Society of American Gastrointestinal and Endoscopic Surgeons (Ankit Sarin, Abhishek Desai)

The Society of American Gastrointestinal and Endoscopic Surgeons (SAGES) submits the attached comments on the FDA discussion paper “Considerations for the Regulation of Generative AI-Enabled Medical Devices” (Docket No. FDA-2026-N-7874).

SAGES represents surgeons who perform minimally invasive, endoscopic, and robotic surgery. Our comments address generative AI functions that direct a physical action on a patient’s tissue, such as those in surgical robots and in energy and stapling devices. The discussion paper considers the AI software component rather than the whole device. For surgical devices, the consequences of an error arise when the device acts on tissue.

The attached comments address 16 of the discussion paper’s 26 questions, including the risk framework, supervision posture, benchmarking and clinical confirmation of functions that act physically, change control, postmarket monitoring, and transparency in public device information. An appendix provides an example framework for characterizing surgical autonomy.

Please see the attached document for the full comments and references.

SAGES Comment on Considerations for the Regulation of Generative AI-Enabled
Medical Devices
10/16/2026

Dockets Management Staff (HFA-305)
Food and Drug Administration
5630 Fishers Lane, Rm. 1061
Rockville, MD 20852

Re: Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled Medical
Devices: Discussion Paper and Request for Feedback

On behalf of the Society of American Gastrointestinal and Endoscopic Surgeons (SAGES), we appreciate
the opportunity to comment on the Food and Drug Administration's (FDA) discussion paper on
generative artificial intelligence (GenAI)-enabled medical devices [1]. SAGES represents surgeons who
perform minimally invasive, endoscopic, and robotic surgery and who use robotic, energy, and stapling
devices in daily practice.

SAGES supports FDA's effort to develop a risk-based, lifecycle approach to GenAI-enabled devices. Our
comments focus on one area the discussion paper raises but does not yet develop: GenAI functions that
direct a physical action on a patient's tissue. The discussion paper considers the GenAI-enabled software
component rather than the whole device. In surgery, an incorrect output matters when the device cuts,
seals, or staples the wrong structure.

Terms used in these comments
• Physical action: movement or energy delivery performed by a device, such as moving a robotic
arm, firing a stapler, or activating an energy device. A physical action may or may not be intended
to affect tissue.
• Supervision posture: what the surgeon must do while a function operates, for example, remain
continuously engaged, approve a specific plan before the device acts, or only monitor and
interrupt.
• Action space: the controls a function can command and the procedural steps, anatomy,
instruments, and tasks to which it applies.
• Closed-loop testing: testing in which the device's own actions change what it observes next, as
they do in clinical use.

Summary of recommendations
SAGES recommends that FDA:
1. Include within its approach GenAI functions whose outputs start, adjust, or gate a physical action
on tissue, including models that generate actions or trajectories (Section 1).
 2. Consider the type of action a function produces as a risk dimension, alongside the two dimensions
already proposed (Section 2).
3. Characterize device autonomy by supervision posture (Section 3).
4. Include physical safety competencies in benchmarking, expect benchmarks to have a defined
context of use and to represent anatomic variation and case complexity, and evaluate functions
that direct physical action with closed-loop testing (Section 4).
5. Specify in advance which modifications require new premarket review, using supervision posture
and action space as the boundaries (Section 5).
6. Maintain the premarket evidence needed to establish safe performance of functions/devices that
act on tissue under supervision alone, even when postmarket monitoring is planned (Section 6).
7. Support intraoperative data capture that allows parties beyond the manufacturer to reconstruct
adverse events, and ask that monitoring plans specify the data to be collected and the triggers for
further review (Section 6).
8. Describe each autonomous function and its supervision posture in public device information, and
draw on surgical societies for clinical expertise in evaluation (Section 7).
The FDA questions addressed are listed by section in the table that follows the signature.

1. Scope (Questions 17 and 26)
SAGES recommends that, when a GenAI function's output starts, adjusts, or gates a physical action on
tissue, FDA evaluate that function within the integrated device.

A model that labels the common bile duct as the cystic duct does not move an instrument, but if that
label permits a clip applier or energy device to fire, the result is a bile duct injury.

The discussion paper notes that an agentic system may meet the device definition when its actions
control another medical device, while limiting its scope to the GenAI-enabled component. For devices
that act on tissue, the software cannot be evaluated apart from the instrument it controls.

GenAI may reach tissue in three ways: through a learned policy that performs surgical tasks on a robot,
through a function that permits or prevents activation of a conventional device, or through a function
that recommends a device setting that the surgeon then applies. In a recent study, a model that
generated language instructions planned and corrected each step while a second learned policy moved
the instruments, completing steps of a cholecystectomy in eight of eight ex vivo gallbladders without
human intervention [2].

