Society of Quality Assurance (Alison Cockrell)

Please see the attached comments from the Society of Quality Assurance.

18 October 2026 via regulations.gov

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

Re: FDA-2026-N-7874

Dear Dockets Manager:

Please find enclosed comments from the Society of Quality Assurance (SQA) on
Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper. SQA appreciates the opportunity to provide this feedback. Please
contact me if you have any questions.

Sincerely,

Alison Cockrell
Alison Cockrell, CAE
Executive Director
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
3 III — N/A Foundation models are AI models trained using • Clarify the sources of the unlabeled datasets The definition describes dataset labeling but does not explain data
Foundation- large, typically unlabeled datasets described in the foundation-model definition, origin. Clarifying the sources would address uncertainty about the
model including whether these may include internet- provenance and reliability of foundation-model training data.
definition
derived data or data obtained from controlled
sources.
9 IV.A — N/A If there are additional dimensions—such as the • Clarify whether additional risk factors should be Device activity and consequences alone may not capture
Question 1 reversibility of a resulting action, the availability of incorporated as modifiers within the two-axis differences in users’ ability to identify and challenge errors, obtain
downstream safeguards, the time pressure of the framework or through an expanded framework, independent review, or operate without the device. A single
deployment setting, or the traceability of the output classification may also obscure variation among functions.
such as a multifactor assessment, and how either
(i.e., to primary source materials)—that should be Connecting these considerations to risk thresholds and expected
represented in a risk framework, please describe approach would determine evidence expectations. evidence would support consistent, actionable assessment.
and provide examples of how they should be • Consider error detectability, intended-user
represented. expertise, effective human oversight, and
operational dependence alongside the factors
already identified, as well as transparency,
explainability, fairness, reliability, safety, privacy,
security, and resilience, consistent with the NIST AI
Risk Management Framework.
• Provide function-specific plots or worked examples
showing how these considerations and ranges of
consequences within an activity category and across
device functions inform risk thresholds and
associated assurance activities.
• Clarify expectations by activity level, including
relevant technical, ethical, and psychological effects.
9 IV.A — N/A the traceability of the output (i.e., to primary source • Clarify expectations for traceability and Source access and explanations may help healthcare professionals
Question 1; materials) explainability that support users’ ability to and patients evaluate recommendations before acting. Clarifying
related Q3 independently evaluate outputs and detect errors, when traceability warrants a reduction in assessed risk would
support consistent application without assuming that traceability
particularly for action-directing outputs.
alone establishes lower risk.
• Address access to primary source materials and
information about which aspects of the prompt or
AI features most influenced the output.
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
• Explain whether, and under what conditions,
traceability should reduce assessed risk.
9 IV.A — N/A If there are additional dimensions • Consider how missing patient-specific context, An output appropriate for the general population may be harmful
Question 1; including underlying medical conditions and when relevant patient context is missing. Collecting that context
related Q4 concomitant medications, affects risk assessment. also introduces privacy and data-protection considerations.
• Clarify expectations for obtaining relevant context,
informing patients of its use, and protecting the
health information collected.
9 IV.A — N/A the two-axis risk framework, organized around • Clarify the relationship between “Device Activity” in Explaining the relationship and any necessary distinctions would
Question 1 device activity and the consequence of relying on an this framework and “Model Influence” in FDA’s support consistent interpretation across FDA frameworks without
incorrect output Considerations for the Use of Artificial Intelligence implying that the terms are interchangeable.
to Support Regulatory Decision-Making for Drug and
Biological Products: Draft Guidance for Industry and
Other Interested Parties (January 2025), including
whether harmonized terminology would be
appropriate given their respective scopes.
9 IV.A — N/A What additional information would provide • Provide criteria and tabulated examples showing Criteria and examples would support consistent classification and
Question 2 manufacturers with sufficient clarity and how wording, specificity, personalization, certainty, clarify acceptable output characteristics. Addressing overreliance
predictability around risk assessment for actionability, intended audience, perceived and safeguards would connect classification to how users may
informational functions while recognizing that interpret and act on the information.
authority, and traceability affect risk across the
directiveness may exist along a continuum rather
than as a binary distinction? continuum from non-directive to action-directing
information.
