

Clinicians remain responsible for exercising independent professional judgment within their
scope of practice, while manufacturers remain responsible for device design, performance,
labeling, monitoring, and correction of device-related failures. FDA’s framework should avoid
shifting liability for failures that manufacturers are positioned to detect and prevent onto
individual clinicians or professional societies.16

CHANGE CONTROL

Questions 22 and 23

ASGE position. Reassessment should be scaled to the reach and expected clinical impact of a
modification. Changes that do not alter clinical outputs may be managed within the
manufacturer’s quality system with appropriate regression testing. A Predetermined Change
Control Plan (PCCP) may address bounded updates, such as guideline or knowledge updates
within scope, tighter guardrails, threshold tuning, added languages, or compatible hardware
changes, when the verification process and acceptance criteria are specified in advance.

FDA review should remain necessary when a change materially expands the risk profile,
including changes to intended use, intended users, clinical domain, input modality, directiveness,
autonomy, human oversight, model family or vendor, or safety controls. Any change capable of
affecting clinical outputs should be evaluated against safety-critical cases. Aggregate
performance should not conceal a clinically unacceptable safety failure.

When future modifications cannot be fully prespecified, the change-control plan should define
boundaries that may not be exceeded, permitted categories of change, required validation, staged
implementation, rollback capability, cumulative-change limits, version records, and user
notification. Higher-risk devices should remain locked between validated updates. Continuously
self-modifying behavior should not be permitted under a change-control plan for systems that
guide procedures, provide highly directive recommendations, or undertake autonomous clinical
actions.

FOUNDATION MODELS

Question 25

ASGE position. Voluntary Foundation Model Device Master Files could reduce duplication if
they are standardized, versioned, maintained, and linked to timely update notification. Useful
American Medical Association. Principles for Augmented Intelligence Development, Deployment, and Use. November 2023.
information may include version identifiers; high-level training-data provenance and knowledge
cutoff; health care evaluation methods and subgroup results; known failure modes and red-team
findings; guardrail behavior; deprecation schedules; security and privacy practices; and incident
history.

A master file should not be interpreted as evidence that a particular medical device is safe or
effective. The device sponsor should remain responsible for evaluating the foundation model as
configured and used within the complete device. When a model developer does not provide
adequate information, the sponsor should disclose the limitations and independently characterize
reliability and clinically relevant failure modes through appropriate testing. FDA may also
recognize qualified independent evaluations and consensus risk-management standards,
including the NIST AI Risk Management Framework.17

AGENTIC ARTIFICIAL INTELLIGENCE

Question 26

ASGE position. Autonomous multi-step action materially changes risk because errors can
compound and become irreversible before a clinician can intervene. FDA should evaluate the
complete action sequence—planning, external tools, handoffs, tool failures, and human-approval
checkpoints—not merely individual components.

Testing should address wrong-patient actions, incorrect tool outputs, prompt injection,
incomplete tasks, loss of connectivity, and actions outside the authorized scope. Acceptance
criteria should include patient-identity confirmation, narrowly defined permissions, audit logs,
stop controls, safe defaults, and a defined fallback workflow.

An agent that adjusts sedation, controls an endoscopic instrument, independently triages
findings, or orders a procedure can directly affect patient care and should be evaluated as a
device function. Clinician authorization should be required for high-consequence or irreversible
steps unless autonomous performance has been supported by evidence and authorized for that
use.

Agents used solely for scheduling, prior authorization, or patient outreach may fall outside
device regulation depending on their intended use and functionality, but they can still affect
access, safety, and equity. These systems require appropriate monitoring, escalation, auditability,
and human oversight. Successful testing of individual components should never substitute for
demonstrating that the full workflow performs safely in realistic conditions.

CONCLUSION

ASGE appreciates FDA’s effort to engage stakeholders before establishing a regulatory approach
for GenAI-enabled medical devices. Gastroenterology provides an important clinical
environment in which to examine the transition from assistive image-based tools to multimodal,
National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence
Profile (NIST AI 600-1). July 2024.
generative, and increasingly agentic systems. ASGE and its Artificial Intelligence Institute for
Gastroenterology welcome the opportunity to provide specialty expertise as FDA develops
clinically meaningful benchmarks, evidence expectations, transparency standards, changecontrol requirements, and postmarket evaluation approaches.

FDA can regulate generative AI more effectively when specialty expertise helps define what
competent, safe, and clinically meaningful performance looks like in actual practice.

Thank you for considering our comments. Please contact Lakitia Mayo, ASGE Chief Policy and
Member Engagement Officer, at lmayo@asge.org with questions or requests for additional
information.

Sincerely,

Jason Domintiz, MD, MHS, MASGE
ASGE President

Prateek Sharma, MD, MASGE
AIGI Chair