Evidence in Context
The FDA’s new consultation is not yet guidance. Its deeper significance is that safe generative AI in medicine will depend on more than market entry: it will require continuous, institution-level governance of the complete human–AI system.
**Author:** Dr Samer Al-Diri, MD, MSc, MPH
**Published:** 20 August 2026
**Article type:** Analytical review and policy commentary
**Evidence status:** FDA discussion paper; not guidance
**Open access:** <u>[CC BY-NC-ND 4.0](creativecommons.org/licenses/by-nc-nd/4.0/)</u>
On 18 August 2026, the US Food and Drug Administration’s Digital Health Center of Excellence, within the Center for Devices and Radiological Health, opened a public consultation on the possible regulation of generative-AI-enabled medical devices. The paper seeks feedback on risk assessment, premarket evaluation, postmarket monitoring, foundation models and agentic AI. Comments are invited until 19 October 2026.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)[[2]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-2)</u>
The document is **not draft or final guidance** and does not change FDA policy or establish new evidentiary requirements. It also leaves open whether the approaches discussed fit the FDA’s existing legal authority. Its purpose is to obtain early external input.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
**Executive judgement**
The paper matters because it reframes the regulatory problem. The question is no longer only whether a model performs well in testing. It is whether a configured generative system can remain clinically bounded, reliable, equitable and accountable throughout real-world use.
**What the FDA is—and is not—addressing**
The paper concerns **generative-AI-enabled medical devices**, not every use of generative AI in healthcare. Its focus is software functions that meet the medical-device definition and fall within the FDA’s regulatory remit; it is not a framework for generative AI as a technology in itself.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
That legal boundary is necessary but not equivalent to a safety boundary. A tool outside device regulation may still influence clinical behaviour, patient understanding, workforce decisions or operational risk. Institutional governance therefore extends beyond regulatory classification.
Table 1. FDA discussion areas and their implications for healthcare organisations
**FDA discussion area** **What the FDA is exploring** **Healthcare-governance implication** **Evidence status**
**Risk assessment** A possible two-axis heuristic combining the activity or independence of the device with the severity of harm if an output is wrong.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u> Governance intensity should reflect clinical consequence, autonomy, user expertise, reversibility and available safeguards. Exploratory FDA proposal; not a final framework.
**Premarket evaluation** A competency-based approach combining device benchmarking with clinical confirmation, tailored to intended use and risk.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u> Hospitals should evaluate the configured human–AI system in the local population and workflow, not rely on a headline benchmark alone. Exploratory FDA proposal; methods and thresholds remain unsettled.
**Postmarket monitoring** Periodic re-benchmarking, sample-based clinician review, degradation monitoring and possible supervisory agents.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u> Implementation requires continuous surveillance, change control, escalation thresholds and the ability to restrict or suspend use. Approaches under consultation; feasibility and proportionality require evidence.
**Foundation and agentic models** Voluntary foundation-model master files and additional scrutiny of systems that plan or execute multi-step actions.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u> Procurement must address model dependencies, updates, permissions, auditability, rollback and accountability across organisations. Emerging regulatory considerations; not established requirements.
**Source:** Author’s synthesis of the FDA discussion paper. Organisational implications are analytical interpretation and do not represent FDA guidance.
**Why generative AI-enabled medical devices create a different evaluation problem**
Generative systems can accept open-ended inputs, conduct multi-turn conversations and produce variable outputs. Their behaviour may depend on prompts, retrieval sources, guardrails, interfaces and third-party foundation models. Some may also change after deployment. These characteristics do not make generative AI inherently unsafe, but they make it difficult to define safety through a single static test.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
**1. Risk should reflect both action and consequence**
The FDA presents a possible two-axis heuristic. One axis reflects what the software does and how independently it acts—from non-directive information to action-directing or autonomous behaviour. The other reflects the severity of harm if a user relies on an incorrect output.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
This is a useful shift from technological novelty to clinical consequence. An inaccurate explanation of a minor symptom is not equivalent to an incorrect recommendation affecting insulin dosing. A function used under continuous specialist supervision also presents a different risk from one acting without meaningful review.
