Cicero Institute (Adam Meier)

Please see the attached formal comment letter submitted on behalf of the Cicero Institute by Adam Meier, Director of Health Policy, regarding Docket No. FDA-2026-N-7874.

CICERO INSTITUTE
2112 Rio Grande St., Austin, TX 78705

October 19, 2026

Dockets Management Staff (HFA-305)
U.S. Food and Drug Administration
5630 Fishers Lane, Room 1061
Rockville, MD 20852
Submitted electronically via Regulations.gov
Re: Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled Medical
Devices: Discussion Paper and Request for Feedback (August 2026)1

Dear Digital Health Center of Excellence:

Introduction and Summary
On behalf of the Cicero Institute, I appreciate the opportunity to comment on the Center for Devices
and Radiological Health’s (CDRH) discussion paper on generative AI (GenAI)-enabled medical devices.
We welcome the Center’s decision to seek stakeholder input before policy is set, and we share its stated
goal of a nimble, least burdensome regulatory approach that gives patients timely access to safe and
effective devices.2

The Cicero Institute is a nonpartisan public policy organization that partners with state leaders across
the country. On September 25, 2026, we published Version 2.1 of our model state legislation, the
Clinical AI Services Act,3 which creates a state-licensed pathway for clinical AI that performs patientspecific functions constituting the practice of medicine or other licensed clinical practice. Many of the
discussion paper’s core ideas track choices we made in that bill. We offer these comments to highlight
that convergence, to share what we learned designing a service-based framework, and to suggest how
federal and state oversight can fit together.

Clinical AI raises two related questions: whether the underlying technology is safe and effective for its
intended use, and whether the clinical service delivering it is competent, accountable, and appropriately
supervised over time. FDA’s device framework is built to answer the first. State oversight of clinical
practice is built to answer the second. FDA need not choose between product regulation and service
regulation; the two can reinforce one another.

Our comments take the current statutory landscape as given. Cicero would welcome congressional
action that clarifies federal roles and reduces unnecessary inconsistency across states where
appropriate, and a few of our recommendations note where additional authority could help. We have
not, however, assumed any significant change in federal authority. Our aim is a framework that works
within the system as it exists today, drawing on FDA’s existing authorities, state authority over clinical
practice, and the tools already available to both.

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 1
Our recommendations, in brief:
1. Keep and formalize the tiered, risk-based approach. The two-axis framework is a significant
step forward and closely parallels the role-by-condition licensure grid in our model bill.
2. Regulate the clinical service and its outcomes, not only the product. The Clinical Laboratory
Improvement Amendments (CLIA) framework for laboratory services, including laboratorydeveloped tests, offers a proven structural model: oversight of accountable entities through
ongoing quality systems, performance verification, and independent accreditation. Clinical AI
would pair that model with explicit clinical competency and real-world outcome standards.
3. Build a complementary federal-state framework. Where FDA authority is limited or uncertain,
FDA can help states integrate clinical AI into their frameworks for the practice of medicine and
other licensed clinical professions through shared benchmarks, reciprocity, and data sharing.
4. Create a federal recognition pathway for independent clinical validators. Qualified third
parties, paid by developers and overseen by FDA, could perform clinical validation the way
accrediting bodies already verify quality across health care. Utah’s healthcare AI sandbox now
uses this model.
5. Recognize committee-approved protocols and standing orders as a delegation model for AI.
Medicine already delegates defined actions to nurses, pharmacists, and electronic order sets
under protocols approved by accountable clinical leaders. AI that follows such protocols
presents a bounded, auditable form of action-taking.
6. Carry the paper’s best ideas through consistently. The competency-based approach, humanperformance comparators, and greater reliance on postmarket evidence point in the right
direction. A few areas would benefit from clarification so that these principles are applied
evenly. We also encourage HHS and FDA to establish a dedicated center for clinical AI.

I. The Tiered, Risk-Based Approach Is the Right Direction
The two-axis framework in the paper’s Section IV is a move in the right direction and the paper’s most
useful contribution.4 Organizing risk around what a function does and what happens if it is wrong is
intuitive, technology-neutral, and scalable. It also mirrors how clinical practice is already governed:
clinicians are supervised more closely, and decisions are scrutinized more closely, as autonomy and
acuity rise.

When we drafted the Clinical AI Services Act, we arrived at a closely parallel structure. The bill pairs four
“AI operational roles” with defined clinical condition categories,5 and licensure obligations attach only as
both autonomy and acuity increase.6 The crosswalk below shows how closely the two frameworks align.

FDA activity axis (Fig. 1) Cicero operational role Cicero treatment by clinical condition
(§ 2(c))
Informational: non- Informational AI Exempt for all conditions
directive

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 2
 FDA activity axis (Fig. 1) Cicero operational role Cicero treatment by clinical condition
(§ 2(c))
Informational: action- Advisory AI Exempt for preventive and chronic/non-critical conditions;
directing license (Modifier L1) for critical or time-sensitive conditions
when relied on to substitute for clinical judgment and
outside the clinical decision support safe harbor
Action-taking: HCP- Supervised Autonomous License (Modifier L2) for chronic, non-critical, critical, and
supervised AI time-sensitive conditions; preventive exempt unless issuing
patient-specific orders
Action-taking: fully Fully Autonomous AI License (Modifier L3) for all conditions
autonomous

