Generative AI Medical Devices: FDA’s 2026 Discussion Paper

FDA has no guidance written for generative AI (GenAI) medical devices yet, so a GenAI device follows the same pathways as any other device. On August 18, 2026, FDA published a discussion paper that proposes a two-axis risk framework and a "competency-based" review, with feedback due by October 19, 2026 [1, 2].

Key takeaways

• The August 18, 2026 paper is not guidance. FDA says it does not propose or set policy, and it asks 26 questions under docket FDA-2026-N-7874 [1, 2, 8].
• FDA suggests a possible two-axis way to think about risk: how independently the device acts, and how serious the harm is if a user relies on a wrong output [2].
• The proposed benchmarking covers 10 elements in 4 groups: safety, clinical proficiency, generalizability and agentic capabilities [2].
• FDA's AI-enabled device list holds 1,614 devices with decisions through June 29, 2026, and it does not yet flag which ones use large language models [4].
• Today, a GenAI device still goes through the normal 510(k), De Novo or PMA pathway, and FDA's final PCCP guidance already covers planned AI changes [5].

What is a generative AI-enabled medical device?

A GenAI-enabled device is a product with at least one device software function that uses generative AI [2]. FDA defines generative AI as AI models that create new synthetic content, such as text, images, audio or video, based on patterns in their input data [2].

The paper also defines related terms. A foundation model is a large model trained on broad data that can be adapted to many tasks. A large language model (LLM) is a type of AI model trained on large text datasets. An agentic AI system plans and carries out multi-step tasks, uses tools or takes actions on its own [2].

FDA is clear on one point: it "does not regulate GenAI as such." It regulates products that meet the device definition in section 201(h) of the FD&C Act [2]. Many software functions are not devices, or fall under FDA enforcement discretion, as the Device Software Functions and Mobile Medical Applications guidance explains [2, 10]. A product can mix device functions and non-device functions, and FDA reviews the device functions [2, 9].

Has FDA issued rules for generative AI devices?

No final or draft guidance targets GenAI devices today. The August 18, 2026 paper is for discussion only, and FDA states it does not communicate CDRH's expected evidence for future submissions [1, 2].

The paper builds on earlier work. In November 2024, FDA's Digital Health Advisory Committee met to discuss total product life cycle questions for GenAI devices [2, 3]. FDA's January 7, 2025 draft guidance on AI-enabled device software functions still lists as a draft in FDA's guidance database [7].

More is planned. On October 1, 2026, CDRH published its fiscal year 2027 guidance list [12]. Its "A-list" of priority documents includes a new draft guidance, "Generative Artificial Intelligence-Enabled Conversational Devices for Mental Disorders: Evidentiary Recommendations" [12]. The A-list also plans a final version of the AI-enabled device software functions guidance [12]. A broader guidance, "Risk Assessment and Evidentiary Considerations for Generative Artificial Intelligence-Enabled Devices", sits on the "Under Construction" list, which CDRH develops as resources permit [12]. The list states intent, not a publication date.

We checked FDA's own AI-enabled device list to see what is already on the market. The list has 1,614 entries, and 1,553 of them (96.2%) are 510(k) clearances [4]. Its six columns show the decision date, submission number, device, company, review panel and product code, so none of them says whether a device uses an LLM [4].

How we counted: We downloaded the CSV version of FDA's AI-enabled device list on October 26, 2026 (page content current as of September 4, 2026). We counted rows by submission number prefix: K for 510(k), DEN for De Novo and P for PMA [4].

These FDA documents shape AI device reviews today:

Document	Status	What it covers
Discussion paper on GenAI-enabled devices	Discussion paper, August 18, 2026	Risk framework, premarket evidence, postmarket monitoring
AI-enabled device software functions	Draft guidance, January 7, 2025	Lifecycle and marketing submission content for AI devices
PCCP for AI-enabled device software functions	Final guidance, August 2025 version	How to plan and pre-authorize future AI changes
Clinical Decision Support Software [6]	Final guidance, January 2026	Which decision support software is not a device
Multiple Function Device Products	Final guidance, July 29, 2020	How FDA reviews products with device and non-device functions

Guidance documents are not binding. They describe FDA's current thinking, and a discussion paper carries even less weight than a draft guidance.