For Question 17, SAGES asks FDA to state that models that generate actions or trajectories, including
vision-language-action models and language-conditioned policies, fall within its approach when their
outputs reach tissue.

2. Risk framework (Questions 1, 2, and 8)
SAGES supports the proposed two-axis framework and recommends adding the type of action as a
further risk dimension.

A function's output may be information, a clinical order, a physical action without intended tissue effect
(such as camera or arm positioning), or a physical action with intended tissue effect (such as cutting,
sealing, or stapling). Physical action can occur at any level of supervision, so it does not correspond to a
single point on the proposed activity axis.

Physical action, especially action with intended tissue effect, differs from a clinical order in three ways
the discussion paper already lists as candidate dimensions. It is irreversible: a divided duct or a stapled
ureter cannot be restored. It is time-constrained: a clip is applied in under a second and a staple line
forms in seconds, leaving little time to stop an error once it begins. It has few downstream safeguards: a
medication order usually passes pharmacist verification before it reaches the patient, whereas in the
operating room the surgeon is the principal check.

For Question 2, FDA should consider whether guidance that a robotic system presents to the surgeon
during an operation, such as identified anatomy or a recommended dissection plane, may function as
action-directing, given that the surgeon may act on it within seconds. In a review of 252 laparoscopic
bile duct injuries, 97% were caused primarily by misperception of the biliary anatomy rather than by a
lapse in technical skill [3]. An overlay that misidentifies anatomic structures could create or reinforce
that misperception.

For Question 8, these distinctions can help FDA scale evidentiary expectations, with the highest
expectations reserved for functions that act on tissue under supervision alone, where the surgeon does
not approve each action or plan.

3. Supervision posture and comparators (Questions 1 and 14)
SAGES recommends characterizing device autonomy by supervision posture rather than by the
capability of the underlying technology.

Autonomous functions in surgical devices span the range between the two poles described in the
discussion paper, continuous clinician supervision and fully autonomous action. A practical dividing line
is whether the surgeon approves each specific action or plan before the device carries it out, or only
defines a region or parameters within which the device selects its own targets or actions. This distinction
can be documented in labeling and instructions for use.

For Question 14, the comparator should reflect the supervision posture. Where the surgeon approves
each action or plan, the appropriate comparator is the surgeon and device working together as used in
practice. Where the surgeon only supervises, standalone device performance should also be assessed,
because the surgeon may not be able to intercept an error in time. Sustained monitoring of an
automated action is itself demanding, and a surgeon's attention may be divided among the operative
field, the assistant, and the anesthesia team. Where FDA uses a panel of qualified clinicians as the
comparator, the panel should consist of surgeons experienced in performing the procedure, including
with the relevant robotic platform.
The presence of a supervising surgeon should not be assumed to mitigate risk. Evaluation should
establish whether surgeons can identify erroneous outputs or actions of the function, intervene before
harm occurs, and safely assume control, accounting for workload and distraction.

4. Benchmarking and clinical confirmation (Questions 7, 9, 11, 17, and 26)
SAGES recommends that benchmarking include physical safety competencies and that functions
directing physical action undergo closed-loop testing.

SAGES supports FDA's focus on evaluating the final device as configured. The internal reasoning of a
GenAI model often cannot be inspected, but its outputs and the supervision posture the function /
device requires can be observed. The benchmarking elements in the discussion paper describe agentic
competencies primarily in software terms. For functions that direct physical action, FDA could also
consider whether the action remains within the approved plan when the operative field changes, stops
promptly when the surgeon intervenes, and leaves tissue and instruments in a safe state afterward.

SAGES recommends that benchmarks used to support regulatory evaluation have a defined context of
use and demonstrate clinical relevance. Surgical benchmarks should represent variations in anatomy
and case complexity.

For Question 11, retrospective and shadow evaluation may be sufficient for informational functions.
These methods cannot show how a GenAI system performs once its actions begin to change what it
observes, allowing small errors to compound. Functions that direct physical action therefore require
closed-loop testing. Bench, simulation, and ex vivo testing can support these assessments, provided
their limitations are stated. Ex vivo tissue does not bleed, move with respiration, or show the
inflammation, adhesions, and anatomic variation that make an operation difficult.

High average performance should not compensate for failures such as performing the wrong action,
acting on the wrong structure, failing to stop, or failing to recover safely. The need for prospective
clinical confirmation should reflect the severity of potential harm, the effectiveness of human
intervention, and the extent to which simulated or ex vivo testing reproduces clinical conditions.

5. Change control (Questions 22, 23, and 24)
SAGES recommends that FDA specify in advance which modifications may proceed without new
premarket review, using supervision posture and action space as the boundaries.