FDA’s consequences axis (limited, moderate, severe) maps broadly to our condition categories
(preventive; chronic and non-critical; critical and time-sensitive). The two are not identical. FDA’s axis
measures the consequence of relying on an incorrect output, while our bill uses the patient’s condition
category as a practical proxy for acuity. FDA’s formulation captures risks a condition category alone can
miss, such as an incorrect insulin dose change for a chronic condition. Our bill captures that risk on the
autonomy side: an AI that changes a medication dose is supervised or fully autonomous AI and requires
a license regardless of condition category. Like FDA, we concluded that the activity axis sets a
presumption of risk and the consequences axis moderates it. That is why the paper’s examples of
hydrocortisone versus insulin, and strep throat antibiotics versus thrombolytics, track our own thinking.7

Three design choices from our bill may be useful as FDA refines the framework:
• Assess foreseeable presentations up front, and individual encounters by what was known at
the time. FDA’s prospective risk assessment should account for the reasonably foreseeable
presentations and consequences within a device’s intended use, including serious conditions
that can initially appear benign. When an individual encounter is reviewed after deployment,
however, the question should be whether the AI responded appropriately to the information
available at the time, not whether a later diagnosis proved more serious. Our bill applies this
second principle, classifying conditions based on the patient’s presentation and the clinical
information reasonably available at the time of service.8 A function that acts reasonably on a
routine sore throat should not be faulted because the patient later proves to have a peritonsillar
abscess, provided its design accounted for that foreseeable possibility.
• Fold time-sensitivity and reversibility into the consequences axis. Question 1 asks whether to
add dimensions such as reversibility and time pressure.9 Our highest condition tier is defined by
the risk of serious irreversible harm or death without prompt intervention, and by care whose
effectiveness depends on timely initiation.10 Building these factors into the definition of the
consequences axis, rather than adding a third axis, keeps the framework simple while capturing
the risk FDA describes.
• Turn the gradient into clear tiers with explicit off-ramps. Figure 1 shows a continuous gradient,
which is conceptually right but hard to apply. Our bill converts the grid into discrete cells, each
with a defined outcome: exempt, or licensed at a stated modifier. FDA could similarly identify

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 3
 cells where it intends enforcement discretion, cells eligible for streamlined review, and cells
requiring full competency-based evaluation. This would give manufacturers the predictability
that Question 2 seeks.

II. Regulate the Service and Its Outcomes: Lessons from CLIA
The paper’s own analysis points toward oversight of a clinical service, not only a product. GenAI-enabled
functions accept open-ended inputs, change over time, and perform tasks that resemble clinical
judgment, and the paper concludes that it may be unreasonable to evaluate every conceivable input
before deployment.11 Those same features are why states govern physicians, nurses, pharmacists, and
other licensed clinicians through licensure, standards of care, and ongoing accountability rather than
one-time product clearance.

The federal precedent already exists. Since 1988, CLIA, administered by the Centers for Medicare &
Medicaid Services (CMS), has regulated clinical laboratory testing as a service.12 CLIA does not pre-clear
each test. It certifies the laboratory, sets personnel and quality-system requirements, requires
laboratories to establish performance specifications for tests they develop or modify,13 and checks
ongoing accuracy through proficiency testing, other accuracy verification, and inspection.14 Laboratorydeveloped tests (LDTs) have been overseen within this framework for decades, with FDA historically
exercising enforcement discretion, and laboratories refine those tests continuously. CLIA is a structural
precedent, not a complete one. It addresses analytical reliability, and CMS has stated that CLIA does not
address the clinical validity of any test.15 Clinical AI would need CLIA’s institutional features plus explicit
clinical competency, patient safety, and real-world outcome standards, which the paper’s competencybased approach and our model bill both supply. Congress is weighing a similar combination for
laboratories: the Enhancing CLIA Act of 2026 (H.R. 8890), introduced in May, would keep laboratorydeveloped tests under an updated CLIA while requiring a reasonable assurance of analytical and clinical
validity and allowing voluntary third-party validation.16

In March 2025, the U.S. District Court for the Eastern District of Texas vacated FDA’s 2024 LDT rule,
concluding that Congress established CLIA as the framework governing laboratory testing services.17 FDA
did not appeal and rescinded the rule in September 2025.18 We raise this not to revisit that question but
because it illustrates a durable principle: when a clinical tool is designed, validated, and used by an
accountable professional entity, oversight of the service can work well alongside, or in some settings in
place of, product clearance.

CLIA’s design maps naturally onto the challenges the paper identifies, including through accrediting
organizations with deemed status and state programs approved for exemption:19
CLIA feature GenAI challenge in the Discussion Parallel in the Model Act
Paper
Entity certification, not test-by- Range of inputs too large to test License held by an accountable entity
test clearance exhaustively (Sec. V) (§ 4)
Laboratory establishes its own test Sponsor-defined benchmarks and Validated competency defines scope of
performance specifications acceptance criteria (Sec. V.B.2) practice (§ 8(a))

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 4
 CLIA feature GenAI challenge in the Discussion Parallel in the Model Act
Paper
Recurring proficiency testing Periodic re-benchmarking (Sec. VI.A) Annual performance reports;
suspension for model drift (§ 7(g)(4))
Accrediting organizations with Role for independent third parties Recognition of third-party
deemed status (Sec. V.D.2) accreditation (§ 7(g)(3))
Approved state exemption Shared ecosystem responsibility (Sec. Federal and interstate reciprocity
programs VI.B) (§ 4(f))

Our model bill was built on these principles. Its Class A license expressly covers clinical AI services that
operate in a manner analogous to LDTs and do not rely on FDA clearance as the basis for their lawful
clinical use.20 It licenses the entity that delivers the service; requires a medical director for supervised
and autonomous services, malpractice coverage, and a surety bond;21 requires logging of clinical outputs
and system configurations;22 and measures performance against a reasonably prudent human provider
in the same or similar specialty.23

Completing the clinician analogy. The paper draws on proposals to evaluate GenAI the way clinicians
are credentialed. As the paper notes in discussing Freyer et al., clinicians also face professional, legal,
and reputational consequences when they fall short.24 Premarket competency testing borrows the first
half of the clinician model. Licensure, the standard of care, malpractice liability, and professional
discipline supply the second half, and they sit largely in state law governing the practice of medicine and
other licensed clinical professions. A service-based framework lets the two halves work together.

