BioPhorum Operations Group

BioPhorum welcomes FDA’s engagement and submits the attached industry feedback on a risk-proportionate, lifecycle-based approach to GenAI-enabled medical devices.

Feedback to FDA on the regulatory
approach for generative AI-enabled
medical devices
Response to FDA Discussion Paper | Docket FDA-2026-N7874 | October 2026
Contents

1 Executive summary.................................................................................................................................................. 5
1.1 Risk assessment ............................................................................................................................................ 5
1.2 System assurance ......................................................................................................................................... 5
1.3 Configuration control and traceability .......................................................................................................... 6
1.4 Human oversight .......................................................................................................................................... 6
1.5 Premarket evidence ...................................................................................................................................... 6
1.6 Postmarket assurance .................................................................................................................................. 6
1.7 Change management .................................................................................................................................... 7
1.8 Third-party components ............................................................................................................................... 7
1.9 Combination products and investigational use ............................................................................................ 7
2 Introduction ............................................................................................................................................................. 8
2.1 Background and scope ........................................................................................................................................... 8
3 Foundational regulatory principles........................................................................................................................ 10
3.1 Risk proportionality .................................................................................................................................... 10
3.1.1 Scaling controls and evidence with risk ................................................................................................. 10
3.2 System-level assurance ............................................................................................................................... 11
3.3 Traceability ................................................................................................................................................. 12
3.4 Reversibility and downstream safeguards .................................................................................................. 13
3.5 Visible failure and graceful degradation ..................................................................................................... 13
3.6 Lifecycle evidence ....................................................................................................................................... 14
3.7 Use case-specific metrics ............................................................................................................................ 14
3.8 Allocation of responsibility ......................................................................................................................... 15
4 Response to FDA’s four discussion areas............................................................................................................... 16
4.1 Risk assessment and human oversight ....................................................................................................... 16
4.1.1 Human oversight .......................................................................................................................................... 16
4.1.3 Foreseeable misuse and unintended use ..................................................................................................... 17
4.2 Premarket evaluation ................................................................................................................................. 18
4.3 Postmarket monitoring ............................................................................................................................... 19
4.4 Additional regulatory considerations ......................................................................................................... 22
4.5 Transparency and explainability ................................................................................................................. 22
4.6 Foundation models and third-party components....................................................................................... 22

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 2
 4.7 Combination products ................................................................................................................................ 23
4.8 Data governance, privacy and cybersecurity .............................................................................................. 24
4.9 Change management .................................................................................................................................. 24
4.10 Total Product Lifecycle................................................................................................................................ 25
4.11 Relationship with other regulatory and policy work .................................................................................. 25
5 Conclusion ............................................................................................................................................................. 27
Sources ........................................................................................................................................................................... 28
Glossary and acronyms ................................................................................................................................................... 30

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 3
© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 4
1 Executive summary
BioPhorum welcomes FDA’s early engagement on the regulation of generative artificial intelligence (GenAI)-enabled
medical devices. BioPhorum’s central position is that regulatory oversight should be risk-proportionate, system-level and
lifecycle-based.

GenAI can support clinically and operationally valuable functions, but its variable and open-ended outputs, rapid
component change, reliance on third-party foundation models and potential for plausible but unsupported responses
create assurance challenges that are not fully addressed by deterministic software validation alone.

BioPhorum recommends that FDA build on established medical-device risk management, software assurance and Total
Product Lifecycle principles, using these foundations to set targeted expectations for the characteristics that are
distinctive to GenAI.

In particular, evaluation should focus on the complete deployed device configuration rather than the model in
isolation, as system-level context may materially affect safety, performance and risk. This includes:

• The generative model, prompts and system instructions
• Retrieval sources
• Connected tools
• Deterministic controls
• User workflow
• Deployment environment
• Downstream actions
The nature and strength of evidence should reflect intended use, severity of consequence, degree of autonomy,
detectability of failure, reversibility and availability of effective controls.

1.1 Risk assessment
Industry considers that established medical-device risk-management methods remain an appropriate foundation. For
GenAI-enabled functions, BioPhorum recommends that FDA consider how those methods may need to address
probabilistic behavior, unsupported outputs, silent failure, instruction-following failures, unsafe tool use and supplier
dependency.

1.2 System assurance
BioPhorum considers that safety and effectiveness are most appropriately assessed for the integrated device system,
while deterministic components remain suitable for conventional verification and validation.

Generative behavior should be evaluated using:

• Representative scenarios
• Repeated sampling
• Challenge testing
• Failure-mode analysis
• Statistically justified acceptance criteria

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 5
1.3 Configuration control and traceability
From an industry perspective, defining the validated state as an identified configuration, rather than merely a model
version, could provide greater regulatory clarity. Manufacturers may need to be able to attribute an output to the
relevant:

• Model or service
• Prompts
• Retrieval corpus
• Inference settings
• Guardrails
• Tools
• Supervisory components
Manufacturers may also need to be able to reconstruct the processing, review and decision pathway needed for
investigation.

1.4 Human oversight
BioPhorum considers that human review should be credited as a risk control only where reviewers have:

• The competence to identify, challenge, override or escalate an unsafe output
• The information to identify, challenge, override or escalate an unsafe output
• The time to identify, challenge, override or escalate an unsafe output
• The authority to identify, challenge, override or escalate an unsafe output
• Oversight design should address workload and automation bias.
Where one AI system is used to verify or monitor another, BioPhorum considers that independent qualification for that
specific role would provide a stronger assurance basis.

1.5 Premarket evidence
Premarket submissions could usefully define:

• The intended-use risk category
• The validated operating envelope
• The device-level architecture
• The human-oversight model
• Foreseeable failure modes
• Evidence supporting each control
BioPhorum considers that evaluation of variable outputs should use pre-specified, risk-justified methods.

It should distinguish development, tuning and final test data, including controls for benchmark contamination.

1.6 Postmarket assurance
From an industry perspective, postmarket monitoring is most useful when designed during premarket development and
linked to residual uncertainty, the approved operating envelope and change control. Metrics, thresholds and cadence
should be justified for the specific use case. BioPhorum considers that clinically critical silent failures require stronger
premarket evidence and should not be deferred to postmarket detection.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 6
1.7 Change management
Changes to models, prompts, retrieval corpora, connected tools, supplier services or deployment environments can alter
device behavior. BioPhorum considers that such changes should be assessed prospectively where they may affect safety,
performance or the validated risk envelope. Postmarket monitoring complements, but does not replace, verification of a
changed configuration before deployment. A Predetermined Change Control Plan (PCCP) may provide a useful
mechanism where changes and evidence requirements can be appropriately bounded.