How would FDA judge the risk of a GenAI device?

FDA suggests a two-axis framework as a possible way to think about risk [2]. One axis is device activity: how far the function directs or takes action on its own. The other is consequence: how serious the harm is if someone relies on a wrong output [2].

Function type in the paper	FDA's example	How FDA is thinking about risk
Non-directive information	A risk score for a future cardiovascular event	Lower risk than action-directing output
Action-directing information	A patient-facing suggestion on whether to seek emergency care	Higher risk, adjusted by how serious the harm could be
Action-taking function	Assigning a diagnosis, prescribing, starting an order set	Typically higher risk than informational functions
Measurement and signal processing	Outputs the user cannot check alone	May be higher risk even though non-directive

Three details matter for product teams. First, a "talk to your doctor" line may not make an action-directing output less directive [2]. Second, FDA asks whether patient-facing outputs should count as higher risk, because patients may not spot a wrong answer, while also noting the benefits of patient access to information [2]. Third, for chat products, FDA is considering risk across a whole conversation, because a session can drift from general information to direct advice [2].

FDA is also considering whether to count both kinds of escalation error. Failing to send a patient to emergency care is a harm, but so is sending too many patients there without need [2].

What evidence might FDA expect before a GenAI device reaches the market?

FDA is considering a competency-based review with two parts: device benchmarking and clinical confirmation [2]. Under this approach, the final device as deployed would be evaluated, not the foundation model on its own [2].

FDA compares this to how doctors are trained and licensed. Doctors are not tested on every possible case. They pass structured exams, then work under supervision with growing independence [2].

Benchmarking is high-volume, non-clinical testing. The paper lists 10 possible elements, chosen by intended use and risk [2]:

Group	Elements FDA lists
Safety	S.1 Safety-critical recognition and escalation; S.2 Scope maintenance and boundary adherence; S.3 Calibration, uncertainty and clinical deferral
Clinical proficiency	E.1 Clinical knowledge and task fidelity; E.2 Information gathering and clinical analysis; E.3 Quantitative and measurement analysis; E.4 Communication quality and user comprehension
Generalizability	R.1 Robustness, reliability and reproducibility; R.2 Subgroup performance
Agentic AI	A.1 Extra competencies for agentic devices

Sponsors would prespecify test methods, scoring rubrics and acceptance criteria before testing [2]. Expert adjudicators would be independent of the sponsor and of any third-party model developer. That still applies when the adjudicator is itself an LLM [2].

Clinical confirmation checks performance in real or clinically representative use. FDA says it "might not require a prospective clinical study in every case" [2]. The paper lists five possible approaches, in approximate order of increasing rigor and patient exposure:

1. Retrospective evaluation on real patient inputs
2. Shadow deployment, where outputs are recorded but not shown
3. Standardized patient interactions with trained actors
4. Clinician adjudication of real cases
5. A prospective clinical study, sometimes a randomized trial

For open-ended outputs, FDA is considering comparing the device with a panel of qualified clinicians, or with a median clinician in practice [2]. FDA also asks whether independent third parties could run parts of the testing, and names its ASCA and MDDT programs as possible models [2].

How would FDA handle changes after clearance?

FDA is considering whether to accept more uncertainty before market in exchange for stronger monitoring after market [2]. It gives three example methods: periodic re-benchmarking, sampled review by independent clinicians, and monitoring for performance drift [2].

Changes are a central problem. A GenAI device can change through a sponsor update, through gradual model evolution, or when a third-party foundation model is updated [2]. FDA notes that a PCCP "could be one mechanism" for some changes, and that the premarket benchmark could be the baseline for re-testing after each change [2, 5].