We recognize the paper expressly sets aside questions of legal authority.25 None of this requires FDA to
claim jurisdiction over clinical services. FDA can borrow CLIA’s strongest structural features within its
existing device authorities, through programs such as ASCA, MDDT, and predetermined change control
plans, while direct regulation of professional clinical services remains with state licensing authorities.26
Within that limit, we encourage FDA to consider three steps:
1. Identify categories of GenAI clinical functions, such as those designed, validated, and delivered
by a single accountable provider entity, that may be better suited to service-level oversight, and
coordinate with CMS on that boundary.
2. Borrow CLIA’s design features within FDA’s own framework where appropriate: entity-level
competency certification, recurring proficiency-style testing, and deemed status for qualified
accreditors (see Section IV).
3. Where service-level oversight falls outside FDA authority, support the state frameworks
described in Section III rather than leaving a gap.

III. A Complementary Federal-State Framework
FDA and the states each hold part of the answer, and clinical AI will be safest where the two fit together.
The paper notes that many software functions do not meet the device definition or are subject to
enforcement discretion,27 and that some agentic functions, such as care coordination, may not be the
focus of FDA’s device oversight.28 Those services need not go unregulated. When they perform patientCicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 5
specific clinical work, they may implicate state laws governing the practice of medicine or other licensed
clinical practice, including nursing, pharmacy, and behavioral health, and states can bring them into a
clear licensing framework through legislation. Federal law recognizes the practitioner’s role: section
1006 of the FD&C Act preserves practitioners’ authority to use legally marketed devices within a
legitimate practitioner-patient relationship, while preserving FDA’s authority over the sale, distribution,
and labeling of devices.29

To be clear, Cicero does not contend that state licensure changes whether a particular AI function is a
medical device under federal law. An AI service can be delivered through a professional practice and still
contain a regulated device function. In those cases, state professional oversight and federal device
oversight would operate concurrently. Our goal is for the two to reinforce one another, minimize
duplicative requirements, and share common competency and postmarket monitoring standards.

Our model bill was written to complement federal oversight, not compete with it. It creates a Class B
license with automatic eligibility for products holding FDA clearance for the relevant use;30 directs the
state to align its benchmarks and postmarket monitoring with federal plans to the maximum extent
practicable;31 and does not authorize conduct prohibited by federal law or that would create an
unavoidable conflict with the FD&C Act.32 We see several concrete ways FDA could build on that
foundation:
• A shared benchmark reference. The benchmarking elements in Appendix A (S.1 through A.1)
could become a common vocabulary that states adopt by reference.33 If FDA recognized
consensus standards or qualified tools through the Medical Device Development Tool (MDDT)
program, states could point to the same assets, and sponsors could test once for both regimes.34
• State sandboxes as a source of clinical confirmation evidence. The paper’s Section V.C
describes clinician adjudication of real cases, shadow deployment, and real-world data.35 Our
bill’s two-year provisional license allows phased deployment with volume caps, geographic
limits, and physician review of a defined number of interactions before human-in-the-loop
requirements are relaxed.36 FDA could state that well-designed state sandbox data may support
clinical confirmation and postmarket evidence. Our bill lets sponsors elect IRB review specifically
to support future federal applications, and it preserves all applicable federal human-subjects
obligations.37
• Harmonized safety reporting. Our bill defines adverse events to include inappropriate triage
and failure to escalate care, mirroring the paper’s attention to under-escalation.38 Aligning state
and federal event definitions, and creating channels to share signals, would strengthen the
shared ecosystem the paper describes in Section VI.B without diffusing manufacturer
accountability.39
• Technical assistance to states. FDA’s scientific expertise in AI evaluation is a national resource.
Convening state licensing officials and clinical licensing boards to share evaluation methods
would help states integrate clinical AI into their existing professional practice frameworks
consistently and reduce the risk of incompatible state rules.
• Clarity at the boundary. A short statement distinguishing three categories would help sponsors,
states, and payers plan with confidence: clinical AI services that contain no device function;

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 6
 regulated device functions delivered as part of a clinical service; and mixed systems in which
state oversight of clinical practice and federal device oversight apply concurrently.

IV. A Federal Recognition Pathway for Independent Clinical Validators
We recommend that FDA establish a process to recognize qualified, independent organizations that
perform clinical validation of GenAI-enabled devices, with validation paid for by the developer and FDA
setting a federal floor and ceiling for the work and auditing the validators. Question 16 asks what roles
independent third parties might play;40 we believe clinical validation is the most valuable one.