1.8 Third-party components
Use of a foundation model, cloud service or supplier-managed component does not remove the legal manufacturer’s
responsibility for the marketed device. BioPhorum recommends that FDA clarify expectations for:

• Version identification
• Supplier evidence
• Change notification
• Audit and investigation support
• Continuity
• Rollback
• Replacement
• Safe retirement

1.9 Combination products and investigational use
Industry considers that greater coordination across Centers could help avoid duplicative or divergent expectations for
GenAI-enabled functions used in drug-device combination products or in support of drug and biologic investigations.
Greater clarity could cover:

Lead and consulting responsibilities across CDRH, CDER and CBER

• The applicability of predetermined change control to device constituent parts of drug-led products
• Proportionate evidence pathways for GenAI-enabled tools used only within controlled clinical investigations and
not intended for standalone commercial use
BioPhorum considers that any future framework would benefit from distinguishing:

• Tools that process study data, support modeling or derive endpoints
• Functions that directly inform diagnosis, treatment or another consequential participant-level decision
BioPhorum also considers that a coordinated approach would help avoid duplicative or divergent requirements for the
same technology.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 7
2 Introduction
2.1 Background and scope
BioPhorum is a globally trusted, industry-led organization that brings biopharmaceutical leaders together to collaborate,
share best practice and shape industry standards.

This response was developed through structured discussion and review by subject-matter experts from BioPhorum
member organizations, drawing on their experience across medical devices, drug-device combination products, clinical
development, regulatory science, quality and digital technologies. It presents a consolidated industry perspective rather
than the position of any individual contributor or member company.

Generative AI introduces regulatory questions that may not be fully addressed by conventional deterministic software
validation. GenAI-enabled systems may produce different outputs from the same or similar inputs and may introduce
new forms of uncertainty, variability and failure. Traditional software, by contrast, is generally expected to produce
bounded and repeatable outputs from defined inputs and algorithms.

This BioPhorum response addresses the regulatory implications of generative models, including large language models, in
medical-device, combination-product and clinical-investigation contexts. It draws on established AI, software, quality,
risk-management and medical-device principles where they remain applicable and identifies areas in which GenAIspecific expectations are needed for evaluation, oversight, monitoring and change control.

It covers GenAI-enabled functions that constitute a medical device or device software function and device constituent
parts of combination products. It also covers GenAI-enabled technologies used in clinical investigations or drug and
biologic development where their outputs may affect data acquisition, endpoint derivation, modeling, safety or
effectiveness assessment, or another regulatory decision.

BioPhorum has a particular interest in the application of GenAI to drug-device combination products and to technologies
used during clinical investigation and drug or biologic development. For the purpose of this response, a GenAI-enabled
medical device is a device, or a device software function, in which a generative model, for example, a large language
model, multimodal model or image-generation model, produces content such as text, images, structured data or
recommendations that inform or initiate a device function.

This includes architectures that combine a generative model with retrieval of external information (retrieval-augmented
generation) or with the ability to call tools or initiate actions (agentic systems). Because risk depends strongly on the
architecture, intended use and context of deployment, BioPhorum considers that explicit distinction between these
configurations would support proportionate oversight.

For clarity, this response considers three regulatory contexts that may warrant distinct treatment:

• A GenAI-enabled investigational device subject to the applicable device framework
• A GenAI-enabled function incorporated into, or used in support of, a drug-device combination product
• A GenAI-enabled tool used only within a clinical investigation or drug-development program and not intended
for standalone commercial use
The third category may process study data, support modeling, derive an endpoint or inform a drug or biologic regulatory
decision without being developed as a marketed medical-device product. FDA is encouraged to clarify when such a tool
falls within the device framework, when a drug or biologic model-credibility framework may be more appropriate, and
what proportionate evidence and controls should apply in the clinical-investigation environment.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 8
This distinction is important because investigational use should not create an assumption that a device must first satisfy
the evidence expectations applicable to a marketed product before it can be used to investigate a drug or biologic.
Clinical investigations may provide controlled access, defined protocols, qualified users, enhanced review and oversight,
and other safeguards that BioPhorum considers relevant to determining the appropriate evidence pathway.

Industry considers that greater coordination across CDRH, CDER and CBER could help avoid duplicative or divergent
requirements where the same GenAI-enabled technology supports a product reviewed by a different Center.

Use of generative AI in pharmaceutical manufacturing and quality operations remains outside the scope of this response
and is subject to separate GMP expectations. This includes emerging international proposals that would restrict
generative and probabilistic models in critical GMP applications. Nothing in this response is intended to address or
preempt those expectations.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 9
3 Foundational regulatory principles
3.1 Risk proportionality
Industry considers that a risk-based approach may be more appropriate because regulatory controls and evidence needs
may vary according to intended use, severity of consequence, degree of autonomy, detectability of failure, reversibility
and availability of effective controls.

Established medical-device risk-management principles provide an appropriate foundation for GenAI-enabled medical
devices. BioPhorum considers that applicable controls and evidence are most appropriately driven by the risk created by
use of the GenAI output within the complete device system, rather than by the model type or broad technology category
alone.

Risk classification should consider not only consequence and autonomy but also the principal sources of risk. These
sources include data quality, representativeness, model design, system integration, deployment environment and user
interaction patterns, together with the failure modes that may arise from those sources.

BioPhorum recommends that FDA consider whether each risk tier should be accompanied by defined examples of
appropriate design controls, supporting evidence, human oversight and boundaries for restricted or enhanced-assurance
use.

BioPhorum considers that risk assessment should occur at the level of the complete deployed system, including the role
of the GenAI function, surrounding software and workflow, intended user, deployment environment and the way an
output informs or initiates downstream action. The same model may therefore justify materially different controls and
evidence in different clinical or operational contexts.

3.1.1 Scaling controls and evidence with risk
Consistent with FDA’s existing credibility assessment framework, risk classification should explicitly consider both model
influence and decision consequence. The level of evidence, validation and monitoring should increase as either the
influence of the AI output on decision-making or the consequence of an incorrect outcome increases.

BioPhorum considers that progression should occur only against agreed evidence that the system, its safeguards and the
surrounding workflow remain safe and reliable within a defined validated risk envelope. Increasing autonomy,
consequence or difficulty of detecting failure may warrant stronger design controls, more robust validation, enhanced
monitoring and clearly defined escalation pathways.

The degree of post-deployment adaptiveness should be evaluated separately from the degree of autonomy. A system
may have limited authority to act but still present material lifecycle risk if it continues to learn, retrain, evolve or
otherwise modify its behavior after deployment.