• Clarify when personalization, including addressing
the user by name, should increase assessed risk
because of potential overreliance, and whether
particular wording or sentence structures should be
permitted or restricted for each output category.
• Manufacturers should demonstrate that outputs are
evidence-based, traceable to reliable sources, and
supported by safeguards against unsupported,
misleading, or overly directive recommendations.
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
9 IV.A — N/A What device characteristics, output features, or • Assess patient-facing risk according to users’ ability Patients vary in their ability to assess clinical information and
Question 3 safeguards might mitigate risks that could arise to evaluate the output, including health and AI recognize an output’s limitations. Safeguards tailored to those
when a user lacks the domain knowledge to literacy, clinical complexity, personalization, differences can address risk while preserving patient
independently evaluate an output, without empowerment, engagement, access to information, and shared
urgency, diagnostic definitiveness, emotional
unnecessarily underestimating patient capability? decision-making.
impact, and actionability, rather than automatically
assigning higher risk to patient-facing use.
• Provide illustrative examples of safeguards,
including evidence-based and traceable outputs,
understandable explanations of the output’s basis
and limitations, uncertainty communication, plainlanguage interpretation of clinical terminology, and
pathways to healthcare-professional review.
9 IV.A — N/A Under what circumstances, if any, might risk be • Assess risk according to intended users’ training, HCP status alone does not establish expertise to evaluate every
Question 4; affected when an HCP who lacks the relevant clinical clinical expertise, responsibilities, and ability to output. Accounting for users’ roles and access to specialist or
related Q3 specialist knowledge receives information that falls identify errors, distinguishing laypersons, allied independent review would address risks arising when information
within a specialist area of practice? What device exceeds their expertise or involves several professionals with
healthcare professionals, generalist physicians, and
characteristics or safeguards might mitigate such different responsibilities.
risks? specialists.
• Address workflows involving multiple HCPs with
different responsibilities, including informational
and action-taking activities. Specify circumstances
warranting specialist consultation, intended-user
restrictions, or independent secondary review,
supported by access to evidence and
communication of uncertainty.
10 IV.A — N/A How could the intended use of such a device be • Assess risk across realistic multi-turn conversational Individual outputs or functions may not reveal risks arising from
Question 5 characterized when its behavior is emergent across pathways, including intended and foreseeable accumulated context, transitions between activities, or interactions
begins p9 a conversation? unintended use, patient context, different HCP among functions and users. Conversation-level assessment,
boundary controls, and monitoring would address these risks
roles, and worst-case scenarios.
throughout the exchange.
• Clarify when assessment should address individual
functions, combinations of functions, or the
complete user experience, accounting for shared
and function-specific risks.
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
• Characterize intended use by the highest level of
function expected across the conversation. Use
pathway maps, specific follow-up questions, and
retained conversational history to support
evaluation.
• Demonstrate that guardrails maintain intended-use
boundaries, including preventing non-directive
devices from progressing to action-directing
recommendations outside their intended use.
• Require routine review of conversational escalation
paths, outputs, and outcomes; consider
manufacturer alerts when defined capability
thresholds are reached or exceeded.
• Clarify expectations for scenario coverage, including
how manufacturers should justify the completeness
of the combinations evaluated.
• Evaluate safeguards that adjust as conversations
become increasingly action-directing. Include
patient-safety, product-quality, and data-integrity
consequences in scenario assessment.
10 IV.A — N/A How could manufacturers characterize and weigh • Evaluate under-escalation and over-escalation Under-escalation and over-escalation can produce different harms.
Question 6 these two directions of error, given that they may separately, with weighting justified by clinical Separate evaluation and context-specific weighting would make
not be commensurable and that acceptable trade- context, severity, public health impact, reversibility, trade-offs explicit and support clinically meaningful acceptance
offs may vary by clinical context? criteria rather than a uniform performance threshold.
and urgency rather than assumed equal weighting.
• Address life-threatening, time-sensitive, and routine
situations, including clinical, operational, and
behavioral consequences.
• Test realistic multi-turn scenarios for time to
escalation, inappropriate reassurance, user
behavior, and automation bias.