The paper further recognises that “action-directing” is a continuum. A conversation may begin with general information and gradually become a personalised recommendation. Risk should therefore be assessed across realistic conversational pathways rather than isolated answers.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
**In my assessment**, the two-axis model is a strong organising concept but not a complete operating model. Reversibility, time pressure, downstream safeguards and traceability to primary evidence may materially change risk; the FDA itself asks whether these dimensions should be added.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
**2. Competency-based evaluation may be more useful than one accuracy score**
The FDA explores a competency-based approach comprising non-clinical benchmarking followed by clinical confirmation. Crucially, the final user-facing device in its intended configuration—not merely its underlying model—would be evaluated.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
Potential domains include safety-critical recognition, scope control, uncertainty communication, clinical knowledge, quantitative reasoning, communication quality, robustness, subgroup performance and, where relevant, agentic capabilities. Clinical confirmation might range from retrospective testing and shadow deployment to standardised-patient interactions, independent adjudication or prospective studies, depending on risk.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
> **The more useful question is not “How accurate is the model?” but “Is this configured system sufficiently competent for this purpose, population, workflow and level of consequence?”**
The analogy with professional competence must remain limited. A device has no ethical duty, professional judgement or personal accountability. “Competency” should describe a rigorous evidence architecture, not imply that a machine inherits a clinician’s responsibilities.
**3. Postmarket monitoring cannot be an afterthought**
Because open-ended outputs, deployment conditions and technical components can change, premarket evidence may not capture every real-world risk. The FDA therefore asks about periodic re-benchmarking, sample-based clinician review, performance-degradation monitoring and possible use of machine-based supervisory agents.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
This consultation does not emerge in isolation. It extends the FDA’s broader total-product-lifecycle approach to AI-enabled device software and its established framework for predetermined change control plans covering specific, prospectively defined modifications.<u>[[3]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-3)[[4]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-4)</u>
**In my assessment**, greater reliance on postmarket evidence may be defensible where harm is limited or reversible, monitoring is reliable, signals can be detected quickly and the system can be paused or rolled back. It is harder to justify for high-consequence, autonomous or time-critical functions.
Monitoring should extend beyond technical accuracy. Health systems should examine clinical outcomes, near misses, escalation failures, subgroup performance, automation bias, user workarounds, workflow burden and unintended effects on access or resource use. Earlier FDA work on real-world AI-device performance similarly recognised that changes in patient populations, clinical practice, infrastructure and user behaviour can alter performance after deployment.<u>[[5]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-5)</u>
**4. Foundation models and agentic AI widen the accountability chain**
Many generative medical devices may depend on third-party foundation models that influence refusal behaviour, content policies, output formatting, version control and other safety-critical characteristics. The FDA is seeking views on voluntary Foundation Model Device Master Files that could provide confidential information to reviewers. Such a file would not authorise the underlying model, and the device sponsor would remain responsible for demonstrating the safety and effectiveness of the finished product.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
For hospitals, this dependency is also a procurement and change-control issue. A prudent institutional approach is to require version identification, update notification, disclosure of known limitations, audit access, incident cooperation, rollback arrangements and continuity planning if an underlying model is changed or withdrawn.
Agentic AI raises the threshold further because it can plan and execute multi-step tasks, use external tools or take actions across a sequence. The FDA notes that some agentic uses may fall outside device oversight, while others may meet the device definition—for example, where an agent controls another medical device.<u>[[1]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-1)</u>
As autonomy rises, human oversight cannot remain symbolic. The responsible professional must have the information, time, authority and expertise to intervene. In my assessment, high-consequence actions should require explicit approval gates, restricted permissions, reliable audit trails and safe interruption.
**Human behaviour is part of the safety evidence**
A technically capable system can still be implemented unsafely if its presentation causes users to over-rely on it.
A 2026 Nature Medicine study examined AI-supported dermatological diagnosis in 623 lay participants and 153 primary-care physicians, with an additional 320 medical students used for comparison. In the experimental tasks, assistance from a fairness-constrained model improved diagnostic performance and reduced skin-tone-related disparities. Large-language-model explanations, however, affected users differently: lay participants were more likely to be helped when the AI was correct but misled when it was wrong, while experienced physicians were more resistant. Presenting AI advice first was also associated with greater deference and possible anchoring.<u>[[6]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-6)</u>
The study does not establish that explainable AI is generally unsafe. It used simplified online tasks, only 12 images per participant, limited clinical context and intentionally sampled cases; the lay and physician experiments also involved different diagnostic tasks and were not directly comparable.<u>[[6]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-6)</u>
Its defensible lesson is narrower and important: the relevant unit of evaluation is often the **human–AI system**, not the model alone. User expertise, interface design, timing and the ability to verify evidence can materially affect safety.