Systems with adaptive capability should therefore be subject to controls proportionate to the nature, frequency and
potential effect of those changes. Such controls may include enhanced performance monitoring, drift detection,
configuration and change management, defined update boundaries, and risk-based revalidation before or following
implementation, as appropriate.

Industry considers that distinguishing between autonomy and adaptiveness would help ensure that evidence and control
expectations appropriately reflect both the system’s decision-making authority and its capacity to change in use.
Risk controls should follow the established hierarchy:

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 10
 • Inherent safety by design
• Engineered or deterministic controls
• Information and labeling used to address only those residual risks that cannot otherwise be adequately
controlled
Human oversight should not be treated as inherent safety by design.

The adequacy of the control strategy should be assessed by whether erroneous output can be detected, whether the
data, model and decision pathway leading to the output can be reconstructed and audited, and whether resulting actions
can be prevented, contained, reversed or otherwise mitigated.
Risk-proportionate oversight should include appropriate governance arrangements, including clearly defined
accountability, multidisciplinary review and lifecycle ownership for GenAI-enabled functions. The level of governance
should scale with risk in the same manner as technical controls and validation evidence.
Risk assessments should remain active throughout the product lifecycle.
BioPhorum considers that a tiered framework would be most credible if controls, evidence and oversight at a higher risk
tier are at least as stringent as those applied at a lower tier, unless a documented rationale supports an alternative
approach. Incorporating this principle into any future tiering framework could support internal consistency, and
manufacturers could apply the same consistency check to their own classification schemes.

3.2 System-level assurance
BioPhorum considers that safety and effectiveness are most credibly demonstrated at the level of the complete GenAIenabled medical-device system, rather than inferred from isolated model-performance metrics. This reflects the way
outputs are generated, controlled, communicated and acted upon within the overall device architecture and intended
context of use.

Evaluation should address:

• The underlying model
• Surrounding software
• Intended use
• User interaction
• Deployment environment
• The way outputs inform, influence or initiate downstream actions
The device-level architecture, including the harness, deterministic controls, routing and escalation logic, can help
constrain inputs and outputs, enforce required output structures, route low-confidence or out-of-scope results
appropriately, and support review, human intervention or escalation when required.

Assurance should consider the effectiveness of these controls in limiting model influence and preventing the AI output
from becoming the sole determinant of consequential actions.

Conventional verification and validation remain appropriate for deterministic components and defined control logic.
Variable generative behavior requires complementary evaluation against intended-use tasks, representative operating
conditions, foreseeable misuse scenarios and known failure modes.

BioPhorum considers that this evaluation should demonstrate whether the integrated system performs reliably within its
validated operating envelope and whether implemented safeguards remain effective under expected real-world

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 11
conditions. The depth of assurance, validation evidence and monitoring should be proportionate to the decision
consequence, model influence and effectiveness of risk controls.

System-level assurance should demonstrate that safeguards remain effective when an underlying model is updated,
reconfigured or replaced. Model-swap testing and device-level performance testing may help demonstrate that the
harness, deterministic controls, routing and escalation logic maintain safety and performance across defined model
variability or replacement. Re-evaluation should be proportionate to the potential effect on intended use, risk controls,
system interactions and validated performance.

For retrieval-augmented architectures, the source corpus, retrieval index, embedding model, system instructions and
associated knowledge resources can materially influence the output presented to the user. From an industry perspective,
these elements may therefore need to be identified, controlled and assessed as part of the deployed device
configuration, particularly where the retrieved content supports regulated decisions or downstream actions.

Where changes to these elements could affect safety, performance, traceability or the effectiveness of controls,
BioPhorum considers that risk-based impact assessment and change control would provide a proportionate assurance
mechanism.

Persistent memory, accumulated user or system context, dynamic knowledge stores and other mechanisms that
influence future outputs may materially affect device behavior even where the underlying model is not retrained,
updated or replaced. BioPhorum therefore considers that these elements should be considered within configuration and
change-control arrangements proportionate to their influence on the device function.

In practice, this may mean defining what information can be retained, how it is validated and versioned, how long it
remains applicable, and the conditions under which it is corrected, refreshed or reset. The rationale is that these
mechanisms may alter device behavior, introduce bias or outdated information, or create drift without an identifiable
model-version change.

Post-deployment monitoring can help detect performance degradation, drift and other conditions that may affect system
behavior.

3.3 Traceability
GenAI outputs may require proportionate citation, provenance and traceability so that users, reviewers and
manufacturers can assess the basis and suitability of a response and investigate failures, particularly where an output
contributes to a regulated decision or downstream action. The record should distinguish source content, generated
content and subsequent human or system actions.

Traceability should extend across the end-to-end device system and lifecycle, including:

• The model and version
• Relevant data and sources
• Prompts and system instructions
• Configuration and system state
• Deterministic controls
• Routing and escalation logic
• Human review or intervention

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 12
Supplier changes, postmarket signals, complaints, monitoring results and corrective actions should be linked to the
affected configuration to support root-cause investigation and demonstrate whether the device remained within its
validated operating envelope.

Because GenAI outputs can be affected by prompts, retrieval sources, inference settings and surrounding controls as well
as by the model itself, BioPhorum considers that defining the validated state as an identified system configuration, rather
than a model version alone, would provide a more reliable basis for traceability and investigation. The configuration
record could include, as applicable:

• The model and version
• Relevant data and sources
• Prompts and system instructions
• Configuration and system state
• Deterministic controls
• Routing and escalation logic
• Human review or intervention
Each released configuration should have a unique and resolvable identity so that an output can be attributed to the
configuration that produced it.

A citation generated by a model remains generated content and may not, by itself, demonstrate source support. Where
citations or source references are relied upon as a control, BioPhorum considers that deterministic or independently
validated verification against the source corpus can provide stronger assurance. Where outputs constitute or inform
regulated records, applicable data-integrity and audit-trail expectations, including 21 CFR Part 11 where relevant, may
also apply. In such circumstances, traceability becomes part of the assurance case rather than an optional design feature.

3.4 Reversibility and downstream safeguards
Reversibility should be considered in relation to severity of consequence and the time available to intervene. An output
that supports a reviewable and recoverable action presents a different risk from one that may trigger an immediate or
difficult-to-reverse consequence.

Lower-risk uses with readily detectable or recoverable failures may permit greater automation. High-risk therapeutic
decisions with limited opportunity for correction should remain subject to tighter system constraints and more
purposeful human oversight.

