American College of Surgeons (Patricia L. Turner)

Please see the attached comment from the American College of Surgeons (ACS).

October 19, 2026

Rick Abramson, MD, MHCDS, FACR
Director
Digital Health Center of Excellence (DHCoE)
Food and Drug Administration
10903 New Hampshire Ave
Silver Spring, MD 20993

Re: Considerations for the Regulation of Generative AI-Enabled Medical
Devices: Discussion Paper (FDA-2026-N-7874)

Dear Dr. Abramson,

On behalf of the more than 95,000 members of the American College of Surgeons (ACS),
we appreciate the opportunity to comment on the Food and Drug Administration’s (FDA)
Discussion Paper, Considerations for the Regulation of Generative AI-Enabled Medical
Devices (Discussion Paper). ACS commends FDA for its leadership and for engaging
stakeholders early to help ensure that generative artificial intelligence (GenAI)-enabled
medical devices are safe, effective, and reliable for patient care. The rapidly evolving AI
ecosystem offers significant opportunities to improve patient care, but it also raises
regulatory challenges that existing frameworks were not designed to address. We offer
the following comments as FDA develops its framework, recognizing that our positions
will evolve as real-world evidence emerges.

ACS welcomes the opportunity to partner with FDA as it designs this framework.
Surgeons have direct experience with how these tools are used in the operating room, and
ACS quality programs and clinical registries offer established methods and data for
evaluating them. ACS is also well positioned to help FDA convene the surgical
community to develop consensus on evaluation and monitoring standards for procedural
AI. The sections below describe specific opportunities for collaboration.

III. Definitions and Context

Regulatory Considerations for Procedural AI

We share FDA’s goal of developing a framework that protects patients while supporting
the timely adoption of beneficial technologies. Although the Discussion Paper focuses on
GenAI-enabled devices, we begin by framing our comments around AI-enabled devices
used during surgical procedures, which we refer to as “procedural AI.” What sets
procedural AI apart is that it ties its output to physical action in real time. It can act
directly, such as by steering a surgical robot, or guiding the surgeon’s actions as they
happen, such as through an anatomy overlay on surgical video. ACS is working to refine this definition
and would welcome doing so jointly with FDA.

The surgical setting also raises the stakes. The patient is typically under anesthesia and relies entirely on
the surgical team. Decisions are made under time pressure, often with little opportunity to verify an
output before acting on it, and errors can cause immediate physical harm that may not be reversible.
ACS urges FDA to evaluate procedural functions according to whether clinicians can detect, interrupt,
and reverse an erroneous action in conditions where the device is intended to be used. For functions
capable of causing immediate, irreversible harm, premarket clinical confirmation should establish a
sufficient safety basis before patient use, and postmarket monitoring should assess outcomes as well as
device outputs.

These considerations are especially relevant to GenAI-enabled tools used during surgery, whose outputs
are open-ended and can vary from one use to the next. The Discussion Paper acknowledges the unique
functionalities of GenAI-enabled devices; however, it does not address considerations for procedural AI
tools, including generative, procedural AI tools. The paper is largely built around outputs that a clinician
can review before they reach the patient. The risk grid, benchmarking elements, and clinical
confirmation methods all describe non-procedural AI functionalities, and little in the paper contemplates
a tool that performs part of the intervention, where the surgeon may have little time to review an output,
or may not recognize an error, before it takes effect. The following sections describe how FDA can
adapt its risk framework, premarket evaluation, and postmarket surveillance to close this gap, matching
the evidence required to the risk of each tool.

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

ACS supports FDA's two-axis framework as a starting point. For procedural GenAI, however, both axes
need adjustments, and other dimensions should be taken into consideration as well.

Question 1
Does the two-axis risk framework, organized around device activity and the consequence of
relying on an incorrect output, appropriately capture the dimensions most relevant to the risk of a
GenAI-enabled software function? If there are additional dimensions—such as the reversibility of
a resulting action, the availability of downstream safeguards, the time pressure of the deployment
setting, or the traceability of the output (i.e., to primary source materials)—that should be
represented in a risk framework, please describe and provide examples of how they should be
represented.