• Consider tiered escalation, clinical red-flag rules,
and reassessment as conversations evolve.
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
11 V.A — N/A Importantly, a competency-based approach could • Clarify how risk-based planning under the Clarification would distinguish their scopes and identify where
Paragraph 2; be tailored to the device’s intended use and be competency-based approach relates to the existing evidence may support both assessments.
related Q7– proportionate to its risk, potentially utilizing the Credibility Assessment Plan described in FDA’s
8,14 two-axis risk framework described in Section IV.
Considerations for the Use of Artificial Intelligence
to Support Regulatory Decision-Making for Drug and
Biological Products: Draft Guidance for Industry and
Other Interested Parties (January 2025), including
differences in applicability and opportunities to
reuse evidence.
13 V.B.2 — N/A prespecify evaluation methods, describe and justify • Select performance comparators according to Intended-use-based comparators support relevant performance
Opening; scoring rubrics, identify qualified expert adjudicators intended use. assessment. Acceptable aggregate performance may mask failures
related Q14– where applicable, and prespecify and justify • Consider defining separate acceptance criteria for in controls such as patient identification or required human
15,26 acceptance criteria. approval; separate criteria would make these failures explicit in
clinical performance and critical system controls,
acceptance decisions.
specifying in advance how critical failures affect
overall acceptance.
14 V.B.2 — Third N/A These considerations would still be applicable when • Consider assessing automated evaluators against An automated evaluator may share the device’s errors.
bullet; related the expert adjudicator is itself an LLM. independent expert judgments and known failure Independent reference judgments, protected test data, and thirdQ10,16–17 cases, with controls for shared blind spots, party participation can strengthen evaluation without transferring
the manufacturer’s responsibilities.
evaluator changes, and protection of test data from
development use and repeated tuning.
• Support qualified, independent third-party
participation in benchmark maintenance, validation
testing, and dataset governance while retaining the
manufacturer’s regulatory responsibilities.
17 V.E — N/A Is the competency-based approach described above, • Support a competency-based approach combining Benchmark results alone may not establish performance in clinical
Questions 7–8; i.e., device benchmarking followed by clinical device benchmarking and clinical confirmation, with use. Changes in device behavior or underlying models warrant
related Q9 confirmation, a useful and appropriate framework evaluation proportionate to intended use and risk. continued assessment, with evaluation frequency proportionate to
for evaluating GenAI-enabled devices? risk.
• Benchmark performance should not be the sole
basis for establishing safety and effectiveness; realworld clinical validation should be conducted for
autonomous systems, diagnostic devices, and
treatment-recommendation devices.
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
• Extend competency assessment through postmarket
monitoring, periodic re-benchmarking, and change
management, with more frequent benchmarking or
confirmation for higher-risk devices.
17–18 V.E — Question N/A Are there elements that are missing, redundant, or • Include security and data-integrity considerations in Clinical-performance benchmarks may not reveal vulnerabilities
9 inappropriately categorized? benchmarking: adversarial resilience testing, that compromise outputs, actions, or sensitive information. Explicit
including prompt injection and tampering; least- evaluation would help identify these weaknesses and clarify how
security, privacy, and data integrity are assessed.
privilege and access-control enforcement; testing
for unintended disclosure of sensitive information;
and clear documentation of stored data and
retention requirements.
• Clarify whether these belong within the existing
benchmarking elements or warrant an additional
category.
18 V.E — Q10; N/A what role should sponsor-developed benchmarks • Prefer independent benchmark repositories Dataset suitability and independence affect whether measured
related Q12–13 play given potential concerns around independence supplemented by external challenge datasets. performance reflects clinical use. Synthetic datasets may reproduce
and optimization to the test? • Sponsor-developed datasets may be acceptable training-data biases, limiting their suitability for evaluating
subgroup differences, equity, and bias.
where clinical representativeness, external
validation, data independence, and generalizability
are demonstrated.
• Use statistically robust clinical datasets when
available.
• Synthetic data may supplement evaluation of rare
conditions, edge cases, and scenario stress testing,
but should not be used for demographic subgroup
analysis, health-equity evaluation, or bias detection.