Higher-risk uses should include downstream safeguards that prevent, interrupt or contain harmful action after a GenAI
output is produced. Depending on the intended use, these may include bounded decision authority, deterministic
eligibility or plausibility checks, confirmation steps, escalation to a qualified reviewer, and the ability to stop, override or
reverse an action.

The device-level architecture may need to prevent operation outside the intended use or validated risk envelope,
particularly where outputs are uncertain, controls are not satisfied or timely human intervention is unavailable.

3.5 Visible failure and graceful degradation
Silent, confident and unsupported failure is a central GenAI risk because a system may continue to produce fluent and
plausible outputs when performance has degraded or the response is incorrect, incomplete or unsupported. BioPhorum
considers that systems should be designed, where feasible and proportionate, to detect uncertainty, unsupported

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 13
output, format or constraint failure, anomalous behavior and performance degradation, and to make these states
apparent to the intended user.

When a relevant failure condition is detected, the system should withhold or qualify the output, notify the user or
escalate to an appropriate reviewer. Evaluation should measure silent-error rates, unsupported responses and both
directions of clinically relevant error. It should also demonstrate that detection, notification and escalation controls
operate under representative and challenge conditions.

BioPhorum considers that distinguishing visible non-performance from silent critical failure would support more
proportionate evaluation. Safe-failure approaches can support graceful degradation and timely human hand-off, while
excessive refusal may compromise the intended clinical utility.

3.6 Lifecycle evidence
Evidence should accumulate across development, preclinical evaluation, clinical use and commercial deployment.
Premarket evaluation should identify residual uncertainty, define the evidence to be collected in use, and establish how
emerging issues will be detected, investigated and managed. Postmarket evidence should verify continued safety and
performance within the approved operating envelope and inform reassessment of the original risk analysis.

Premarket evidence should define expected performance, known limitations, residual uncertainties and the specific
monitoring commitments required to manage them. A pre-agreed progression plan may specify the evidence needed to
expand the operating envelope, increase autonomy or reduce human oversight. Progression should depend on meeting
defined safety and performance criteria under representative conditions rather than on elapsed time or deployment
volume alone.

This approach may draw on the logic of a Predetermined Change Control Plan (PCCP), with an agreed operating envelope,
staged evidence thresholds and defined reassessment triggers. Triggers should include material changes to:

• Model
• System Configuration
• Intended Use
• User Population
• Deployment Environment
• Safeguards
• Monitored Performance
These triggers should determine when additional validation, restored oversight, restricted use or regulatory engagement
is required.

3.7 Use case-specific metrics
BioPhorum considers that a framework in which manufacturers define and justify monitoring metrics, baselines,
thresholds and cadence for the specific intended use may provide a more proportionate approach than prescribing a
universal performance threshold.

Metrics should reflect the clinical or operational consequence of error, detectability of failure, role of the GenAI output
and surrounding workflow. Generic accuracy alone is insufficient. Manufacturers should justify measures against the
intended use, risk, user, environment, availability of reliable ground truth, volume of use and feasibility of obtaining
representative postmarket data.

Evaluation should include:

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 14
 • Unsupported-output and silent-failure rates
• Confidence calibration
• Latency and timeliness
• Low-confidence routing and escalation frequency
• Compliance with required output structures
• Performance under foreseeable variation
• Performance after model or system change
For each metric, manufacturers should define its relevance, baseline, acceptance criteria, trigger threshold, assessment
cadence and linkage to investigation, escalation, corrective action, regulatory reporting and change control.

Where a GenAI-enabled function provides citations or source references, monitoring may need to distinguish citation
presence from citation support. A citation-support metric can assess whether the cited source exists, is authoritative and
applicable to the intended use, and contains evidence that substantively supports the specific claim, recommendation or
decision presented. BioPhorum considers that citation accuracy should not be inferred from a plausible reference or left
solely to end-user verification, particularly where the output may influence dosing, diagnosis or another consequential
decision.

Where citations are relied upon as a risk control, BioPhorum considers that source-to-claim alignment should be verified
by deterministic or independently validated methods where feasible, with unsupported, mismatched or fabricated
citations captured within the unsupported-output and silent-failure metrics and linked to predefined investigation and
change-control thresholds.

Where a monitored metric supports a PCCP, BioPhorum considers that it should define the applicable threshold, the
changes permitted in response, the evidence required before implementation and the conditions that instead require
additional revalidation or regulatory engagement. Exceeding a threshold should not by itself authorize an adaptive
change. The response should remain within the pre-specified change protocol, validated risk envelope and associated
verification and monitoring commitments.

3.8 Allocation of responsibility
Reliance on foundation models, cloud services, external data sources or supplier-managed updates does not remove the
legal manufacturer’s responsibility to demonstrate control of the marketed device system. The regulatory framework
should nevertheless recognize which party possesses the information, authority and technical capability to perform each
control activity.

BioPhorum would welcome FDA guidance that clarifies responsibility for control and validation of each component and
interface, supplier evidence and assurances, and receipt and assessment of supplier changes. Clarity would also be
helpful on expectations for monitoring real-world performance, drift and emerging risk, and for investigation, escalation,
corrective action and regulatory reporting.

Supplier contracts, controls and deployment agreements should provide timely access to the information needed to
discharge these responsibilities. Accountability should remain explicit when a third-party model changes, information is
unavailable, responsibilities overlap or a GenAI output contributes to harm, so that no safety-critical activity is left
unassigned or obscured by the supply chain.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 15
4 Response to FDA’s four discussion areas
The following sections respond directly to the four principal areas identified in FDA’s discussion paper:

1. Risk assessment and human oversight
2. Premarket evaluation
3. Postmarket monitoring
4. Additional regulatory considerations

4.1 Risk assessment and human oversight

Regulatory position Risk assessment for GenAI-enabled medical devices should remain grounded in established
medical-device risk-management principles but should evaluate the complete deployed
system. The nature and strength of controls and evidence should reflect intended use,
severity of consequence, degree of autonomy, detectability of failure, reversibility and
availability of effective controls.

Recommendation to FDA could provide a flexible, risk-proportionate framework, supported by representative
FDA examples, showing how device-level architecture, purposeful human oversight, traceability,
reversibility and downstream safeguards influence premarket evidence, deployment
constraints and lifecycle monitoring.

Any streamlined pathway for lower-risk GenAI functions should be bounded by explicit exclusion criteria (for example
autonomous action, direct therapeutic impact or inability to detect failure) rather than by a general risk judgment alone.

Additional FDA clarity on what constitutes sufficient controls and evidence at each stage would help manufacturers
determine when implementation should be constrained, oversight increased or progression withheld. Clearer
expectations would reduce inconsistent review outcomes and avoid unnecessary delay without lowering the standard of
assurance for higher-risk uses.

