Public comment periodCloses Oct 19, 2026

7days20hours59minutes26seconds

Research by

The FDA’s 26 GenAI questions.What the public is saying.

Public feedback on the regulation of generative AI-enabled medical devices.

Read the FDA paper ↗Public docket on regulations.gov ↗

Explore responses ↓The 26 questions →

95 filings screened · Latest review Sep 16, 2026 Comments close Oct 19, 2026

FDA paper ↗Public docket ↗

FDA GenAI public feedbackAll audiences

123 responses so far to the FDA’s GenAI questions.

Over the first 55 days, industry contributed 75 of 123 submissions (61%), more than all other groups combined. Academia / other contributed the fewest, with 10. There are 7 days remaining before comments close on October 19.

View all submissions ↗

One dot per submission. Hover or tap for details.

Source: FDA public docket · FDA-2026-N-7874 ↗

Filings through Oct 11, 2026

Research by

recovry.ai/fda-ai-feedback-tracker

FDA GenAI public feedbackAll audiences

Views on the FDA’s approach to GenAI.

IndustryCliniciansPublic / patientsAcademia / other

One dot per filing. Select a column to read what they said.

Source: FDA public docket · FDA-2026-N-7874 ↗

Filings through Oct 11, 2026

Research by

recovry.ai/fda-ai-feedback-tracker

FDA GenAI public feedbackAll audiences

The FDA GenAI questions getting the most attention.

Of the FDA’s 26 questions, these three appear most often across 123 public submissions so far. Counts show references, not agreement; each card also highlights leading recommendations or positions in the coded filings.

FDA Q19 · Postmarket performance evaluation

How should an AI device be monitored after launch, and what sets the cadence?

55 of 123 filings name it

Most mentioned · 3 of 4 · 21 coded

Repeat performance testing on a schedule14

Reassess after changes or safety signals12

Have clinicians review samples of outputs9

A filing can propose more than one.

FDA Q1 · The two-axis risk framework

Does a two-axis framework, AI device activity and the consequence of relying on an incorrect output, capture the dimensions of risk?

45 of 123 filings name it

Positions · 21 coded

Keep it, but add or change elements20

The framework is insufficient1

Keep the framework as proposed0

FDA Q24 · Third-party foundation model changes

When the foundation model’s developer changes the model, how does the device maker detect it and respond, so safety and effectiveness are not compromised?

44 of 123 filings name it

Most mentioned · 3 of 4 · 15 coded

Identify and control the model version in use13

Detect supplier updates or unexpected behavior changes12

Retest changed models or provide rollback12

A filing can propose more than one.

IndustryCliniciansPublic / patientsAcademia / other

Explore all 26 questions ↓

Source: FDA public docket · FDA-2026-N-7874 ↗ · Preliminary analysis

Filings through Oct 11, 2026 · Responses: Sep 13 snapshot · Citing a question does not imply agreement.

Research by

recovry.ai/fda-ai-feedback-tracker

Themes across the public feedbackAll audiences

Top themes shaping the FDA GenAI discussion.

A synthesis of the top emerging themes across public submissions. These themes connect several FDA questions and will evolve as new submissions are reviewed. Counts show submissions raising each theme, not agreement.

IndustryCliniciansPublic / patientsAcademia / other

Source: FDA public docket · FDA-2026-N-7874 ↗ · Preliminary analysis

Filings through Oct 11, 2026 · Filings may appear in several themes.

Research by

recovry.ai/fda-ai-feedback-tracker

FDA GenAI public feedbackAll audiences

Public feedback on all 26 FDA GenAI questions.

Explore each question and the submissions that discuss it.

IndustryCliniciansPublic / patientsAcademia / other

Assessing clinical riskQ1–Q6

FDA Q1 · The two-axis risk framework

Does a two-axis framework, AI device activity and the consequence of relying on an incorrect output, capture the dimensions of risk?