Context as an added dimension

The two axes do not fully capture the unique risk associated with procedural AI. Therefore, ACS
recommends that FDA add context as a third dimension for procedural care. Context asks whether,
under actual operative conditions, the surgeon can recognize an error, intervene in time, and prevent or
reverse injury. It captures the conditions under which the surgeon is using the tool, including the acuity
of the patient's condition, the urgency of care, the time available to verify and intervene, the reversibility
of the resulting action, workload and fatigue, and intraoperative conditions. The same intraoperative
guidance tool carries different risk in an elective procedure than in an emergent trauma operation on an
unstable patient. Because care for sicker patients is more complex, the same error can cause greater
harm, and urgency shortens the time available to intervene.

Two elements of context deserve particular attention. The first is reversibility. A wrong recommendation
can often be reconsidered before anyone acts on it, but a wrong intraoperative action often cannot be
undone, so two errors of similar immediate severity can carry very different risks. ACS recommends
that FDA define irreversible harm for purposes of the framework, for example by whether the harm can
be corrected within the same episode of care.

The second is downstream safeguards. Much of the framework assumes an error can be caught before it
reaches the patient, but in surgery there may be no interception point. A key measure is the interception
window, which is the time between an AI output and its physical effect, compared with the surgeon's
ability to verify and override it. Time alone, however, is not enough. A surgeon may have five seconds
to intervene yet not recognize that an error is occurring. The relevant construct is therefore not simply
response time but effective human control.

In procedural settings, that window is often very short. Sponsors should specify the range of contexts in
a device's intended use, including urgency, care setting, and the patient's clinical state, and state it in
labeling. They should also demonstrate, under realistic intraoperative conditions including workload,
fatigue, and decision windows, that intended users can recognize errors and intervene before harm
occurs.

Adjustments to the existing axes

Context also explains why the activity axis is an unreliable proxy for risk in surgery. The axis treats
clinician supervision as reducing risk, but in the operating room there is often little or no opportunity to
verify an output before the action takes effect. A tool classified as "health care professional (HCP)-
supervised action-taking" may therefore carry the risk of a fully autonomous function while being held
to a lower evidentiary bar. Conversely, greater autonomy does not always mean greater risk. For
example, a robotic function that autonomously positions a camera may carry less risk than a supervised
function that applies energy near a critical structure. Because the acceptable and achievable level of
supervision varies by procedure and task, FDA must work with surgeon experts to define what
meaningful supervision means for intraoperative functions. The axis also does not capture functions that
constrain the surgeon, such as no-dissection zones, which can fail by blocking a needed action or by
signaling a boundary that is not there.

On the consequences axis, ACS recommends that FDA define the criteria for each consequence level
since a function’s placement on the grid may determine the evidence and postmarket surveillance FDA
requires. ACS welcomes the opportunity to work with FDA on all aspects of the risk-based framework.

Question 2
Center for Devices and Radiological Health (CDRH) seeks input on how to account for the
spectrum of GenAI-enabled informational functions that vary in the degree to which they direct a
user to a particular action (i.e., between “non-directive” and “action-directing”). What
characteristics of an output—such as its wording, specificity, personalization, or context—could
be considered as modifiers of the risk associated with an informational function, after accounting
for the device’s overall functionality and intended use? What additional information would
provide manufacturers with sufficient clarity and predictability around risk assessment for
informational functions while recognizing that directiveness may exist along a continuum rather
than as a binary distinction?

The Discussion Paper's treatment of directiveness focuses on language, that is, how strongly the wording
of an output steers the clinician. Some intraoperative tools direct action through color and salience
instead. For example, an overlay that highlights a dissection plane directs the surgeon even if no text
says to cut there. ACS recommends that FDA judge whether a function is action-directing by its effect
on the user, evaluated through human factors testing, rather than by its wording.

Question 4
CDRH seeks input on whether and how the distinction between generalist and specialist
physicians might inform the assessment of risk for HCP-facing GenAI-enabled functions. Under
what circumstances, if any, might risk be affected when an HCP who lacks the relevant clinical
specialist knowledge receives information that falls within a specialist area of practice? What
device characteristics or safeguards might mitigate such risks?