Additional FDA clarity on the GenAI-specific hazards relevant to an ISO 14971 risk analysis, such as unsupported content,
instruction-following failures and unsafe tool actions, would help support consistent implementation.

4.1.1 Human oversight
Human oversight should be purposeful and risk-based rather than treated as a universal safeguard. Its design should
reflect the decision being supported, the likelihood that an error will be detected, user competence, the time available to
intervene, automation bias, workload and the consequences of accepting, rejecting or failing to review an output.

Early deployment may place a human in the loop for consequential decisions, low-confidence outputs and detection of
silent or unsupported failures. Any later reduction in oversight should be supported by evidence that reviewers’ safety

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 16
functions have been replaced by effective device-level controls and that the remaining oversight model is feasible under
representative conditions of use.

Human oversight design

For systems producing very high volumes of outputs, review of every output may be impractical and may itself create
workload-related risk. Industry considers that layered controls, representative sampling, automated verification,
confidence thresholds, exception-based review and defined escalation pathways can provide a more scalable approach
when selected and justified for the intended use.

The manufacturer should demonstrate that human reviewers have sufficient information, time, authority and
competence to intervene effectively, understand the system’s limitations and resist automation bias. Nominal review
should not be credited as a risk control where the reviewer cannot realistically identify, challenge, override or escalate an
unsafe output.

Oversight is only effective if it is independent of the output it reviews. By analogy to the established quality principle that
an author should not be the sole approver of their own work, output generated by a model should not be verified solely
by the same model or configuration, or by a model from the same family with correlated failure modes. Where a second
AI system is used to verify, filter or monitor outputs, it should be qualified for that role against its own intended use, and
its failure modes should be included in the risk assessment. Human-factors evaluation of the oversight step, consistent
with IEC 62366-1, should include representative workload, time pressure and the risk of automation bias.

4.1.2 Reducing human oversight
The risk of automation bias tends to increase as system performance improves, because reviewers who rarely encounter
errors become less attentive. Good observed performance alone should therefore not be sufficient grounds for reducing
human oversight; a reduction should also be supported by evidence that the remaining safeguards detect the failures the
reviewer would otherwise have caught.

Closed-loop medical-device development, including the staged evolution of insulin-pump decision support, offers a useful
analogue. The lesson is not that critical decisions can never be automated, but that autonomy should expand only within
a defined validated risk envelope and against agreed evidence that safety, reliability and escalation controls remain
effective.

4.1.3 Foreseeable misuse and unintended use
Risk assessment should address reasonably foreseeable reliance on apparently confident outputs, use outside the
intended population or workflow, automation bias, attempts to bypass safeguards, inappropriate prompting, and use in
contexts where connectivity, data quality or human review is inadequate.

Controls should anticipate both incorrect outputs and the possibility that users treat incomplete, unsupported or out-ofscope outputs as authoritative. The interface, workflow and access controls should discourage reliance outside the
intended use and make material limitations and uncertainty apparent at the point of use.

Testing should include realistic misuse scenarios, workflow pressures and boundary conditions, including prompt
variation, adjacent or off-scope questions, attempts to circumvent constraints, degraded data or connectivity, and
situations in which timely human review is unavailable. The scope should be risk-based and cover reasonably foreseeable
patterns without requiring manufacturers to anticipate every possible prompt or hypothetical interaction. Use-related
risk analysis consistent with IEC 62366-1 provides an established basis for this assessment.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 17
Labeling and user information should communicate limitations clearly but should not substitute for inherent safety by
design or engineered controls where foreseeable misuse could result in significant harm. Safeguards should be designed
to prevent or contain unsafe action even where users misunderstand, overlook or deliberately test the stated limitations.

4.2 Premarket evaluation
Regulatory position Premarket evaluation should demonstrate the safety and effectiveness of the complete
GenAI-enabled medical-device system for its intended use. Conventional verification and
validation should apply to deterministic components and controls, while variable generative
behavior should be evaluated through risk-justified statistical methods, representative
scenarios, repeated sampling, challenge testing, failure-mode analysis and human-factors
evidence.

Recommendation to FDA could clarify system-level evaluation strategies for variable and open-ended outputs,
FDA including expectations for sampling, acceptance criteria, benchmark integrity, subgroup and
environment testing, validation of automated evaluators, configuration control and the
documentation of residual uncertainty and associated postmarket monitoring
commitments.

Submissions should define the intended-use risk category, complete device-level architecture, model and supplier
dependencies, validated operating envelope, human-oversight model, foreseeable failure modes, escalation pathways
and the evidence supporting each control. Documentation should explain how deterministic and generative components
interact and what assumptions underpin safe performance. It should also explain which uncertainties remain, and how
the system is prevented from operating outside its intended use or validated risk envelope.

Testing and evaluation

Testing should be use-case-specific and proportionate to risk.

Testing should include:

• Representative populations, users and environments
• Edge cases and boundary conditions
• Adversarial, misuse and foreseeable-unintended-use scenarios
• Repeatability and response variability
• Unsupported-output and silent-failure detection
• Confidence thresholds and calibration
• Latency and timeliness
• Drift sensitivity
• The clinical or operational consequences of an unavailable, withheld or escalated output
Human-factors evaluation should confirm that users can understand limitations, recognize relevant failure states and
intervene effectively.

Evidence generation and validation

For variable outputs, evidence should be statistical rather than anecdotal.

For variable outputs, BioPhorum considers that a risk-justified evaluation approach could include:

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 18
 • Repeated sampling of the same input to characterize output variability
• Pre-specified and risk-justified sample sizes, agreement measures and confidence bounds for critical error rates
(including silent and unsupported outputs)
• Documentation of the inference configuration under which performance was established
• Separation of development, tuning and final test data
Configuration settings intended to reduce variability (for example sampling temperature) should not be assumed to
guarantee deterministic behavior. Acceptance criteria should be derived from the consequence of error for the intended
use, not from the performance the model happens to achieve.

Where automated evaluators, including model-based graders, are used to score outputs in premarket testing or
postmarket monitoring, their agreement with qualified human assessment should itself be established for the context of
use.

For benchmarking, performance should translate to real-world safety and clinical or operational utility, not merely a high
benchmark score. Independent or sequestered test sets should be considered, particularly where public or synthetic
datasets may introduce bias, leakage or contamination, including where the manufacturer cannot verify whether public
benchmarks were included in a foundation model’s training data. Test-set provenance and representativeness should be
documented.