Accreditation is how American health care already verifies quality at scale. Joint Commission
accreditation can confer Medicare deemed status on hospitals,41 CARF accreditation is widely relied
upon for behavioral health and rehabilitation providers, and CLIA relies on approved accrediting
organizations for laboratories (Section II). In each case, the organization being reviewed pays for the
review, while government sets the bar and oversees the accreditors. FDA has its own precedents: the
Accredited Persons program for third-party review of premarket notifications for eligible devices,42 and
the ASCA and MDDT programs the paper describes.43

The market and the states are already moving. On October 5, 2026, Utah’s Office of Artificial
Intelligence Policy announced agreements with six independent evaluators for its healthcare AI sandbox:
the Coalition for Health AI (CHAI), Clarion AI Partners, Glacis Technologies, mpathic AI, Vega Health, and
the Stanford Clinical Excellence Research Center.44 Utah’s framework distinguishes three roles: clinical
and technical evaluation (mpathic, Stanford CERC, and Vega Health); runtime attestation that specified
safety checks ran (Clarion AI Partners and Glacis Technologies); and an evaluator network through which
CHAI qualifies partner evaluators rather than evaluating pilots itself. The company being evaluated pays
the evaluator directly. Evaluators must certify that each evaluation was independent, report any patient
safety concern to the state, and allow the state to audit their work. No evaluator has an exclusive role,
and the state retains every decision about whether a pilot starts, continues, or stops.45 Utah’s design
closely mirrors the elements we propose below. It is also likely that health systems are already engaging
outside firms to validate AI tools before deployment, a practice we expect to grow. In June 2026, the
Joint Commission launched a voluntary Responsible Use of AI in Healthcare certification whose
standards include monitoring, evaluating, and validating AI safety and performance; that program
certifies organizations, not individual tools.46 What is missing is a federal framework that lets this market
scale across states without creating a single gatekeeper.

A recognition pathway could include the following elements:
• A federal floor and ceiling, with competition in between. FDA would set a floor: minimum
qualifications, independence requirements, and evidence expectations scaled to a function’s
place on the two-axis grid and tied to the benchmarking elements in Appendix A and the clinical
confirmation methods in Section V.C of the paper. FDA would also set a ceiling: limits on what
validators, or regulators relying on them, may require beyond what a function’s risk tier
warrants, so that recognition does not become a back door to new barriers or escalating

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 7
 compliance costs. Between those bounds, validators would compete on methods, speed, and
price rather than converge on a single prescribed approach.
• The developer pays. As with CLIA accreditation, Joint Commission surveys, third-party
premarket review, and Utah’s evaluator network, the cost falls on the company seeking
validation rather than on taxpayers, and review capacity can grow with demand rather than
with FDA staffing.
• Independence safeguards. Validators would have no ownership or financial ties to the
developer beyond the validation fee, would disclose conflicts, and could not validate a product
they helped design. This extends the paper’s own principle that expert adjudicators be
structurally independent from the sponsor and from any foundation model developer.47
• Evidence, not delegated approval. Recognized validators would generate and certify evidence:
competency assessments, clinical confirmation, re-benchmarking after modifications, and realworld monitoring. FDA would continue to make marketing decisions, relying on that evidence as
appropriate. Existing programs such as ASCA, MDDT, and third-party premarket review offer
partial models; broader reliance on validator determinations may require additional authority,
which FDA could identify for Congress.
• FDA oversight of validators. FDA would audit a sample of validations, review validator
performance over time, and suspend or withdraw recognition when warranted, much as CMS
oversees CLIA accrediting organizations.
• An open, competitive market. Multiple recognized validators, published criteria, no exclusivity,
and no requirement to join any particular consortium would address the paper’s concern that
third-party roles not limit competition or innovation.
• Test once, use many times. A recognized validation could support FDA review, including rebenchmarking after modifications; satisfy state requirements; and inform health system
procurement, sparing developers and providers duplicative testing.
Our model bill already lets a state recognize independent third-party accreditation or certification as
evidence that a licensee meets, exceeds, or maintains its benchmarks.48 A federal floor and ceiling would
give states, health systems, and payers a trusted basis for relying on recognized validators, and would let
developers validate once against a common baseline rather than navigate fifty separate regimes.

V. Committee-Approved Protocols and Standing Orders as a Delegation Model
We recommend that FDA recognize the execution of clinical actions under protocols approved by an
accountable clinical committee as a risk-mitigating characteristic and a more bounded form of actiontaking, and that evaluation of such functions focus on whether the AI follows the protocol faithfully.

Medicine already delegates this way. Medicare’s hospital Conditions of Participation permit preprinted
and electronic standing orders, order sets, and protocols when the medical staff and nursing and
pharmacy leadership have approved them, they are consistent with nationally recognized and evidencebased guidelines, they are reviewed periodically for continued usefulness and safety, and resulting

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 8
orders are promptly authenticated in the medical record.49 Nurses, pharmacists acting under stateauthorized collaborative practice agreements, and other clinicians act on such protocols every day once
defined criteria are met, and electronic health records already surface or carry out electronic order sets
when conditions are satisfied.