18 V.E — Q11; N/A How might a sponsor select and justify a • Provide a tiered approach linking confirmation rigor A tiered approach can build evidence progressively while limiting
related Q12 confirmation approach tailored to a device’s to clinical and device risk: retrospective evaluation patient exposure. Higher-consequence uses warrant stronger
intended use and proportionate to the device’s risk or shadow deployment with clinician review for clinical evidence, while security controls remain relevant across
profile? confirmation approaches.
lower-risk assistive uses; prospective observation
with clinician adjudication for intermediate-risk
uses; and prospective studies for higher-risk
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
autonomous uses. Higher-risk diagnosis, treatment,
and autonomous actions should require real-world
performance evidence and clinically meaningful
endpoints.
• Clarify the limitations of retrospective or synthetic
approaches for dose- or therapy-controlling devices,
medication selection, and devices with write access
to clinical records, including when prospective
confirmation is required.
• Apply stronger security controls to these higher-risk
functions.
• Across tiers, address secure pipelines, least-privilege
access, version rollback, change logs and audit trails,
runtime guardrails, and adversarial testing.
20 VI.A — After N/A CDRH seeks feedback on these approaches, as well • Consider requiring continuous drift monitoring, real- Monitoring can reveal changes in performance and emerging risks
monitoring as the feasibility and practicality of implementing world performance reporting, periodic safety after deployment. Linked, protected records support investigation
approaches; these or other approaches to postmarket reviews, adverse-event trend analysis, cybersecurity and risk response, including when exact output reproduction is not
related Q19 monitoring for GenAI-enabled devices. feasible.
monitoring, and bias management.
• Address changes in input data and prompts relative
to training data, and the model’s access to realworld data.
• Monitoring should support risk response as well as
risk detection.
• Risk-proportionate records could link relevant
inputs, sources, configurations, tool activity, human
decisions, outputs, and resulting actions, with
controls for integrity, privacy, retention, and
retrieval.
20 VI.A — Final N/A whether some aspects of postmarket monitoring • Consider assessing supervisory agents for missed Shared dependencies may cause the device and supervisor to fail
sentence; can be enabled by machine-based supervisory detections, false alarms, detection and response together. Demonstrated monitoring effectiveness and appropriate
related Q20 agents times, and shared dependencies. fallback would help justify reliance on supervisory agents.
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
• Monitoring plans could define responsibility for
alerts, escalation, change control, and fallback when
monitoring is unavailable or unreliable.
21 VI.C — Devices N/A Devices built on third-party foundation models raise • Document the foundation-model version within the Supplier-initiated changes can affect safety, intended use, and the
built particular change-control considerations because evaluated configuration. applicability of prior evaluation evidence. Version pinning alone
paragraph; changes to the underlying model may be initiated by • Establish advance update-notification agreements, may not control all service-side changes. Combining contractual and
related Q24 the foundation model developer rather than the technical controls with impact assessment would support continued
version controls, change-detection processes,
p22 device manufacturer. assurance and make limitations in manufacturer control explicit.
revalidation triggers, and rollback capability.
• Assess updates through change control before
deployment, using risk assessment, rebenchmarking, predefined acceptance criteria, and
available supplier performance and safety
information.
• Where supplier control is limited, evaluate
compensating controls, residual risk, and conditions
for restricting or suspending use.
• Credit human review or AI monitoring as
compensating controls only where effectiveness is
demonstrated.
21 VI.D — N/A Under what conditions might such an approach be • Limit consideration of greater premarket This distinction would prevent reliance on postmarket detection
Question 18 appropriate, and what characteristics of a uncertainty to low-risk functions. from displacing evidence needed before high-risk clinical use.
monitoring program would need to be in place to • High-risk uses, including autonomous diagnosis,
justify reduced premarket evidence?
autonomous treatment, and critical care, should not
receive reduced premarket evidence expectations
on the basis of postmarket monitoring.
• Postmarket controls should supplement, rather than
replace, premarket evidence.
21 VI.D — N/A how can these roles be structured without diffusing • Clarify that healthcare institutions may contribute Clearly distinguishing data contributions from regulatory obligations
Question 21 manufacturer accountability? data to postmarket monitoring without assuming would support institutional participation while preserving
the manufacturer’s regulatory obligations. manufacturer accountability.