The level of autonomy of a device or function can change without any change to its hardware or model,
for example when a requirement for continuous surgeon engagement is removed. A modification that
changes supervision posture should be treated as one that may affect safety and effectiveness. If such a
modification is managed through a predetermined change control plan, the plan should include human
factors validation in the new supervision posture and closed-loop performance testing.

Model updates, including those supplied by a third party, could proceed under a predetermined change
control plan when the supervision posture and action space are unchanged and the device meets
prespecified acceptance criteria on closed-loop retesting of the competencies described in Section 4. An
update that expands the action space, for example to new procedural steps, anatomy, instruments, or
tasks, or that does not meet those criteria, would warrant new review. Retesting should be most
extensive for functions that act on tissue under supervision alone.

6. Postmarket monitoring (Questions 18, 19, and 21)
SAGES recommends that FDA maintain the premarket evidence needed to establish safe performance
of functions that act on tissue under supervision alone, even when postmarket monitoring is planned.

Errors by these functions/devices can cause irreversible injury before a surgeon is able to intervene, so
postmarket detection cannot prevent the initial injuries. Many of these injuries, such as a delayed
thermal injury to bowel or a bile leak, are recognized only after the operation and are then attributed to
surgical technique rather than to the device. Without an intraoperative record, the device's contribution
to the injury may never be identified.

Effective monitoring also depends on intraoperative data. Device logs are generally held by
manufacturers, and adverse event reports are not designed to capture AI-specific failures. SAGES
recommends that functions acting on tissue record their inputs, their outputs, and each surgeon
intervention in a form that allows an adverse event to be reconstructed and investigated by parties
beyond the manufacturer, including FDA and the investigating hospital. Multiport intraoperative data
capture has already been used to characterize errors and events during laparoscopic surgery [4].
Surgeon video review can complement outcome measures in assessing whether a function performs as
intended.

Postmarket monitoring data should capture device and model versions, exposure denominators,
surgeon overrides and takeovers, near misses, and adverse events. Monitoring plans should specify
triggers for additional review, revalidation, corrective action, or restriction of use. Professional societies
such as SAGES could support common event definitions and clinical adjudication, while manufacturers
retain responsibility for monitoring, investigation, reporting, and corrective action.

7. Transparency and professional societies (Questions 16 and 21)
SAGES recommends that FDA consider describing, in labeling and public decision summaries, what
each autonomous function does, the type of action it produces, and its supervision posture.

A 2024 systematic review of 49 FDA-cleared surgical robots found that only 2 were recognized by FDA as
having machine learning-enabled capabilities [5]. In applying the framework in the Appendix to cleared
devices, members of the SAGES Robotics Committee found that the required supervision posture often
could not be confirmed from public FDA records. Consistent with the agency's existing transparency
principles for machine learning-enabled devices [6], this information is necessary to support safe
adoption. Hospitals need it to set privileging criteria and training requirements for each function, and
surgeons need it to know what they are expected to do while the function operates.

SAGES would welcome the opportunity to discuss with FDA how surgical societies might contribute to
independent evaluation, including surgeon adjudication of device performance and the development of
surgical video benchmarks, with manufacturers retaining responsibility for their devices. Additional
collaboration could include developing surgical human factors assessment protocols, a common
postmarket event dataset, and a clinical adjudication methodology. SAGES developed and validated the
Fundamentals of Laparoscopic Surgery program [7], which the American Board of Surgery requires for
certification in general surgery; published consensus recommendations on the annotation of surgical
video and on surgical video data use and structure [8, 9]; and organized the Critical View of Safety
Challenge, a collaboration of 54 institutions in 24 countries that produced 1,000 videos annotated by 20
surgical experts as a benchmark for AI models [10].

Thank you for considering these comments. SAGES would be glad to serve as a resource as the agency's
approach develops. Dr. Ankit Sarin and Dr. Abhishek Desai of the SAGES Robotics Committee prepared
this submission. For questions, please contact Dr. Sarin (axsarin@health.ucdavis.edu) or Dr. Desai
(abhishek@headclone.com).