AI can follow the same model. Clinical judgment is exercised in advance by an accountable committee
that defines inclusion criteria, exclusion criteria, permitted actions, reassessment intervals, and
escalation triggers. The AI’s task is to determine whether the criteria are met, act within those bounds,
and escalate when they are not. This turns the open-ended behavior the paper is concerned about into a
bounded and auditable task.50 The approach fits the paper’s framework in several ways:
• Placement on the activity axis. Protocol-bounded action sits between HCP-supervised and fully
autonomous action. Human judgment is exercised before the encounter rather than case by
case. FDA could treat this as its own position on the activity axis or as a modifier that lowers
presumptive risk.
• A clearer reference standard. Because the protocol defines the correct action, protocol fidelity
can be measured directly. This eases the comparator problem the paper identifies for openended outputs, where no single correct response exists.51 Testing would focus on correctly
recognizing when criteria are and are not met, honoring exclusions, and escalating
appropriately, which aligns with benchmarking elements S.1, S.2, and A.1, including compliance
with human-oversight checkpoints.52
• A clean division of responsibility. The clinical committee owns the content of the protocol,
which is governed by the Conditions of Participation and state practice law. FDA and recognized
validators evaluate whether the AI follows protocols reliably. Where protocol updates remain
within the device’s authorized intended use and validated operating boundaries, FDA could offer
streamlined change control, for example through a predetermined change control plan. Updates
that change the intended use, patient population, or safety profile would still warrant review.
• Built-in reassessment. Well-designed protocols include reassessment intervals and escalation
triggers. For example, a medication renewal protocol should require periodic clinical
reassessment and monitoring. These safeguards keep protocol-bounded AI tied to ongoing
clinical oversight.
A hospital-approved protocol does not by itself authorize AI to exercise powers ordinarily delegated to
licensed professionals; that depends on state law and on how the technology is authorized. Our model
bill supplies that authorization and reflects this approach. It directs the state to develop frameworks for
delegated, collaborative practice, and supervision agreements;53 its Modifier L2 covers supervised AI
operating under human oversight or a collaborative practice agreement;54 and each supervised or
autonomous licensee must have a medical director responsible for clinical scope, safety protocols, and
escalation pathways.55 This approach may also help with Question 26, since protocol-bounded execution
with defined human-oversight checkpoints offers a practical model for agentic systems that take multistep actions.56

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 9
VI. Additional Elements We Support
Beyond the risk framework, several of the paper’s choices align with our own approach.
• Regulating uses, not technology. The paper states that FDA does not regulate GenAI as such,
only devices.57 Our bill takes the same position: it does not regulate AI as a general technology,
only clinical services delivered to patients.58
• Evaluating the system as deployed. The competency-based approach would evaluate the final
user-facing device as configured, rather than the foundation model alone.59 We agree. Our bill
requires logging of prompt templates, system instructions, and configurable rules applied to a
model, without requiring disclosure of a third-party model’s weights.60 This tracks the paper’s
observation that device behavior can change through prompts, retrieval strategies, guardrails,
and orchestration logic, not only the underlying model.61
• A human-performance standard. Comparing performance to qualified clinicians is the right
anchor.62 Our bill requires performance that meets or exceeds a reasonably prudent human
provider in the same or similar specialty, both to graduate from the sandbox and to expand
autonomy.63
• Measuring against the realistic alternative. Question 15 asks whether performance could be
assessed relative to the care likely to occur without the device.64 We strongly support this. For
many patients the realistic alternative is not a specialist; it is a long wait, a long drive, or no care.
Provider shortages and access gaps are present harms, and benefit-risk assessment should
count them.
• Proportionate clinical confirmation. Not every device needs a prospective trial.65 A range of
approaches scaled to risk is consistent with least burdensome principles.
• Greater reliance on postmarket evidence. The paper’s willingness to consider accepting more
premarket uncertainty in exchange for robust monitoring reflects how medicine has always
advanced: supervised use, measured outcomes, and adjustment.66 Our sandbox applies the
same logic.
• Counting both directions of error. The attention to over-escalation and over-refusal, not only
their opposites, is important.67 An overly cautious tool that sends every patient to the
emergency department or declines in-scope questions imposes real costs on patients and the
health system.
• Respect for patient capability. We support the paper’s caution against medical parentalism that
underestimates patient capability.68 Our bill addresses patient-facing risk through transparency
rather than restriction: patients have a right to know when AI is delivering their care, a right to
request human review of supervised AI decisions, and notice that they may seek care from a
human provider at any time.69
• Independent third parties and Foundation Model Master Files. Both are practical, voluntary
mechanisms that can lower review costs.70 We support them, and Section IV describes how a
recognition pathway for clinical validators could work. We ask that FDA preserve the master file

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 10
 program’s voluntary character in any future guidance. Many developers fine-tune open-weight
models on their own infrastructure, where no third-party developer may file a master file. A
device built on such a model should not be disadvantaged if the deployed system meets the
applicable benchmarking and clinical confirmation expectations, consistent with the paper’s
statement that sponsors remain responsible for demonstrating the safety and effectiveness of
their own devices.

VII. Areas for Refinement and Internal Consistency
The suggestions below aim to apply the paper’s better principles more evenly across its sections,
drawing on lessons from drafting our bill.

1. Connect the risk framework, evidence, and monitoring into one ladder
The paper’s Sections IV, V, and VI are presented separately, and Question 8 asks how they connect.71 We
suggest each cell of the two-axis grid carry a default evidence package and a default monitoring
intensity. Higher-autonomy, higher-consequence functions could receive a staged or provisional
authorization with phased deployment, volume limits, and a defined number of clinician-reviewed
interactions before oversight steps down. This is how our sandbox works,72 and it mirrors the paper’s
own description of clinician training: supervised practice with progressively greater independence.73
Greater reliance on postmarket evidence should also come with prespecified stop rules, escalation
thresholds, and the ability to restrict or withdraw a function when performance falls short. Our bill
authorizes suspension of a license when data show model drift or degraded safety outcomes.74

2. Use one functional test for directiveness
The paper considers that directiveness depends on the substance and context of an output, not on
words like “consider,” and that disclaimers may not make an output less directive.75 To make that
workable, we suggest a single anchor: whether the output is intended, represented, or reasonably relied
upon to substitute for independent clinical judgment. This is the test our bill uses for its clinical decision
support safe harbor.76 It is consistent with the clinical decision support exclusion in section 520(o)(1)(E)
of the FD&C Act, which the paper already cites, and with FDA’s January 2026 clinical decision support
guidance, which emphasizes whether software supports, rather than replaces, a health care
professional’s independent judgment.77 The test is meant to complement the statute, not displace it: it
is a way to assess directiveness within the risk framework, not a standalone determination of whether
software is a device. Functions that meet all of the statutory criteria would remain outside the device
definition. The test would help classify informational functions that fall outside the exclusion, including
patient-facing functions, which the exclusion does not reach. One test across these functions would give
manufacturers the predictability Question 2 seeks.