**From regulatory authorisation to accountable healthcare use**
Figure 1. A healthcare-governance pathway for generative-AI-enabled medical devices
[image: Six sequential governance stages: FDA regulatory assessment, institutional due diligence, local human-AI validation, controlled implementation, continuous monitoring, and governance decision.] 
1→2→3→4→5→6
1**FDA regulatory assessment**Market-entry evidence for the defined device and intended use.
2**Institutional due diligence**Purpose, vendor, risk, accountability and contractual controls.
3**Local human–AI validation**Population, language, workflow, subgroup and user performance.
4**Controlled implementation**Training, permissions, oversight, escalation and transparency.
5**Continuous monitoring**Safety, equity, drift, workload, incidents and outcomes.
6**Governance decision**Scale, modify, restrict, pause or suspend.
**Source:** Author’s original synthesis informed by FDA, WHO, NIST and IMDRF materials. This is not an FDA pathway or regulatory requirement.
**What healthcare leaders should do now
The following recommendations are the author’s interpretation. They are not FDA requirements.**
**Define purpose and ownership for AI-enabled medical devices**
Every AI-enabled clinical function should have a named executive sponsor, clinical owner, defined intended use, prohibited uses and a clear escalation route.
**Validate the human–AI system locally**
External evidence should be supplemented by evaluation in the local population, language, pathway and infrastructure. Testing should include subgroup performance, human factors and combined human–AI performance.
**Procure for transparency and controlled change**
Contracts should require version identification, notice of material changes, access to relevant performance information, incident cooperation and the ability to restrict or roll back deployment.
**Monitor outcomes, not only outputs**
An institutional AI register should connect each system to safety, effectiveness, equity, workflow and patient-experience measures, with predefined triggers for review, restriction or suspension.
**Make oversight operational**
Clinicians need enough information, training, time and authority to challenge an output. Patients should understand when AI materially contributes to their care and where professional responsibility remains.
**Apply stricter controls to agentic functions**
Use least-privilege access, human approval for consequential actions, complete logging, tested failure modes and an accessible emergency stop.
These recommendations are consistent with WHO’s emphasis on risk–benefit assessment, transparency, accountability, evaluation and performance monitoring; NIST’s voluntary Govern–Map–Measure–Manage model; and IMDRF principles covering representative data, human–AI team performance, user information and post-deployment monitoring.<u>[[7]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-7)[[8]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-8)[[9]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-9)[[10]](drsameraldiri.com/wp-admin/post.php?post=9473&action=edit#ref-10)</u>
**What remains unresolved**
• Can open-ended performance be tested rigorously without narrowing the evaluation until emergent behaviour is missed?
• How much premarket uncertainty is acceptable before postmarket monitoring becomes a substitute for evidence?
• Who is accountable when harm arises across a foundation-model developer, device sponsor, healthcare organisation and clinician?
• Will voluntary foundation-model disclosure provide enough transparency where commercial incentives favour opacity?
• How should agentic systems be assessed when errors can accumulate across multiple steps before human review?
These questions are not arguments against innovation. They are conditions for making innovation trustworthy.
**Conclusion: authorisation is a threshold, not an operating model**
The FDA paper is exploratory and deliberately non-binding, but it captures a real shift in regulatory thinking. Generative medical AI cannot be evaluated only as a static model producing a single answer. Safety depends on the configured device, the user, the conversation, the workflow, the underlying model, the possibility of change and the consequences of action.
Regulatory authorisation remains essential. It is not sufficient. Healthcare organisations must still decide whether a system is suitable for their population, whether users can challenge it, whether changes remain controlled, whether performance stays equitable, whether incidents are detectable and whether the technology can be restricted when confidence is no longer justified.
> **Market authorisation may open the door. Accountable governance determines how far that door should open—and whether it should remain open.**
**Method and limitations**
This is a focused narrative and policy analysis, not a systematic review or legal opinion. Regulatory claims were checked against primary FDA materials through 20 August 2026. The human-factors evidence is selective, and the organisational recommendations are the author’s synthesis informed by FDA, WHO, NIST and IMDRF materials. Regulatory policy may change after consultation.