45 of 123 filings name it

FDA Q2 · The spectrum of device activity

How should the continuum from non-directive to action-directing outputs, and the risk that changes along it, be accounted for?

32 of 123 filings name it

FDA Q3 · When an output becomes directive

When clinical information goes straight to the patient, does the risk change, and what safeguards help without underestimating patients?

31 of 123 filings name it

FDA Q4 · Generalist and specialist users

Should it matter whether the clinician using the AI is a generalist or a specialist?

24 of 123 filings name it

FDA Q5 · Multi-turn conversations that migrate

How is risk assessed when a conversation starts with non-directive information and drifts into action-directing?

31 of 123 filings name it

FDA Q6 · Care escalation functions

How should under-escalation be weighed against over-escalation?

35 of 123 filings name it

Demonstrating clinical competenceQ7–Q17

FDA Q7 · The competency-based approach

Is the two-step approach, benchmark the AI, then confirm it in clinical use, the right way to evaluate these devices?

33 of 123 filings name it

FDA Q8 · Mapping the risk grid to evidence

Should an AI device’s position on the risk map help decide how much evidence it must bring before market?

21 of 123 filings name it

FDA Q9 · The benchmarking structure

Do the ten benchmark competencies, from clinical knowledge to generalizability, add up to enough evidence of safety and effectiveness?

40 of 123 filings name it

FDA Q10 · Benchmark contamination and saturation

How can a benchmark score be shown to predict real-world behavior?

36 of 123 filings name it

FDA Q11 · Clinical confirmation without a prospective trial

When can a device be confirmed without a prospective clinical study, and what earns that lighter path?

31 of 123 filings name it

FDA Q12 · Statistically meaningful performance

How do you get statistically meaningful performance numbers when synthetic inputs are mixed with real ones?

28 of 123 filings name it

FDA Q13 · Synthetic data

Where is synthetic data good enough, and where is it not?

28 of 123 filings name it

FDA Q14 · Comparators and acceptance criteria

For open-ended AI outputs, who is the performance comparator: a clinician panel, generalists, specialists, or the human-AI team?

32 of 123 filings name it

FDA Q15 · Performance against usual care

Could the AI be measured against what would have happened without it: unaided judgment, a delayed specialist, or no intervention?

29 of 123 filings name it

FDA Q16 · Independent third parties

What role should independent third parties play?

35 of 123 filings name it

FDA Q17 · Devices with many functions

Does the approach still work for devices built on other model architectures, such as multimodal vision-language models and world models?

19 of 123 filings name it

Monitoring after launchQ18–Q24

FDA Q18 · Trading premarket certainty for postmarket monitoring

Can greater premarket uncertainty about a GenAI device’s benefit-risk profile be accepted through greater reliance on postmarket monitoring?

41 of 123 filings name it

FDA Q19 · Postmarket performance evaluation

How should an AI device be monitored after launch, and what sets the cadence?

55 of 123 filings name it

FDA Q20 · Machine-based supervisory agents

Could AI supervisory agents help carry out postmarket monitoring?

42 of 123 filings name it

FDA Q21 · Clinicians, institutions and societies

What roles should clinicians and institutions play in monitoring, without diluting manufacturer accountability?

38 of 123 filings name it

FDA Q22 · Re-benchmarking after a modification

With the premarket competency assessment as the baseline, which post-deployment changes need re-evaluation, and how much?

40 of 123 filings name it

FDA Q23 · PCCPs for GenAI devices

How can a change-control plan cover changes that cannot be fully specified in advance?

30 of 123 filings name it

FDA Q24 · Third-party foundation model changes

When the foundation model’s developer changes the model, how does the device maker detect it and respond, so safety and effectiveness are not compromised?

44 of 123 filings name it

Source: FDA public docket · FDA-2026-N-7874 ↗ · Preliminary analysis

Filings through Oct 11, 2026 · Responses: Sep 13 snapshot · Citing a question does not imply agreement.