The same approach helps with multi-turn conversations (Question 5).78 Rather than classifying a
conversational system by its opening turn, FDA could look to the highest-risk function the system is
marketed, represented, or reasonably used to perform. Our bill takes a similar approach in defining
which services it covers.79

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 11
3. Extend the clinician analogy to change control
Clinicians are credentialed for a scope of practice and then update their knowledge continuously
without seeking new licensure for each change, provided they stay within scope and meet the standard
of care. A parallel approach for GenAI would authorize a validated scope of practice, then treat changes
that stay within that scope and maintain benchmark performance as quality-system matters rather than
new submissions. Re-benchmarking against the premarket baseline, as the paper’s Section VI.C
proposes, is the right verification tool.80 Our bill ties scope of practice to validated competency in the
same way.81 This could help answer Question 23 for modifications that cannot be fully prespecified.82

4. Generalist use and the comparator standard
Question 4 asks whether risk changes when a clinician without specialist training uses a function that
reaches into a specialist’s domain.83 We would not treat that use as presumptively higher risk. Helping
primary care physicians, nurse practitioners, and rural clinicians handle questions that would otherwise
mean a months-long specialist wait is one of the largest benefits clinical AI offers, and for many patients
the realistic alternative is no specialist input at all. Our model bill reflects this priority by extending
access to patients who cannot obtain timely care from an appropriate clinician.84 Risk is better managed
through design: clear labeling of the function’s validated scope, prompts to refer or escalate when a
case exceeds that scope, and validation showing that the generalist working with the AI performs at
least as well as the care patients would otherwise receive. Where a function is intended to substitute for
a specialist’s judgment rather than support a generalist’s, it should be held to the specialist standard,
consistent with the comparator approach below.

On the related comparator question (Question 14), we suggest the familiar professional standard of
care: a reasonably prudent clinician acting under the same or similar circumstances, meaning the
clinician who would realistically provide that service to that patient in that setting. For a screening
function deployed in primary care, the comparator should be a reasonably prudent primary care
clinician, not a subspecialist most patients never see. A specialist standard should apply only where the
device takes the place of specialist care the patient would otherwise actually receive. This builds on the
reasonably prudent provider benchmark in our bill85 and pairs naturally with Question 15: the question
is whether patients are better off with the device than with the care realistically available to them.
Human-AI team performance fits supervised functions; device-alone performance fits fully autonomous
functions. Tying the comparator to the activity axis keeps the risk and evaluation sections consistent.

5. Pair flexibility with predictable timelines
The paper’s goal of a nimble, least burdensome approach would be reinforced by clear review timelines,
particularly for re-benchmarking and change-control submissions where speed directly affects whether
improvements reach patients. Our bill uses a 30-day completeness review and a 90-day decision
deadline so that uncertainty does not become delay.86 Recognizing that FDA’s review goals for
marketing submissions are set through the MDUFA process, we suggest FDA pilot an expedited pathway
for GenAI submissions that rely on re-benchmarking against an established baseline, including PCCPrelated modifications, with completeness and decision targets shorter than standard review.

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 12
6. Consider three elements not yet addressed
In response to Question 9’s invitation to identify missing elements:87
• Commercial integrity of outputs. As patient-facing tools scale, undisclosed commercial steering
toward affiliated pharmacies, products, or services becomes a safety and trust issue. Our bill
establishes a duty of loyalty for supervised and autonomous services, bans paid content within
the clinical encounter, and requires disclosure of financial affiliations.88 A benchmarking check
for undisclosed steering could fit within element E.4.
• Output logging. Sample-based clinician review and drift monitoring both depend on retained
records. Our bill requires retention of inputs, outputs, and active configurations for at least six
years.89 A baseline logging expectation would make the paper’s postmarket approaches
workable.
• Continuity of care. The paper addresses changes to third-party foundation models but not their
discontinuation. If a foundation model is deprecated or a sponsor exits the market, patients who
rely on a service need continuity. Our bill requires a wind-down plan, a successor data
custodian, and transfer of patient data within 72 hours.90 A similar expectation could be among
the mechanisms Question 24 asks about.91

7. Build dedicated institutional capacity for clinical AI
The paper acknowledges that existing regulatory frameworks were not designed to address the
challenges GenAI-enabled devices present.92 We agree, and we believe adjustments within the current
structure may not be enough. Clinical AI is software that changes continuously, performs tasks
resembling clinical judgment, and is often built on general-purpose models. Evaluating it well requires
sustained expertise in software, machine learning, and clinical practice working together.

We encourage HHS and FDA to establish a dedicated center for clinical AI, separate from CDRH’s
traditional device and radiological-health portfolio, with responsibility for evaluating supervised and
autonomous clinical AI. Such a center would be a natural home for the competency-based approach, the
validator recognition pathway described in Section IV, protocol-bounded functions, and coordination
with CMS and the states. FDA and HHS may be able to take initial steps under existing organizational
authority; where they cannot, they could identify for Congress the authority needed.