Additional FDA clarity would be helpful on how premarket expectations apply to GenAI-enabled technologies used only in
clinical studies or drug-development contexts. The applicable evidence pathway may differ from that for a marketed
medical device. However, the sponsor should still justify the intended use, level of control, participant or user risk, data
integrity protections and the transition criteria for any later clinical or commercial deployment.

Two situations should be distinguished. For a GenAI-enabled investigational device, additional clarity on how these
expectations apply under 21 CFR Part 812, including what evidence is proportionate for early feasibility use and what
should be demonstrated before transition to a marketing application, would support consistent implementation.

Where a GenAI-enabled tool is used within a clinical investigation without being the investigational product, or generates
evidence intended to support a drug or biologic regulatory decision, BioPhorum would welcome clarity on how the
device-oriented approach relates to the risk-based credibility framework in FDA’s January 2025 draft guidance on the use
of AI to support regulatory decision-making for drug and biological products. The framework assesses model risk from
model influence and decision consequence within a defined context of use. Clarity on how the two approaches align
would support consistent implementation.

A consistent approach across Centers would avoid parallel and potentially divergent evidence expectations for the same
technology.

4.3 Postmarket monitoring
Regulatory position Postmarket monitoring should be planned during premarket development and operated as
an active component of the lifecycle assurance case. Monitoring should remain linked to
the intended use, validated operating envelope, residual uncertainty and safety-critical
controls, using justified, use case-specific metrics, thresholds, cadence and escalation
pathways.

Recommendation to FDA could provide a structured but flexible framework identifying the categories of
FDA information and action manufacturers should define, including data quality, metric

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 19
 definitions, baselines, thresholds, assessment cadence, investigation and intervention rules,
regulatory reporting and linkage to corrective action and prospective change control.

Where reliable ground truth is delayed or unavailable, BioPhorum considers that appropriately justified surrogate signals,
such as override, escalation and user-correction rates, could provide useful evidence, provided that their relationship to
safety and performance is defined and periodically reassessed.

Monitoring should cover:

• Harmful, incorrect and unsupported outputs
• Silent failures
• Withheld or unavailable outputs
• Low-confidence routing and escalation frequency
• Latency and timeliness
• Complaints and user feedback
• Model, data and population drift
• Changes in prompting, workflow or use patterns
• Performance across relevant populations and environments
• Supplier or model changes
• Whether the complete device system continues to operate within its validated risk envelope and intended
controls
Premarket and postmarket assurance should consider both clinical criticality and detectability of failure. Clinically critical
silent failures require stronger premarket evidence and should not rely on postmarket monitoring as a substitute.

Early deployment may generate apparent performance complaints because the system withholds an output, routes a
low-confidence result for human review or otherwise fails visibly. These events should be distinguished from unsafe silent
failures, in which plausible-looking outputs continue despite degradation or error. Both should be measured. Silent failure
may create direct safety risk, while persistent visible non-performance may undermine the intended clinical or
operational benefit of the device.

Postmarket assessment should consider the complete deployed system, not only the underlying model. Monitoring
results should be interpreted alongside model behavior, configuration record, routing and safeguard performance, user
interaction, human interventions, downstream outcomes and supplier changes. This means that new hazards, changed
detectability or increased consequences can be identified and traced to the affected device configuration.

Predetermined thresholds should trigger proportionate action, including:

• Investigation
• Corrective or Preventive Action
• Increased Human Oversight
• Restriction of Use
• Rollback to a Previously Qualified Model or Configuration
• Suspension
• Correction or Removal
• Regulatory Engagement

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 20
The monitoring plan should define who reviews each signal, how quickly it is assessed, what evidence is required to close
an investigation and how decisions link to complaint handling, change control and regulatory reporting (for example
under 21 CFR Parts 803 and 806).

A supervisory agent may support postmarket monitoring, but it should be independently qualified and validated for that
role. Otherwise, reliance on the supervisory agent may transfer the trust and failure-detection problem to another AI
system.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 21
4.4 Additional regulatory considerations

Regulatory position Additional regulatory expectations should address the complete lifecycle and supply-chain
context of GenAI-enabled medical devices. Requirements should provide proportionate
transparency and explainability, maintain legal-manufacturer accountability for third-party
and foundation-model components, support coordinated oversight of combination
products, and integrate data governance, privacy, cybersecurity, change management and
Total Product Lifecycle controls. These expectations should align with existing medicaldevice and international frameworks while addressing the distinctive risks of variable
outputs, rapid component change and limited access to supplier-controlled technical
information.

Recommendation to BioPhorum would welcome FDA clarity on:
FDA
• The information, controls and evidence expected across users and risk levels
• The responsibilities of legal manufacturers and third-party suppliers
• Cross-Center pathways for combination products
• Requirements for data governance, privacy and GenAI-specific cybersecurity
• Prospective assessment of model and configuration changes
• Lifecycle monitoring, continuity and retirement
• Alignment with recognized standards and international regulatory approaches

4.5 Transparency and explainability
Transparency should be proportionate to the intended user and risk. Regulators, healthcare professionals, patients and
other users may require different information, but each should be able to understand the system’s role, intended use,
material limitations, confidence or uncertainty, known failure modes, level of autonomy, update status and escalation
behavior to the extent relevant to their decisions.

Full explanation of a foundation model’s internal reasoning may not be technically feasible or necessary.

Device-level evidence should instead show:

• How outputs are grounded, checked, traced and constrained
• How unsupported or low-confidence outputs are identified
• What source or provenance information is available
• How users can recognize limitations and invoke review or escalation

4.6 Foundation models and third-party components
Where direct qualification of a third-party or frontier model is not feasible, model-swap testing and device-level
performance testing may help demonstrate that the harness, deterministic controls, routing and escalation logic maintain
safety and performance across defined model variability or replacement. The acceptable approach should be justified by
the intended use, risk and degree of supplier control.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 22
Manufacturers should be able to pin and verify the exact model version in use and to detect provider-side changes made
under an unchanged identifier. Supplier arrangements should secure access to the information needed for audit,
inspection and investigation, advance notice of version deprecation or service withdrawal. They should also include a
documented continuity or exit strategy that allows the manufacturer to maintain, revert and qualify a replacement for or
safely retire the affected device function.

Reliance on supplier evidence should be risk-based and justified by an assessment of the supplier.

4.7 Combination products
Many GenAI-enabled functions of interest to the biopharmaceutical industry will be constituent parts of combination
products, for example software that supports dose preparation, injection guidance or adherence for a drug-device
combination.