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
21 VI.D — Q22; N/A How might the extent of re-benchmarking or other • Scale re-benchmarking and supporting evidence to Small technical changes can meaningfully alter outputs or safety. An
related Q23 evidence be scaled to the nature and expected expected behavioral and clinical impact, rather than impact-based approach would distinguish changes requiring
p22 impact of a given modification? the technical size of a change. renewed evidence from those appropriately managed within the
quality management system and clarify the scope of PCCPs.
• Assess changes to models, prompts, retrieval,
context handling, safety guardrails, and thresholds.
• Clarify when minimal-risk, non-behavioral changes
may be managed within the manufacturer’s quality
management system and when FDA review is
needed.
• Consider including model, prompt, retrieval, safetyguardrail, and foundation-model changes within
Predetermined Change Control Plans (PCCPs),
where appropriate.
23 VII.C — N/A Would voluntary Foundation Model MAFs be • Support voluntary Foundation Model MAFs with Standardized, independently reviewed foundation-model
Question 25 practical for foundation model developers to standardized content covering model architecture information could improve transparency and reduce duplicated
provide and to keep current, and what content and version, training methodology and data assessment across manufacturers, while preserving responsibility
might they include? for device-specific safety and effectiveness.
provenance, known limitations, bias assessments
and performance across relevant populations,
cybersecurity controls, safety guardrails,
performance testing, change-management
processes, update-notification commitments, and
monitoring requirements.
• Consider a mechanism for independent review and
shared reference by multiple device manufacturers,
drawing on the qualified trust-service provider and
trusted-list approach under the European Union’s
eIDAS framework as a potential organizational
analogy.
• Manufacturers should remain responsible for
evaluating the foundation model within their
device’s intended use.
 SQA RRT Template for FDA Draft Guideline Comments – Discussion Paper: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Page Section/ Line Current Wording Proposed Wording Rationale
Paragraph (clarifications, spelling/grammatical
corrections, etc.)
23 VII.C — Q26; N/A How should the elevated risk associated with • Apply enhanced scrutiny to agentic systems, Autonomous action sequences may propagate errors and reduce
related autonomous multi-step action, tool use, and including evaluation of planning safety, tool opportunities for intervention. Evaluation should establish that
Appendix A A.1 reduced opportunity for human review be reflected governance, action authorization, cascading errors, safeguards and human decisions remain effective throughout
p26 in acceptance criteria and oversight? execution, rather than relying on the presence of a checkpoint
and operational boundaries. Include technical
alone.
safeguards, logging and auditing of agent actions,
and explanations supporting review of each action.
• Assess whether human reviewers have sufficient
information, time, and authority to intervene at
required decision points.
• Verify that actions requiring human approval
execute only within that approval’s scope, including
after retries, restarts, or material changes.
26 Appendix A — N/A “accurate tool use and recognition of erroneous tool • Consider testing interoperability and recovery Accurate individual tool calls do not establish safe end-to-end
A.1; related outputs” and “graceful handling of tool failures” across connected systems, tools, and agents, operation. An action may complete even when confirmation is lost,
Q9,26 including incomplete, delayed, or duplicate creating risks during recovery or retry. Testing the connected
workflow would address both information-transfer failures and
exchanges, partial execution, lost context, and
unsafe recovery behavior.
uncertain action status.
• Verify preservation of clinical meaning, patient
identity, workflow state, and applicable approvals;
prevention of omitted, duplicated, or incorrectly
sequenced actions; and clinically appropriate
fallback when execution status remains unresolved.
26 Appendix A — N/A resistance to prompt injection through user inputs, • Consider documenting tested prompt-injection Identifying injection sources does not establish effective resistance.
A.1; related retrieved content, and tool outputs pathways, demonstrated effectiveness of Evidence of prevention and containment would support assessment
Q9,26 prevention and containment controls, testing of whether connected workflows remain within their authorized
and intended operation.
limitations, and residual risk across connected
workflows.
• Assess whether injected instructions can alter
clinical outputs or trigger unauthorized actions.