VIII. Responses to Selected Discussion Questions
For reviewers’ convenience, the table below maps our comments to the paper’s numbered questions.
Q Topic Cicero response See
1 Dimensions of risk Two axes are sufficient; fold time-sensitivity and reversibility into the I, V
consequences axis, assess foreseeable presentations up front and
individual encounters by what was known at the time, and recognize
protocol-bounded action as a distinct activity level
2 Directiveness Use one functional test: whether output is intended, represented, or VII.2
spectrum reasonably relied upon to substitute for independent clinical judgment

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 13
Q Topic Cicero response See
3 Patient-facing Address risk through transparency, a right to know, and access to human VI
functions review, rather than restricting patient access
4 Generalist use of Not presumptively higher risk; manage through scope labeling, VII.4
specialist-domain escalation and referral prompts, and validation against care realistically
functions available; specialist standard where the function substitutes for specialist
judgment
5 Multi-turn Classify by the highest-risk function the system is marketed, represented, VII.2
conversations or reasonably used to perform
6 Under- and over- Count both directions of error; treat failure to escalate and inappropriate III, VI
escalation triage as reportable events
7 Competency-based Yes; it is the right framework and fits naturally with service-level II, VI
approach oversight
8 Linking risk to Assign each grid cell a default evidence package and monitoring VII.1
evidence intensity, with staged authorization at the top
9 Missing benchmark Add checks for undisclosed commercial steering; set baseline logging and VII.6
elements continuity expectations
11 Clinical confirmation Recognize well-designed state sandbox data and independently validated III, IV
evidence as sources of confirmation
14 Comparators Reasonably prudent clinician under the same or similar circumstances, V, VII.4
reflecting the care patients would realistically receive; specialist standard
only where the device replaces specialist care actually received; humanAI team for supervised tiers, device alone for autonomous tiers; the
protocol itself for protocol-bounded functions
15 Counterfactual Strongly support; count access gaps and delays as present harms VI, VII.4
comparator
16 Third parties Recognize developer-funded, FDA-audited clinical validators under a II, IV
federal floor and ceiling, with competing methods and independence
safeguards; their evidence informs FDA decisions
18 Postmarket reliance Support, paired with phased deployment, defined step-down criteria, VI, VII.1
and prespecified stop rules
21 Ecosystem roles States can share benchmarks, event data, and sandbox evidence while III
manufacturers remain accountable
22–23 Change control Authorize a validated scope; manage within-scope changes, including V, VII.3,
protocol updates within the authorized intended use, through the quality VII.5
system or a PCCP
24 Third-party model Include continuity and wind-down planning among the mechanisms VII.6
changes
25 Foundation Model Support; preserve voluntariness and ensure devices built on open-weight VI
MAFs models are not disadvantaged when no master file exists
26 Agentic devices Protocol-bounded execution with defined human-oversight checkpoints V
offers a practical model for multi-step action

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 14
Conclusion
The discussion paper’s tiered, risk-based approach is a move in the right direction. But as the paper itself
recognizes, existing regulatory frameworks were not designed for these technologies, and incremental
adjustments may not be enough. We encourage FDA to build on the paper by focusing oversight on
accountable clinical services and their real-world outcomes, drawing on CLIA’s proven structural model,
building dedicated institutional capacity for clinical AI, and working with states on a complementary
framework that brings clinical AI inside the health care system rather than leaving it outside.

The Cicero Institute would welcome the opportunity to discuss these comments, share our experience
working with state policymakers, and support any federal-state collaboration FDA undertakes.

Respectfully submitted,

Adam Meier
Director of Health Policy
Cicero Institute
adam@ciceroinstitute.org
U.S. Food & Drug Admin., Ctr. for Devices & Radiological Health, Considerations for the Regulation of Generative AIEnabled Medical Devices: Discussion Paper and Request for Feedback (Aug. 2026),
https://www.fda.gov/media/194242/download (hereinafter “Discussion Paper”). Page citations refer to the paper’s
printed page numbers.
Discussion Paper at 1.
Cicero Institute, The Clinical AI Services Act (Model Legislation, Ver. 2.1, Sept. 25, 2026), https://ciceroinstitute.org/wpcontent/uploads/2026/09/AI-Clinical-Services-Act-9-25-2026.pdf (hereinafter “Model Act”). The executive summary
accompanying the model legislation refers to it as “The AI Clinical Services Act”; the short title in section 1(a) is “Clinical
AI Services Act.”
Discussion Paper at 5–6 & fig. 1.
Model Act § 2(c), (f).
Model Act § 4(a)–(b), (h); see also the Regulatory Reference Table in the Model Act’s executive summary.
Discussion Paper at 7.
Model Act § 2(f).
Discussion Paper at 9 (Question 1).
Model Act § 2(f)(4)–(5).
Discussion Paper at 2, 10.
42 U.S.C. § 263a; 42 C.F.R. pt. 493.
42 C.F.R. § 493.1253(b)(2).
42 C.F.R. pt. 493, subpt. H (proficiency testing); 42 C.F.R. § 493.1236 (evaluation of proficiency testing performance
and verification of accuracy).
Ctrs. for Medicare & Medicaid Servs., LDT and CLIA FAQs, https://cms.gov/regulations-andguidance/legislation/clia/downloads/ldt-and-clia_faqs.pdf.
Enhancing Clinical Laboratory Innovation and Access Act of 2026, H.R. 8890, 119th Cong. (introduced May 19, 2026),
https://www.govinfo.gov/app/details/BILLS-119hr8890ih; see also Covington & Burling LLP, Enhancing Clinical