Particular consideration is needed where a GenAI-enabled function is used within a clinical investigation of a drug or
biologic but is not intended to be developed or marketed as a standalone medical-device product. Such functions may
collect or process study data, support modeling, derive an endpoint, inform trial conduct or contribute evidence used in a
regulatory decision.

The applicable pathway should reflect the function’s intended role, its influence on participant safety and regulatory
decision-making, and the controls available within the clinical-investigation environment, rather than relying solely on
whether the technology could meet a device definition.

Investigational use should not create an assumption that a GenAI-enabled function must first satisfy the evidence
expectations applicable to a marketed medical device before it can be used to investigate a drug or biologic. Defined
protocols, controlled access, qualified users, enhanced oversight and restricted deployment may provide additional
safeguards that support a proportionate investigational evidence pathway.

Additional FDA clarity would be helpful on the evidence and controls expected where a GenAI-enabled function processes
study data, derives an endpoint or supports modeling, and on how these uses should be distinguished from functions that
directly inform diagnosis, treatment or another consequential decision for an individual participant.

Coordination across CDRH, CDER and CBER will be essential where the GenAI-enabled function supports a drug or biologic
program or forms part of a combination product.

BioPhorum would welcome greater clarity on how lead and consulting review responsibilities might be coordinated
across CDRH, CDER and CBER, how the applicable model-credibility and device-assurance frameworks might be aligned,
and how duplicative or divergent evidence expectations for the same technology could be avoided. International
alignment, including consideration of European Medicines Agency (EMA)’s work on AI across the medicinal-product
lifecycle, would further support consistent development expectations.

BioPhorum would welcome clarity on:

• How GenAI-specific expectations apply when the primary mode of action is attributed to the drug or biologic
and a drug Center leads the review, and how evidence expectations are coordinated between the lead and
consulting Centers
• Whether, and through which mechanism, predetermined change control can be used for a GenAI-enabled
device constituent part of a drug-led combination product, given that current PCCP guidance is framed around
device marketing submissions

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 23
 • How GenAI-related design controls integrate with the streamlined cGMP approach for combination products
under 21 CFR Part 4
• How risk assessment should address interactions between the GenAI function and the drug constituent, for
example dose-related consequences of an erroneous output
• How evidence expectations differ where a GenAI-enabled function is used only within a controlled clinical
investigation and is not intended for standalone commercial use
• How CDRH, CDER and CBER will coordinate classification, review responsibility and evidence expectations where
the function supports a drug or biologic investigation or regulatory decision

4.8 Data governance, privacy and cybersecurity
BioPhorum would welcome clarity on expectations for the quality, provenance, representativeness and governance of
data used to develop, validate and operate GenAI-enabled medical devices.

Sponsors should address patient privacy, data protection, bias and population coverage, data drift, retention and access
controls, and traceability of prompts, outputs and supporting sources, with requirements proportionate to the intended
use and risk.

Traceability expectations should be reconciled with data-minimization obligations, because prompts and outputs may
contain personal health information. FDA should clarify the minimum record set needed for investigation and how it can
be retained without unnecessary storage of personal health information.

Cybersecurity controls should address prompts, outputs, interfaces, retrieval sources, connected services and external or
cloud-hosted models. The manufacturer should demonstrate how risks are managed when data, services or model
updates occur outside its direct control. This should include supplier access controls, timely notification, impact
assessment, monitoring, incident response, service continuity and contingency arrangements.

GenAI-specific threats, including direct and indirect prompt injections through retrieved or user-supplied content,
jailbreaking of safeguards and leakage of sensitive information through outputs, should be addressed within the
premarket cybersecurity risk assessment and threat model.

4.9 Change management
Changes to models, configuration records, prompts, retrieval systems, data sources, supplier services, deployment
environments and intended use may affect whether the device remains within its validated risk envelope and whether
existing controls remain effective. Change assessment should therefore consider potential effects on performance,
detectability of failure, degree of autonomy, purposeful human oversight, traceability, cybersecurity and downstream
safeguards.

A risk-based change framework can help identify which changes may be implemented under a PCCP or established
quality-system process, which may require enhanced monitoring or documented retrospective review, and which may
require prior regulatory engagement. Particular attention is needed when a supplier makes or deploys a change before
the manufacturer can fully assess its effect on the medical-device system.

The established approach for urgent cybersecurity patches may provide a useful analogue for security-driven changes
only, where clear guardrails, notification thresholds and time limits apply. BioPhorum considers that this rationale should
not be extended automatically to changes that alter model behavior or intended-use performance. Changes affecting
high-risk functions, intended use, detectability, autonomy, model behavior, supplier dependency or system safeguards
may warrant prompt impact assessment and proportionate revalidation, monitoring or regulatory engagement.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 24
BioPhorum considers that assurance of a changed model or configuration should generally be completed before it is used
for its intended purpose, unless a justified urgent safety or security rationale supports a controlled alternative.

Postmarket monitoring complements, but does not substitute for, prospective verification of a change where that change
may affect the validated risk envelope or effectiveness of controls.

Changes to retrieval corpora or knowledge bases can alter device performance without any change to the underlying
model. BioPhorum considers that the potential impact of such changes should therefore be assessed within the device
change-control process, with the level of review proportionate to intended use, risk and the role of the retrieved content
in the device function.

Simplified or expedited change pathways within a quality system may be appropriate for some bounded changes, but
BioPhorum considers that they should not be assumed to apply automatically to changes affecting a GenAI function.
Eligibility for such pathways should be justified against the validated risk envelope and the potential effect on model
behavior, controls, detectability and downstream safeguards.

An inventory of deployed GenAI functions, with their intended use, risk tier and current configuration record, would
support consistent change assessment and postmarket oversight.

4.10 Total Product Lifecycle
The regulatory lifecycle approach should connect intended use, design controls, risk management, verification and
validation, human factors, supplier management, deployment controls, premarket evidence, postmarket monitoring,
change management, corrective action and eventual retirement of the device. This includes retention of the records
needed to investigate outputs generated before retirement.

The assurance case should evolve as real-world evidence accumulates while preserving traceability to the original
hazards, assumptions, controls, validated operating envelope and acceptance criteria. Monitoring outputs, complaints,
investigations, supplier changes and corrective actions should feed back into risk management and determine whether
the device remains suitable for its intended use.

4.11 Relationship with other regulatory and policy work
Alignment between the GenAI-enabled medical-device framework and established medical-device risk management,
software assurance, Good Machine Learning Practice (GMLP), PCCPs, cybersecurity and quality-system expectations
would help support consistent implementation. Relevant work on digitally derived endpoints, patient-focused drugdevelopment principles, the FDA TEMPO pilot, postmarket evidence and established AI risk-management guidance could
be used where directly applicable to the final recommendations.