Sincerely,

Society of American Gastrointestinal and Endoscopic Surgeons

FDA questions addressed
Question Topic in the discussion paper Section of these comments

1 Two-axis risk framework and additional dimensions 2, 3

2 Directive versus non-directive informational functions 2

7 Competency-based approach to premarket evaluation 4

8 Risk framework and level of evidence 2

9 Benchmarking elements 4

11 Clinical confirmation approaches 4

14 Comparators and acceptance criteria 3

16 Role of independent third parties 7

17 Other model architectures 1, 4

18 Premarket uncertainty and reliance on postmarket monitoring 6

19 Postmarket performance evaluation 6

21 Roles of clinicians and professional societies 6, 7

22 Scaling re-evaluation to the nature of a modification 5

23 Predetermined change control plans 5

24 Changes to third-party foundation models 5

26 Agentic GenAI-enabled devices 1, 4
7
Appendix: Example framework for characterizing surgical autonomy
The following framework, SURGE (Surgical Understanding, Responsiveness, Guidance and Execution),
developed by the Autonomous Surgery Working Group of the SAGES Robotics Committee, is included as
one illustration of how the principles above can be applied. It has not yet been formally validated. Each
autonomous function is characterized on one of three axes, from Level 0 (no autonomous capability) to
Level 5 (no surgeon in the loop). At Level 3 and below, the surgeon approves or engages with each
output or action. At Level 4, the surgeon supervises and can interrupt.

The examples in the table are cleared devices that, to our knowledge, do not use GenAI. Because the
framework describes what a function does and what the surgeon must do, it applies equally when a
function is GenAI-enabled.

The framework separates systems that a single-level scale groups together. Lee et al. place TSolution
One and ARTAS iX at the same level [5]. On the three axes of this framework, the working group scored
TSolution One at 2, 0, and 3, and ARTAS iX at 4, 0, and 4. TSolution One carries out a plan the surgeon
has approved, whereas ARTAS iX selects its own targets within a region the surgeon defines.

Level Surgical cognition Physical action without intended Physical action with intended tissue
tissue effect effect

0 No autonomous capability No autonomous capability No autonomous capability

1 Display or measurement only (e.g., Passive assistance (e.g., endoscope Passive assistance (e.g., tremor
fluorescence overlay) stabilization) filtration, motion scaling)

2 Interpretation (e.g., identification of Bounded action on explicit surgeon Action requiring approval and
anatomic structures) command (e.g., guided instrument continuous engagement (e.g.,
exchange) powered stapling)

3 Planning or decision support Adaptive action following a surgeon Action requiring approval, then
requiring approval (e.g., signal (e.g., eye-tracking camera proceeding while the surgeon
intraoperative ablation plan) control) monitors (e.g., active bone milling)

4 Decisions made under supervision Action started by the system under Adaptive tissue manipulation under
only (e.g., selection of individual supervision only (no cleared supervision only (e.g., automated
targets in a defined region) example) follicular unit extraction)

5 Complete autonomy Complete autonomy Complete autonomy

References
1. U.S. Food and Drug Administration. Considerations for the Regulation of Generative AI-Enabled
Medical Devices: Discussion Paper and Request for Feedback. August 2026.
2. Kim JW, Chen JT, Hansen P, et al. SRT-H: a hierarchical framework for autonomous surgery via
language-conditioned imitation learning. Sci Robot. 2025;10(104):eadt5254.
doi:10.1126/scirobotics.adt5254.
3. Way LW, Stewart L, Gantert W, et al. Causes and prevention of laparoscopic bile duct injuries:
analysis of 252 cases from a human factors and cognitive psychology perspective. Ann Surg.
2003;237(4):460-469. doi:10.1097/01.SLA.0000060680.92690.E9.
4. Jung JJ, Jüni P, Lebovic G, Grantcharov T. First-year analysis of the Operating Room Black Box
study. Ann Surg. 2020;271(1):122-127. doi:10.1097/SLA.0000000000002863.
5. Lee A, Baker TS, Bederson JB, Rapoport BI. Levels of autonomy in FDA-cleared surgical robots: a
systematic review. NPJ Digit Med. 2024;7:103. doi:10.1038/s41746-024-01102-y.
6. U.S. Food and Drug Administration, Health Canada, Medicines and Healthcare products Regulatory
Agency. Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles. June
2024.
7. Peters JH, Fried GM, Swanstrom LL, et al. Development and validation of a comprehensive
program of education and assessment of the basic fundamentals of laparoscopic surgery. Surgery.
2004;135(1):21-27. doi:10.1016/s0039-6060(03)00156-9.
8. Meireles OR, Rosman G, Altieri MS, et al. SAGES consensus recommendations on an annotation
framework for surgical video. Surg Endosc. 2021;35(9):4918-4929. doi:10.1007/s00464-02108578-9.
9. Eckhoff JA, Rosman G, Altieri MS, et al. SAGES consensus recommendations on surgical video data
use, structure, and exploration (for research in artificial intelligence, clinical quality improvement,
and surgical education). Surg Endosc. 2023;37(11):8690-8707. doi:10.1007/s00464-023-10288-3.
10. Alapatt D, Eckhoff J, Lyu Z, et al. The SAGES Critical View of Safety Challenge: a global benchmark
for AI-assisted surgical quality assessment. arXiv. 2025. arXiv:2509.17100.