Research by

recovry.ai/fda-ai-feedback-tracker

Who’s saying what?

All audiences shown for comparison.

Provisional mapping: industry includes startups, large companies and consultants; practitioners includes clinicians and health systems. Academia is shown separately. Filings are not a representative population survey.

The conditions matter.

How much autonomy is accepted?

Highest accepted autonomy, compared across two consequence levels.

Autonomy →

Two debates worth watching.

Two themes drawn from the FDA’s questions.

FDA GenAI public feedbackAll audiences

The public submissions behind the FDA GenAI discussion.

Browse the filings, contributors and original source documents.

Source: FDA public docket · FDA-2026-N-7874 ↗

Filings through Oct 11, 2026

Research by

recovry.ai/fda-ai-feedback-tracker

SearchCoverage

Explore the 26 FDA questions

Counts show coverage among the selected audience, not agreement.

The FDA’s 26 GenAI questions.What the public is saying.

123 responses so far to the FDA’s GenAI questions.

Views on the FDA’s approach to GenAI.

The FDA GenAI questions getting the most attention.

How should an AI device be monitored after launch, and what sets the cadence?

Does a two-axis framework, AI device activity and the consequence of relying on an incorrect output, capture the dimensions of risk?

When the foundation model’s developer changes the model, how does the device maker detect it and respond, so safety and effectiveness are not compromised?

Top themes shaping the FDA GenAI discussion.

How should a device be watched after launch, and can another AI do the watching?

How many dimensions does it take to evaluate the risk of an AI?

Is a benchmark score sufficient to predict how a device behaves with real patients?

Public feedback on all 26 FDA GenAI questions.

Does a two-axis framework, AI device activity and the consequence of relying on an incorrect output, capture the dimensions of risk?

How should the continuum from non-directive to action-directing outputs, and the risk that changes along it, be accounted for?

When clinical information goes straight to the patient, does the risk change, and what safeguards help without underestimating patients?

Should it matter whether the clinician using the AI is a generalist or a specialist?

How is risk assessed when a conversation starts with non-directive information and drifts into action-directing?

How should under-escalation be weighed against over-escalation?

Is the two-step approach, benchmark the AI, then confirm it in clinical use, the right way to evaluate these devices?

Should an AI device’s position on the risk map help decide how much evidence it must bring before market?

Do the ten benchmark competencies, from clinical knowledge to generalizability, add up to enough evidence of safety and effectiveness?

How can a benchmark score be shown to predict real-world behavior?

When can a device be confirmed without a prospective clinical study, and what earns that lighter path?

How do you get statistically meaningful performance numbers when synthetic inputs are mixed with real ones?

Where is synthetic data good enough, and where is it not?

For open-ended AI outputs, who is the performance comparator: a clinician panel, generalists, specialists, or the human-AI team?

Could the AI be measured against what would have happened without it: unaided judgment, a delayed specialist, or no intervention?

What role should independent third parties play?

Does the approach still work for devices built on other model architectures, such as multimodal vision-language models and world models?

Can greater premarket uncertainty about a GenAI device’s benefit-risk profile be accepted through greater reliance on postmarket monitoring?

How should an AI device be monitored after launch, and what sets the cadence?

Could AI supervisory agents help carry out postmarket monitoring?

What roles should clinicians and institutions play in monitoring, without diluting manufacturer accountability?

With the premarket competency assessment as the baseline, which post-deployment changes need re-evaluation, and how much?

How can a change-control plan cover changes that cannot be fully specified in advance?

When the foundation model’s developer changes the model, how does the device maker detect it and respond, so safety and effectiveness are not compromised?

Would voluntary Foundation Model Master Files be practical, and useful in premarket review?

What extra oversight does an AI that plans and acts in multiple steps need?

Who’s saying what?

The conditions matter.

How much autonomy is accepted?

Two debates worth watching.

The public submissions behind the FDA GenAI discussion.