Safe use of an AI-enabled tool can depend on the user's ability to recognize when the tool is wrong. A
tool that helps a generalist reason about operative versus non-operative management may extend access
to specialist knowledge, but it raises risk if the generalist cannot detect its errors. A user who cannot
recognize an error has no effective control over it.

ACS recommends that sponsors define the user population and level of expertise with which a tool was
validated, test the tool with that population, and state it clearly in labeling. Human factors testing should
measure not only whether the tool is accurate, but whether intended users detect its errors, whether the
tool flags when its output conflicts with the clinician's judgment, and whether users override it
appropriately. Clear labeling, rather than restrictions on use, preserves access for rural and low-volume
providers, who may benefit most from tools that extend specialist knowledge.

Question 6
For GenAI-enabled care escalation functions, CDRH is considering how the evaluation may
account for both under-escalation and over-escalation. How could manufacturers characterize
and weigh these two directions of error, given that they may not be commensurable and that
acceptable trade-offs may vary by clinical context?

ACS agrees that the two directions of error are not commensurable and recommends weighing them by
the time sensitivity of the condition that could be missed. After surgery, failure to escalate a bleed,
anastomotic leak, or wound infection delays care when timing is critical and should be treated as a
severe consequence.

For post-discharge check-in tools, which are already in use, ACS recommends benchmark scenarios
built on standardized postoperative complication definitions, such as those used in the ACS National
Surgical Quality Improvement Program (NSQIP), and on the expected recovery course for each
procedure.
V. A Competency-Based Approach for Premarket Evaluation of GenAI-Enabled Devices

FDA's Discussion Paper describes a premarket approach modeled on how physicians are trained and
evaluated, pairing non-clinical device benchmarking with clinical confirmation. ACS agrees that
benchmarking alone cannot establish how a device will perform in real clinical use. Our
recommendations focus on adapting this framework to account for procedural AI.

Question 7
Is the competency-based approach described above, i.e., device benchmarking followed by clinical
confirmation, a useful and appropriate framework for evaluating GenAI-enabled devices?

Device benchmarking tests whether a device behaves consistently and safely across a defined set of
inputs, which FDA is well positioned to assess. Clinical confirmation asks whether the device performs
as intended for its intended population under real-world or clinically representative conditions, which is
a judgment for clinical experts. This is especially important for GenAI-enabled tools because GenAI
outputs are open-ended and non-deterministic, meaning the same input can produce different outputs,
and there is often no single correct answer to test against.

ACS urges FDA to build a role for clinicians and specialty societies into clinical confirmation as part of
premarket review. Specialty societies can contribute in two ways. First, they can convene qualified
surgeon reviewers and define the clinical standards those reviewers apply. Second, they can help define
the clinical reference frame a tool is measured against, drawing on the clinical guidelines and validated
data they maintain, along with knowledge of relevant patient populations and known failure modes.

Benchmarking Elements

Question 9
Would a benchmarking structure such as the one described above be likely to provide adequate
evidence of clinical knowledge, analytic capabilities, safety behavior, communication, and
generalizability to support a reasonable assurance of safety and effectiveness? Are there elements
that are missing, redundant, or inappropriately categorized? Are there externally developed
standards that could be leveraged?

Many of the ten benchmarking elements in Appendix A assume a conversational, text-based AI-enabled
device. Elements that address capabilities central to procedural AI, such as spatial and anatomic
reasoning, interpretation of video over time, instrument and tissue recognition, or intraoperative timing
are missing.

FDA’s benchmarking elements do not appear to contemplate AI-enabled devices that act autonomously
or semi-autonomously during procedures, such as functions that guide or control a surgical robot. The
agentic element addresses multi-step planning, tool use, and human-oversight checkpoints in software,
but not functions that act on a patient in real time. ACS recommends that FDA recognize that separate
benchmarking elements may be needed for the agentic and autonomous components of AI-enabled
devices used during procedures, and work with surgeons and other stakeholders to develop them.
Clinical Confirmation

Question 11
CDRH is considering that clinical confirmation for a GenAI-enabled device might not require a
prospective clinical study in every case, and has described above a range of approaches of
increasing rigor and patient exposure. How might a sponsor select and justify a confirmation
approach tailored to a device’s intended use and proportionate to the device’s risk profile? Are
there device types or risk profiles for which one or more of these approaches would be insufficient
or inappropriate? How might the anticipated distribution of real-world inputs be taken into
consideration? Are there other methods of clinical confirmation that might help inform the
evaluation of GenAI-enabled devices?