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 15
Laboratory Innovation and Access Act (Enhancing CLIA Act) of 2026 (June 2026), https://www.cov.com/en/news-andinsights/insights/2026/06/enhancing-clinical-laboratory-innovation-and-access-act-enhancing-clia-act-of-2026; Am. Soc’y
for Clinical Pathology, New Legislation Clarifies CMS Authority Over Laboratory Developed Tests (June 10, 2026),
https://www.ascp.org/news/news-details/2026/06/10/new-legislation-clarifies-cms-authority-over-laboratorydeveloped-tests.
Am. Clinical Lab’y Ass’n v. U.S. Food & Drug Admin., No. 4:24-cv-479-SDJ (E.D. Tex. Mar. 31, 2025) (consolidated with
Ass’n for Molecular Pathology v. U.S. Food & Drug Admin., No. 4:24-cv-824-SDJ).
U.S. Food & Drug Admin., final rule rescinding the May 6, 2024 LDT rule, Fed. Reg. Doc. No. 2025-18239 (Sept. 19,
2025), https://www.federalregister.gov/documents/2025/09/19/2025-18239/regulation-identification-number-0910aj05-medical-devices-laboratory-developed-tests-implementation; see also FDA Rescinds Final Rule on LaboratoryDeveloped Tests, 360Dx, https://www.360dx.com/node/105636.
42 C.F.R. pt. 493, subpt. E (accreditation by private, nonprofit accreditation organizations and exemption under
approved state laboratory programs).
Model Act § 4(e)(1)(A).
Model Act § 7(e)(1), (3), (5).
Model Act § 6(d).
Model Act § 7(g)(2).
Discussion Paper at 10 & n.16 (citing Freyer O, Jayabalan S, et al., Overcoming Regulatory Barriers to the
Implementation of AI Agents in Healthcare, Nature Medicine 2025;31(10):3239–3243).
Discussion Paper at 1.
Discussion Paper at 17, 20 & n.22.
Discussion Paper at 4.
Discussion Paper at 22.
21 U.S.C. § 396.
Model Act § 4(e)(1)(B), (f)(2).
Model Act § 7(g)(3), (5).
Model Act § 12(b)–(c).
Discussion Paper at 24–26.
Discussion Paper at 17.
Discussion Paper at 15.
Model Act § 7(a)–(b).
Model Act § 7(d)(3)–(6).
Model Act § 2(a), (l).
Discussion Paper at 20.
Discussion Paper at 18–19 (Question 16).
42 U.S.C. § 1395bb.
21 U.S.C. § 360m.
Discussion Paper at 17.
Utah Dep’t of Commerce, Utah Enhances Pro-Human AI Initiative with New Healthcare Pillar, Adds New Pilots to its
Regulatory Sandbox, and Partners with Third-Party Evaluators (Oct. 5, 2026),

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 16
https://www.einnews.com/pr_news/947788475/utah-enhances-pro-human-ai-initiative-with-new-healthcare-pillaradds-new-pilots-to-its-regulatory-sandbox-and-partners-with-third-party-evaluators.
Utah Off. of A.I. Pol’y, Third-Party Evaluators, https://commerce.utah.gov/ai/third-party-evaluators/ (last visited Oct.
19, 2026).
Joint Comm’n, Responsible Use of AI in Healthcare Certification (2026), https://www.jointcommission.org/enus/knowledge-library/news/2026-05-responsible-use-of-ai-in-healthcare-certification; Cailey Gleeson, Joint Commission
Launches Voluntary AI Certification Program for Healthcare Organizations, Fierce Healthcare (June 2, 2026),
https://www.fiercehealthcare.com/ai-and-machine-learning/joint-commission-launches-voluntary-ai-certificationprogram-healthcare.
Discussion Paper at 14.
Model Act § 7(g)(3).
42 C.F.R. § 482.24(c)(3); see Medicare and Medicaid Programs; Reform of Hospital and Critical Access Hospital
Conditions of Participation, 77 Fed. Reg. 29,034 (May 16, 2012).
Discussion Paper at 2.
Discussion Paper at 16.
Discussion Paper at 24, 26.
Model Act § 3(f)(3).
Model Act § 4(e)(2)(C).
Model Act §§ 2(k), 7(e)(3).
Discussion Paper at 23 (Question 26).
Discussion Paper at 5.
Model Act § 12(a).
Discussion Paper at 11.
Model Act § 6(d)(2).
Discussion Paper at 2.
Discussion Paper at 16.
Model Act §§ 4(f)(5), 7(a)(2), 7(g)(2).
Discussion Paper at 18 (Question 15).
Discussion Paper at 14–15.
Discussion Paper at 19.
Discussion Paper at 8–9, 24.
Discussion Paper at 8.
Model Act § 5(a)–(c).
Discussion Paper at 16–17, 22.
Discussion Paper at 17 (Question 8).
Model Act § 7(b).
Discussion Paper at 11.
Model Act § 7(g)(4).
Discussion Paper at 7.

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 17
Model Act § 4(a)(1), (h).
21 U.S.C. § 360j(o)(1)(E); Discussion Paper at 8 n.13; U.S. Food & Drug Admin., Clinical Decision Support Software:
Guidance for Industry and Food and Drug Administration Staff (Jan. 2026) (superseding the September 2022 guidance).
Discussion Paper at 8, 9–10 (Question 5).
Model Act § 2(d)(4).
Discussion Paper at 20–21.
Model Act § 8(a).
Discussion Paper at 22 (Question 23).
Discussion Paper at 8, 9 (Question 4).
Model Act § 7(c)(6).
Model Act §§ 7(g)(2), 8(a)(3).
Model Act § 4(i).
Discussion Paper at 17–18 (Question 9).
Model Act § 6(a)–(c).
Model Act § 6(d)(1)–(2).
Model Act § 7(f).
Discussion Paper at 22 (Question 24).
Discussion Paper at 23.

Cicero Institute Comments, Docket No. FDA-2026-N-7874 | Page 18