The paper also floats voluntary Foundation Model Device Master Files (MAFs), built on FDA's existing Device Master File program. Model developers could share model cards and other details with FDA in confidence, and device sponsors could reference them [2]. FDA states that such a file would not authorize the model for any device use [2].

What should a GenAI device team do now?

These steps are our suggested practice, built around the paper's structure. They are not FDA requirements, because the paper's ideas are still proposals:

1. Write a narrow intended use. List every function and mark which are device functions. FDA's multiple function policy explains how this split works [9].
2. Place each function on both risk axes. Note whether outputs are non-directive, action-directing or action-taking, and who the user is [2].
3. Design a benchmark set. Cover the safety, proficiency and subgroup elements in the paper that fit your use, and fix rubrics and pass criteria before you test.
4. Plan clinical confirmation. Pick a method whose rigor matches your device's risk, as the paper suggests, and justify it [2].
5. Control the foundation model. Pin model versions, ask the model vendor for update notices, and re-run your benchmark after any change. The paper raises this as an open question [2].
6. Consider a PCCP for changes you can describe in advance [5].
7. Ask FDA early. A Pre-Submission (Q-Sub) lets you test your plan with FDA before you spend on studies [11].

If your device also fits FDA's existing AI patterns, read our analysis of 1,614 FDA-authorized AI devices and our PCCP implementation guide. For decision support software, check the 2026 CDS exemption rules first.

Choosing a pathway for a GenAI device is a decision with real risk, because no predicate or guidance may fit. Complizen's free gap assessment reviews your intended use and evidence plan with regulatory experts who have direct FDA submission experience.

What should manufacturers outside the US know?

The same US rules apply to every manufacturer, wherever the device is built. A marketing authorization from another country does not replace FDA review.

The paper asks how data collected outside the United States could support clinical confirmation [2]. That is a question, not a promise, so plan to show that your data fits US patients and US clinical practice.

If your team builds on a foundation model from another company, note that FDA's paper says changes made by the model developer can affect your device [2]. Your contracts with the model vendor become part of your regulatory plan. For the full AI device picture, see our SaMD and AI learning path.

Frequently asked questions

Has FDA cleared any generative AI medical devices?

FDA's public list cannot answer this yet. The AI-enabled device list holds 1,614 devices with decisions through June 29, 2026. FDA says it will explore ways to tag devices that use LLMs or other foundation models. Until then, check each device's decision summary for a description of its AI.

Is FDA's generative AI discussion paper a guidance document?

No. FDA states the paper is for discussion only and is not draft or final guidance. It does not set policy or tell sponsors what evidence to submit. It asks 26 questions to collect feedback. CDRH's FY2027 list plans a draft guidance on GenAI conversational devices for mental disorders, but no GenAI guidance is published yet.

When did the comment period for the GenAI paper close?

FDA asked for feedback under docket FDA-2026-N-7874 on Regulations.gov by October 19, 2026. FDA published the paper on August 18, 2026. You can still read submitted comments on the docket. FDA has not said when, or whether, it will publish draft guidance based on this feedback.

What is the two-axis risk framework?

It is a way to think about the risk of a GenAI function. One axis is how independently the device acts, from giving information to taking action. The other is how serious the harm could be if a user relies on a wrong output. FDA calls it a possible heuristic, not a final rule.

Does a "talk to your doctor" disclaimer lower the risk of an AI chatbot?

Not necessarily. FDA is considering whether a patient-facing function stays action-directing even with a "talk to your doctor" or "I am not a medical professional" line. The content and context of the output may matter more than its wording. This is a proposal, not current policy.

What does competency-based evaluation mean?

It is FDA's proposed way to test GenAI devices, inspired by how doctors are trained and licensed. It has two parts: non-clinical benchmarking across many inputs, and clinical confirmation in real or realistic use. The final deployed device would be evaluated, not the foundation model alone. The approach is still only a proposal.