ACS agrees that the rigor of clinical confirmation should be proportionate to the risk of the device, and
that this risk should drive both the approach a sponsor selects and the justification for it. Applying that
principle to procedural AI, however, exposes a gap. The five clinical confirmation approaches in the
Discussion Paper appear to have been developed with software-based functions in mind, and the paper’s
examples of action-taking functions reflect that. They do not address GenAI-enabled functions used
during procedures, which may act on a patient in real time.

For procedural functions capable of immediate, irreversible harm, the less rigorous approaches are
unlikely to be sufficient on their own. Standardized patient interactions have little relevance to
intraoperative tools. Shadow deployment can work for planning and preoperative tools, but not for tools
that act during surgery, because the only way to observe a tool’s effect is to let it act. Retrospective
evaluation on real inputs is also limited for video and spatial data. Methods that are standard in surgical
device evaluation, such as simulation validated for the device’s intended use, cadaveric and ex vivo
evaluation, and structured video review by qualified peer surgeons, can establish a minimum safety
signal before a device is used in patients.

Whatever the method, clinical confirmation of procedural AI should answer two questions. Does the tool
act in a way that makes sense to a qualified surgeon reviewing it? And can the evaluation distinguish
device malfunction from user error? The second question is absent from all five of FDA’s approaches,
and ACS recommends that it be a design requirement for any procedural confirmation method. Clinical
confirmation should also reflect the most demanding conditions in which a device may be used. For
example, a sponsor should not be able to confirm performance only in controlled, elective settings if the
device may be used in emergent, high-acuity, or anatomically complex cases. Where prospective studies
are needed, clinical registries such as ACS NSQIP can support the collection of risk-adjusted outcomes,
but responsibility for generating this evidence rests with the manufacturer.

Question 14
For open-ended device outputs, how should performance comparators and acceptance criteria be
selected? When a panel of qualified clinicians serves as the comparator, how should the applicable
standard (for example, the standard of care versus the performance of a median clinician in
practice) be defined and justified? Should generalist or specialist physicians be used as a
performance standard? When should human-AI team performance, rather than the device
operating alone, serve as the basis for evaluation?
FDA suggests comparing device performance to a panel of qualified clinicians or to a median clinician
in practice. Where a specialist standard of care exists, the comparator should reflect that. Measuring
against a median clinician, however, could set the bar below the care patients would otherwise receive.

Procedural AI is most often used by a surgeon within a workflow, so evaluation should reflect how the
device performs in the hands of its intended users. FDA already takes this approach for radiology AIenabled tools, where sponsors provide both standalone performance data and reader studies comparing
clinicians with and without the device. ACS recommends a similar two-part approach for procedural AI.
First, device-alone performance should be measured through laboratory, simulation, or cadaveric testing,
since these devices cannot be tested on their own in clinical use. Second, performance should be
measured with surgeons who are representative of the intended users, not only experts, whose skill can
mask a weak device.

Patient-reported outcomes (PROs) are also a candidate endpoint FDA does not mention. PRO collection
and validation methods can support acceptance criteria that capture patient-relevant harm and benefit,
particularly for patient-facing functions.

Question 17
CDRH is exploring whether a competency-based approach could be applied to devices that
incorporate a variety of underlying model architectures, such as multimodal vision-language
models and generative or predictive world models. Are there aspects of the described approach
that might be ineffective or inapplicable to these devices? What additional or distinct
considerations might need to be addressed?

The competency approach can apply to multimodal models and world models, but only with the
adjustments described above. These are the types of models expected to power intraoperative video
interpretation and robotic action. The benchmarking elements should account for procedural AI by
incorporating surgical confirmation methods, such as simulation and video review, and by evaluating
these functions as part of the integrated device rather than as standalone software. FDA should build this
category now, before these products arrive, rather than retrofit the framework afterward.