**Related analysis by Dr Samer Al-Diri
<u>[Healthcare AI](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Governance](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[:](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Institutional](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Leadership](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[,](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Workforce](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Readiness](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[, and Patient-](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Centred](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Accountability](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[as](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Foundations](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[for](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Sustainable](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)[Healthcare Transformation](drsameraldiri.com/healthcare-ai-governance-institutional-leadership-workforce-readiness-and-patient-centred-accountability-as-foundations-for-sustainable-healthcare-transformation/)**</u>
<u>[ResearchGate](doi.org/10.13140/RG.2.2.13405.99041)[preprint](doi.org/10.13140/RG.2.2.13405.99041)[and DOI](doi.org/10.13140/RG.2.2.13405.99041)[record](doi.org/10.13140/RG.2.2.13405.99041)[for Healthcare AI](doi.org/10.13140/RG.2.2.13405.99041)[Governance](doi.org/10.13140/RG.2.2.13405.99041)</u>
**<u>[Building Resilient Health Systems: A Healthcare Management Framework for](drsameraldiri.com/building-resilient-health-systems-healthcare-management-framework/)[Governance](drsameraldiri.com/building-resilient-health-systems-healthcare-management-framework/)[,](drsameraldiri.com/building-resilient-health-systems-healthcare-management-framework/)[Workforce](drsameraldiri.com/building-resilient-health-systems-healthcare-management-framework/)[](drsameraldiri.com/building-resilient-health-systems-healthcare-management-framework/)[Readiness](drsameraldiri.com/building-resilient-health-systems-healthcare-management-framework/)[and Digital Transformation](drsameraldiri.com/building-resilient-health-systems-healthcare-management-framework/)**</u>
**Selected references**
1. US Food and Drug Administration. <u>[Considerations](fda.gov/media/194242/download)[for the](fda.gov/media/194242/download)[Regulation](fda.gov/media/194242/download)[of](fda.gov/media/194242/download)[Generative](fda.gov/media/194242/download)[](fda.gov/media/194242/download)[AI-Enabled](fda.gov/media/194242/download)[Medical Devices:](fda.gov/media/194242/download)[Discussion](fda.gov/media/194242/download)[](fda.gov/media/194242/download)[Paper](fda.gov/media/194242/download)[and Request for Feedback](fda.gov/media/194242/download)</u>. 18 August 2026.
2. US Food and Drug Administration. <u>[FDA](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[Seeks](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[Public Feedback to](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[Inform](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[Regulatory](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[Approach](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[for](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[Generative](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[AI-Enabled](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)[Medical Devices](fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices)</u>. 18 August 2026.
3. US Food and Drug Administration. <u>[Artificial](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[Intelligence-Enabled](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[Device Software](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[Functions](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[:](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[Lifecycle](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[Management and Marketing](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[Submission](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)[Recommendations](fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing)</u>. Draft guidance. January 2025.
4. US Food and Drug Administration. <u>[Marketing](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[Submission](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[Recommendations](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[for a](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[Predetermined](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[Change](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[Control Plan for](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[Artificial](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[Intelligence-Enabled](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[Device Software](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)[Functions](fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)</u>. Final guidance. August 2025.
5. US Food and Drug Administration. <u>[Request for Public](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[Comment](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[:](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[Measuring](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[and](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[Evaluating](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[Artificial](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[Intelligence-Enabled](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[Medical Device](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[Performance](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)[in the Real-World](fda.gov/medical-devices/digital-health-center-excellence/request-public-comment-measuring-and-evaluating-artificial-intelligence-enabled-medical-device)</u>. 2025.
6. Xu XO, Hu H, Zhang H, et al. <u>[Divergent](doi.org/10.1038/s41591-026-04553-w)[](doi.org/10.1038/s41591-026-04553-w)[impacts](doi.org/10.1038/s41591-026-04553-w)[of](doi.org/10.1038/s41591-026-04553-w)[explainable](doi.org/10.1038/s41591-026-04553-w)[AI for](doi.org/10.1038/s41591-026-04553-w)[dermatological](doi.org/10.1038/s41591-026-04553-w)[](doi.org/10.1038/s41591-026-04553-w)[diagnosis](doi.org/10.1038/s41591-026-04553-w)[on](doi.org/10.1038/s41591-026-04553-w)[clinicians](doi.org/10.1038/s41591-026-04553-w)[versus](doi.org/10.1038/s41591-026-04553-w)[lay](doi.org/10.1038/s41591-026-04553-w)[](doi.org/10.1038/s41591-026-04553-w)[people](doi.org/10.1038/s41591-026-04553-w)</u>. Nature Medicine. 2026;32:3000–3009. doi:10.1038/s41591-026-04553-w.