Will a GenAI device always need a clinical trial?

Not always, according to the paper. FDA says clinical confirmation might not need a prospective study in every case. Options range from retrospective testing on real patient inputs to shadow deployment, standardized patients, clinician review and, depending on the device, a prospective or randomized study. Rigor would match risk and intended use.

What is a Foundation Model Device Master File?

It is a voluntary idea in the paper. Foundation model developers could share confidential details, such as model cards, training data sources and known failure modes, with FDA. Device sponsors could then reference the file in their submissions. FDA says such a file would not authorize the model for any device use.

Can a PCCP cover updates to the underlying foundation model?

Possibly, for changes you can describe in advance. FDA's final PCCP guidance lets sponsors pre-authorize planned AI changes with a modification protocol. The GenAI paper says a PCCP could be one tool, but asks how it would work when future changes cannot be fully specified. Discuss this with FDA in a Pre-Submission.

Are AI scribes and documentation tools regulated as devices?

They may not be, but it depends on the function. FDA's paper says agentic tools for care coordination, clinical documentation and patient outreach may not be functions FDA focuses on as devices. A tool becomes a device function when it meets the device definition, for example when its actions control another medical device.

Does FDA regulate the large language model itself?

No. FDA says it does not regulate GenAI as such. It regulates medical devices, including devices that use GenAI. FDA would evaluate the final user-facing device for its intended use. The model developer could share information through a voluntary master file, but the device sponsor stays responsible.

What should I do now if I am building a GenAI device?

Define a narrow intended use, place each function on FDA's two risk axes, and design a benchmark set with fixed pass criteria. Plan how you will confirm performance in real use and control foundation model updates. Then request a Pre-Submission meeting to test your plan with FDA before you start costly studies.

References

1. FDA press release, FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices, August 18, 2026. https://www.fda.gov/news-events/press-announcements/fda-seeks-public-feedback-inform-regulatory-approach-generative-ai-enabled-medical-devices
2. FDA CDRH, Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback, August 18, 2026. https://www.fda.gov/medical-devices/digital-health-center-excellence/considerations-regulation-generative-ai-enabled-medical-devices-discussion-paper-and-request
3. FDA, Executive Summary for the Digital Health Advisory Committee Meeting: Total Product Lifecycle Considerations for Generative AI-Enabled Devices, November 2024. https://www.fda.gov/media/182871/download
4. FDA, Artificial Intelligence-Enabled Medical Devices list, content current as of September 4, 2026. https://www.fda.gov/medical-devices/artificial-intelligence-enabled-medical-devices/list-artificial-intelligence-enabled-medical-devices
5. FDA, Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions, final guidance (guidance page dated August 18, 2025). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence
6. FDA, Clinical Decision Support Software, final guidance, January 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software
7. FDA, Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations, draft guidance, January 7, 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing
8. Regulations.gov, docket FDA-2026-N-7874, accessed October 26, 2026. https://www.regulations.gov/docket/FDA-2026-N-7874
9. FDA, Multiple Function Device Products: Policy and Considerations, final guidance, July 29, 2020. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/multiple-function-device-products-policy-and-considerations
10. FDA, Policy for Device Software Functions and Mobile Medical Applications, final guidance, September 28, 2022. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/policy-device-software-functions-and-mobile-medical-applications
11. FDA, Requests for Feedback and Meetings for Medical Device Submissions: The Q-Submission Program, final guidance, May 29, 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/requests-feedback-and-meetings-medical-device-submissions-q-submission-program
12. FDA, CDRH Proposed Guidances for Fiscal Year 2027 (FY 2027), October 1, 2026. https://www.fda.gov/medical-devices/guidance-documents-medical-devices-and-radiation-emitting-products/cdrh-proposed-guidances-fiscal-year-2027-fy-2027