VI. Postmarket Monitoring for GenAI-Enabled Devices

ACS appreciates FDA's recognition that effective postmarket surveillance requires coordinated effort.
Specialty societies can steward clinical validity, health systems can assure safe deployment, clinicians
can identify and report problems in practice, and patients and real-world outcomes show whether a
device delivered benefit. These roles support, but do not replace, the manufacturer's responsibility,
including its postmarket monitoring, under FDA oversight. Participating in surveillance should not shift
device-related risk to clinicians, health systems, or societies, particularly when a device fails despite use
in accordance with its labeling and the standard of care. We provide comments on issues raised in the
Discussion Paper, but some of our recommendations involve other stakeholders and authorities beyond
FDA.

As described above, AI-enabled devices used during surgical procedures raise distinct considerations
because many surgical actions take effect in real time and cannot be undone. ACS is well positioned to
convene surgical stakeholders to develop consensus on evaluation and monitoring standards for
procedural AI and will need dedicated support to sustain these efforts. We see many potential paths for
ongoing collaboration, starting with an initial meeting with FDA to discuss unique considerations
around AI in surgery, and over the longer term, a surgical AI collaborative community.

Because there is some overlap in our responses to questions in Section VI and because the evidence on
GenAI-enabled devices in surgical care is still limited, ACS is responding with overarching
considerations rather than question by question. As mentioned above, GenAI is rapidly developing, and
its use in surgery is still nascent, so these considerations reflect our current thinking and will continue to
evolve as experience and evidence accumulate.

Proposed Considerations for Postmarket Monitoring for GenAI-Enabled Devices

For ACS, the central question for postmarket surveillance is: does the device still deliver the benefit it
was authorized to deliver, across the patients, surgeons, and hospitals it now reaches, and across each
new version of the device?

Consideration 1. Postmarket surveillance should be required in proportion to risk, and for higher-risk
functions it should show that patients continue to benefit
(Responds to FDA Question 19)

ACS recommends that FDA identify which GenAI-enabled functions warrant required postmarket
surveillance, based on their risk and context of use. For the highest-risk functions, including actiontaking procedural functions, FDA must ensure that surveillance includes patient outcomes, including
patient-reported outcomes, not only output benchmarks.

FDA establishes a device’s safety and effectiveness at authorization, but that determination is a
snapshot. For GenAI-enabled devices, whose outputs can vary for the same input, can be plausible but
wrong, and can change when an underlying foundation model is updated, the question becomes whether
the device continues to perform as expected over time. Yet most devices reach the market without a
required postmarket study. Unless FDA imposes conditions of approval, special controls, or a Section
522 order, a manufacturer’s postmarket obligations are limited to general requirements, such as adverse
event reporting, complaint handling, and reporting of corrections and removals. These requirements are
largely reactive and respond to problems once they are identified rather than measure how the device
performs in real-world use. FDA classifies devices by intended use and risk, not by technology, so
GenAI-enabled functions may span device classes.

The approaches in FDA's paper, namely re-benchmarking, sample-based clinician review, and
degradation monitoring, test whether a device's outputs remain as good as they were at authorization.
They are necessary. But output quality is a proxy. ACS's experience in surgical quality measurement
shows that a process can meet every technical standard and still fail to improve outcomes. We recognize
that FDA's standard is reasonable assurance of safety and effectiveness, not proof that a device improves
outcomes relative to alternatives. However, real-world patient outcomes are how continued effectiveness
is confirmed.

As noted in our response to Question 14, the surgeon-AI team is the relevant unit of performance. A
device may perform well alone yet degrade team performance through distraction, automation bias, or
overreliance, or it may help some users more than others. Real-world measurement should cover both
the device alone and the device in the hands of its intended users and should look for differences by
surgeon experience and setting.

Consideration 2. Postmarket monitoring cannot always substitute for premarket evidence
(Responds to FDA Questions 1, 18, and 26)

Accepting greater premarket uncertainty in exchange for postmarket monitoring may not be appropriate
for action-taking functions in procedural settings. ACS recommends that FDA define the minimum
safety evidence required before market entry, weighed against benefit. For higher-risk procedural
functions, FDA should consider a staged-entry approach in which rigorous preclinical confirmation is
followed by controlled initial use at sites that can track outcomes reliably, before broad deployment.