Consistency with international approaches would reduce duplication, support a common evidence language and improve
the feasibility of global development and lifecycle management. Alignment should not mean importing wider AI policy
wholesale. The framework should remain focused on medical-device intended use, safety and effectiveness.

Any aligned approach should still address the distinctive characteristics of GenAI, including variable and open-ended
outputs, supplier and foundation-model dependency, limited access to underlying technical information, rapid
component change, automation bias and the possibility of confident or silent failure.

Recent standards and regulatory work confirm that existing ML-specific instruments do not yet fully cover generative
models. ISO/TS 24971-2:2026, for example, excludes medical devices employing large language models or generative AI

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 25
from its scope, while the European Commission’s draft EU GMP Annex 22 excludes probabilistic and generative models
from critical GMP applications.

FDA’s framework would therefore address a recognized gap. FDA is encouraged to work through the International
Medical Device Regulators Forum (IMDRF)

so that GenAI-specific expectations build on the IMDRF GMLP guiding principles and remain interoperable with
applicable requirements under the EU Artificial Intelligence Act.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 26
5 Conclusion
BioPhorum welcomes FDA’s early engagement on GenAI-enabled medical devices. Established medical-device risk
management, software assurance and Total Product Lifecycle principles provide a sound regulatory foundation. However,
targeted GenAI-specific expectations are needed where variable and open-ended outputs, rapid component change,
reliance on third-party foundation models and plausible but unsupported responses create distinctive assurance
challenges.

FDA action would be most valuable in five areas:

1. A risk-proportionate framework supported by representative examples.
2. Clear expectations for system-level evaluation of variable outputs.
3. Explicit requirements for configuration control, purposeful human oversight and lifecycle monitoring.
4. Clarity on supplier and third-party responsibilities while preserving the legal manufacturer’s accountability.
5. Coordinated, cross-Center treatment of combination products and clinical-investigation use, with clear lead and
consulting responsibilities across CDRH, CDER and CBER.
A clear, internationally aligned and operationally implementable framework would support consistent regulatory
submissions and review, enable proportionate control of change across the product lifecycle, and provide manufacturers
with a predictable basis for responsible innovation. This would protect patients while enabling timely access to GenAIenabled medical-device functions whose safety and effectiveness have been demonstrated for their intended use.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 27
Sources
This paper does not use formal in-text citations. The numbered sources below comprise published information considered
during the review and discussion process and drawn upon in developing the positions and recommendations presented.

No. Source

1 FDA, Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback, August 2026, Docket FDA-2026-N-7874.

2 FDA, Technology-Enabled Meaningful Patient Outcomes (TEMPO) for Digital Health Devices Pilot, launched
December 2025.

3 BioPhorum, Artificial Intelligence Risk Guidance for the Pharmaceutical Industry: Harmonizing Frameworks
for Practical Implementation, 2 June 2026.

4 Duke-Margolis Institute for Health Policy and FDA, Statistical Considerations for Digitally-Derived Endpoints
in Clinical Trials, virtual public workshop, 27 August 2026.

5 EMA, Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle (adopted
September 2024).

6 FDA, Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing
Submission Recommendations. Draft guidance, January 2025.

7 FDA, Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial
Intelligence-Enabled Device Software Functions. Final guidance, December 2024.

8 FDA, Health Canada and MHRA, Good Machine Learning Practice for Medical Device Development: Guiding
Principles (2021); Predetermined Change Control Plans for Machine Learning-Enabled Medical Devices:
Guiding Principles (2023); Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles
(2024).

9 IMDRF, Good Machine Learning Practice for Medical Device Development: Guiding Principles (N88).

10 FDA (CDER/CBER), Considerations for the Use of Artificial Intelligence to Support Regulatory DecisionMaking for Drug and Biological Products. Draft guidance, January 2025.

11 FDA, Computer Software Assurance for Production and Quality Management System Software, final
guidance, February 2026.

12 21 CFR Part 4 (combination products); 21 CFR 3.2(m) (primary mode of action); 21 CFR Part 820 (Quality
Management System Regulation); 21 CFR Part 812; 21 CFR Part 11.

13 ISO 14971:2019; ISO/TS 24971-2:2026 (application of ISO 14971 to machine learning; excludes LLMs and
generative AI from scope); AAMI TIR34971:2023.

14 IEC 62304; IEC 62366-1; IEC 81001-5-1; FDA, Cybersecurity in Medical Devices: Quality Management
System Considerations and Content of Premarket Submissions, final guidance, February 2026.

15 ISPE, GAMP 5: A Risk-Based Approach to Compliant GxP Computerized Systems, Second Edition (2022).

16 ISO/IEC 42001:2023; ISO/IEC 23894:2023.

17 Regulation (EU) 2024/1689 (Artificial Intelligence Act), as amended.

18 European Commission, EU GMP Annex 22: Artificial Intelligence. Draft for consultation, 2025.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 28
19 FDA, Digital Health Technologies for Remote Data Acquisition in Clinical Investigations. Final guidance,
December 2023.

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 29
Glossary and acronyms
Term or acronym Definition

Agentic system A system in which a generative model can select and invoke tools or initiate actions,
not only generate content

AI Artificial intelligence

CBER Center for Biologics Evaluation and Research

CDER Center for Drug Evaluation and Research

CDRH Center for Devices and Radiological Health

Configuration record The identified set of components and settings that together constitute the validated
state of a GenAI-enabled device

Context of use The specific role and scope of a model output in addressing a defined question or
decision

FDA United States Food and Drug Administration

Foundation model A model trained on broad data that can be adapted to a wide range of tasks, often
supplied by a third party

GenAI Generative artificial intelligence

GxP Good practice quality guidelines and regulations

HITL Human in the loop

IDE Investigational Device Exemption (21 CFR Part 812)

LLM Large language model

Operating envelope The defined range of inputs, users, environments and configurations within which
performance has been established

PCCP Predetermined Change Control Plan

PMOA Primary mode of action (21 CFR 3.2(m))

QMSR Quality Management System Regulation (21 CFR Part 820)

RAG Retrieval-augmented generation: an architecture in which the output is conditioned
on information retrieved from a defined corpus at the time of use

Silent failure An incorrect or degraded output presented without any indication of error or
uncertainty

TPLC Total Product Lifecycle

Unsupported output Generated content not supported by the input, the retrieved sources or established
fact (sometimes called hallucination)

© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 30
© BioPhorum Operations Group Ltd | 2026 Feedback to FDA on the regulatory approach for generative AI-enabled medical devices | 31