7. World Health Organization. <u>[Ethics](who.int/publications/i/item/9789240084759)[and](who.int/publications/i/item/9789240084759)[Governance](who.int/publications/i/item/9789240084759)[of](who.int/publications/i/item/9789240084759)[Artificial](who.int/publications/i/item/9789240084759)[](who.int/publications/i/item/9789240084759)[Intelligence](who.int/publications/i/item/9789240084759)[for Health:](who.int/publications/i/item/9789240084759)[Guidance](who.int/publications/i/item/9789240084759)[on](who.int/publications/i/item/9789240084759)[Large](who.int/publications/i/item/9789240084759)[](who.int/publications/i/item/9789240084759)[Multi-Modal](who.int/publications/i/item/9789240084759)[](who.int/publications/i/item/9789240084759)[Models](who.int/publications/i/item/9789240084759)</u>. Geneva: WHO; 25 March 2025.
8. World Health Organization. <u>[Regulatory](who.int/publications/i/item/9789240078871)[](who.int/publications/i/item/9789240078871)[Considerations](who.int/publications/i/item/9789240078871)[on](who.int/publications/i/item/9789240078871)[Artificial](who.int/publications/i/item/9789240078871)[](who.int/publications/i/item/9789240078871)[Intelligence](who.int/publications/i/item/9789240078871)[for Health](who.int/publications/i/item/9789240078871)</u>. Geneva: WHO; 19 October 2023.
9. Autio C, Schwartz R, Dunietz J, et al. <u>[Artificial](doi.org/10.6028/NIST.AI.600-1)[](doi.org/10.6028/NIST.AI.600-1)[Intelligence](doi.org/10.6028/NIST.AI.600-1)[](doi.org/10.6028/NIST.AI.600-1)[Risk](doi.org/10.6028/NIST.AI.600-1)[Management Framework:](doi.org/10.6028/NIST.AI.600-1)[Generative](doi.org/10.6028/NIST.AI.600-1)[](doi.org/10.6028/NIST.AI.600-1)[Artificial](doi.org/10.6028/NIST.AI.600-1)[](doi.org/10.6028/NIST.AI.600-1)[Intelligence](doi.org/10.6028/NIST.AI.600-1)[Profile](doi.org/10.6028/NIST.AI.600-1)</u>. NIST AI 600-1. National Institute of Standards and Technology; 26 July 2024.
10. International Medical Device Regulators Forum. <u>[Good Machine Learning](imdrf.org/documents/good-machine-learning-practice-medical-device-development-guiding-principles)[Practice](imdrf.org/documents/good-machine-learning-practice-medical-device-development-guiding-principles)[for Medical Device Development:](imdrf.org/documents/good-machine-learning-practice-medical-device-development-guiding-principles)[Guiding](imdrf.org/documents/good-machine-learning-practice-medical-device-development-guiding-principles)[](imdrf.org/documents/good-machine-learning-practice-medical-device-development-guiding-principles)[Principles](imdrf.org/documents/good-machine-learning-practice-medical-device-development-guiding-principles)</u>. IMDRF/AIML WG/N88 FINAL:2025. 29 January 2025.
**About the author**
Dr. Samer Al-Diri, MD, MSc, MPH, is a UK-trained ophthalmologist, public-health professional and strategic healthcare consultant. His work spans healthcare management, health-system transformation, clinical governance, prevention, responsible digital health and eye health.
**<u>[Official](drsameraldiri.com/)[](drsameraldiri.com/)[Website](drsameraldiri.com/)**</u> | **<u>[ORCID](orcid.org/0009-0004-1908-0714)**</u> | **<u>[ResearchGate](researchgate.net/profile/Samer-Al-Diri)**</u> | **<u>[LinkedIn](linkedin.com/in/dr-samer-al-diri/)**</u>
ResearchGate DOI: <u>[10.13140/RG.2.2.31927.07845](doi.org/10.13140/RG.2.2.31927.07845)</u>
**Open-access licence:** © 2026 Dr Samer Al-Diri. This article is licensed under <u>[CC BY-NC-ND 4.0](creativecommons.org/licenses/by-nc-nd/4.0/)</u>.
Professional and educational analysis only. This article does not constitute legal, regulatory or clinical advice.