Consistent with our response to Question 1, risk should be judged by the device's activity, the
consequence of an error, and the context in which it is used. That risk should determine both the
premarket evidence required and the postmarket requirements described in Consideration 1. Procedural
AI can act in real time with little or no opportunity to verify an output and can cause harm that cannot be
undone. Postmarket monitoring is essential for detecting drift where errors are recoverable, but it is not a
substitute for premarket evidence where the first error is the injury.

The same concern extends beyond action-taking functions. Directive functions, such as guidance that
identifies a dissection plane, can shape the surgeon’s next action as strongly as a device that acts, and
GenAI heightens this risk because a fluent, confident output can be persuasive even when it is wrong.

Consideration 3. Performance should be measured independently and in the real world
(Responds to FDA Questions 19, 20, and 21)

Monitoring that relies solely on the sponsor is not sufficient. Performance reported by a tool’s own
developer can diverge sharply from real-world results: the Epic Sepsis Model, a predictive tool deployed
outside FDA review, reported an internal area under the curve (AUC) of 0.76 to 0.83, while external
validation found 0.63.1 While the Epic model example is not GenAI, the same risk is even greater for
GenAI-enabled devices, whose outputs vary and whose performance may change after deployment.

For higher-risk functions, FDA should expect sponsors’ postmarket monitoring plans to include
measurement of performance in real-world clinical settings, conducted or verified independently of the
sponsor. Independence does not require excluding the sponsor. It means that review methods are
prespecified and accepted by FDA, adjudicators have documented independence from the sponsor and
any foundation model provider, and FDA can audit the underlying data. FDA can build these
expectations into conditions of approval, special controls, Section 522 orders, and predetermined change
control plans.

Monitoring plans should also define re-evaluation triggers that include changes to an underlying thirdparty foundation model, not only sponsor-initiated updates, and should address how changes to the
surgical platform or workflow will be identified. If sponsors use machine-based supervisory agents to
screen outputs at scale, FDA should expect those agents to be validated, version-controlled, and
Wong A, Otles E, Donnelly JP, et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in
Hospitalized Patients. JAMA Intern Med. 2021;181(8):1065-1070.
periodically audited by humans. They may help direct clinician review, but they should not replace
independent clinician adjudication for higher-risk functions.

Because these tools inform or act on clinical judgment, the clinician belongs in the monitoring loop.
Where FDA requires postmarket monitoring, sponsors’ plans should include FDA's proposed samplebased clinician review in the form of ongoing, structured peer review of real-world outputs. This
approach could be modeled on surgical peer coaching rather than repeating premarket methods such as
simulation. That review should examine complete episodes of care, including risk-adjusted clinical
outcomes and PROs, not only cases flagged for complications.

Clinical registries such as ACS NSQIP can support this measurement. Registries capture risk-adjusted
outcomes but do not record whether an AI tool was used; to compare outcomes with and without the
device, they would need additional data elements, such as device identity, model version, intended use,
and exposure, linked to complications and longitudinal outcomes. Some of this work, such as structured
clinician review within health systems and new registry data elements, falls outside FDA’s authority
over sponsors and would depend on partnerships among sponsors, health systems, and specialty
societies.

For higher-risk functions, sponsors’ postmarket monitoring plans should address real-world
performance, adverse events and near misses, performance across patient subgroups, clinician overrides,
errors the surgeon intercepted or missed, degradation and drift, material model or knowledge changes,
cybersecurity events, patient outcomes, and continuing evidence of clinical benefit, drawing on clinical
registry data and clinician review obtained through partnerships with health systems and specialty
societies.

Consideration 4. Changes must be transparent to the clinicians who use the device and the patients it
affects
(Responds to FDA Questions 22 through 24)

FDA should require manufacturers to notify users of material changes to a device, including changes
that originate with a third-party foundation model. This should include making plain-language
summaries of postmarket performance, known limitations, and material changes publicly available,
whether published by FDA or manufacturers. Patients should also be able to learn what evidence
supports a device used in their care and what remains uncertain.

ACS has consistently supported transparency for AI tools used in clinical care, including in our recent
comments urging the Office of the National Coordinator for Health Information Technology (ONC) to
retain its decision support intervention transparency requirements. Clinicians cannot exercise sound
judgment about a tool whose behavior has changed without their knowledge. Change control should
make material model or knowledge changes visible, with the device's version and technical
dependencies documented.

Responsibility for device-related risk, including risk introduced by changes to an underlying foundation
model, remains with the manufacturer under FDA oversight.
Consideration 5. Reporting should be built for learning
(Responds to FDA Questions 19 and 21)

Deaths, serious injuries, and reportable malfunctions involving GenAI-enabled devices should continue
to be reported to FDA and to manufacturers through existing channels, so that they are captured in the
Manufacturer and User Facility Device Experience (MAUDE) database. Where FDA requires
postmarket monitoring, sponsors’ monitoring should also capture near misses and clinician overrides,
which reveal problems before patients are harmed.

Existing reports, however, rarely contain enough information to determine what role an AI tool played
in an event. Reconstructing an AI-involved adverse event requires data that are not routinely collected
today, including interoperable, time-synchronized logs of AI outputs aligned with clinical records and
device data. Because GenAI outputs can vary for the same input, an event may not be reproducible from
outputs alone. FDA should require sponsors to retain version-stamped logs of AI inputs, outputs, and
model settings, or to build in logging capability so deployment sites can retain them, and to make these
logs accessible to deployment sites and available to FDA on request. FDA should use findings from
event reviews to inform labeling, guidance, and device updates, and should ensure clear mechanisms for
review, recall, or corrective action that cover ongoing updates and adaptive changes.

ACS supports structured, non-punitive pathways for reconstructing AI-involved adverse events, led by
stakeholders such as health systems, patient safety organizations, and specialty societies. These efforts
face barriers outside FDA’s authority. Effective surveillance depends on clinicians being able to report
malfunctions and adverse events confidentially and without penalty when tools are used appropriately,
but the risk of penalty comes mainly from hospital review processes, licensing boards, and malpractice
liability rather than from FDA. FDA can still support learning by treating reports as signals about device
performance.

Consideration 6. Clinical roles should strengthen, not replace, manufacturer accountability
(Responds to FDA Question 21)

Specialty societies can supply methodology, data, and clinical expertise, but regulation must still set the
guardrails. Specialty societies cannot absorb the cost of sponsor-obligated surveillance, and their
involvement cannot substitute for manufacturer accountability. ACS supports postmarket surveillance,
but surveillance should not become a means of shifting responsibility for device performance to
surgeons. When a device fails despite use in accordance with its labeling and the standard of care, that
failure should be addressed as a device performance issue under FDA oversight.

Within that framework, specialty societies can steward clinical validity: reference standards, specialty
evidence, meaningful outcomes, relevant populations, known failure modes, benchmark test cases, and
clinically important performance thresholds. Board-certified surgeons can serve as the independent
clinical adjudicators that Section V of the Discussion Paper calls for but does not identify.

Building on the convening role described in our introduction, ACS asks FDA to work with ACS and
surgical stakeholders to establish a process for developing postmarket assurance for GenAI-enabled
devices used in surgery. Where FDA requires postmarket surveillance, sponsors’ plans should provide
for independent clinical adjudication with safeguards to ensure its independence. FDA should also
establish mechanisms to support specialty-led convening to develop consensus on clinical reference
standards, outcome measurement, and performance thresholds for postmarket evaluation.

Conclusion

ACS appreciates FDA’s attention to GenAI-enabled medical devices. As FDA develops its regulatory
framework, we urge it to account for the distinct considerations around procedural AI, including GenAIenabled procedural functions, incorporate context into its assessment of risk, and build postmarket
monitoring capable of detecting drift in tools whose performance depends on the surgeon, the team, and
the patient. The ACS looks forward to partnering with FDA on these issues as real-world evidence
continues to develop. Please contact Vinita Mujumdar, Chief of Regulatory Affairs, at
vmujumdar@facs.org or Jill Sage, Chief of Quality Affairs, at jsage@facs.org, with questions.

Sincerely,

Patricia L. Turner, MD, MBA, FACS
Executive Director and CEO