Filers agreed with the FDA’s core proposals and asked for changes. More than half asked for more oversight than the FDA proposed.
This research analyzes every public comment on the FDA’s discussion paper on generative AI-enabled medical devices: 530 comments from 509 filers. AI models read and sorted every comment. A reviewer at RecovryAI checked the results, and the published figures and quotes, against the filings. RecovryAI, a company developing FDA-regulated AI for post-operative recovery, prepared this research to bring greater understanding, visibility and discussion to the regulation of AI in health care, and also filed a comment on this docket.
OverviewKey takeaway 1 of 10
All 509 filers, and comment counts on 11 earlier FDA consultations on AI
What’s behind this530 commentsGenerative AI medical devices, 2026
239AI in early-phase clinical trials, 2026
155AI/ML device software framework, 2019
124AI in drug regulatory decisions, 2025
509 filers sent 530 comments, about 1.7 million words, in 63 days. The next largest, a 2026 request on AI in clinical trials, drew 239 comments.
* Largest by comments posted to a single docket on regulations.gov, checked November 8, 2026. The 530 leaves out 10 exact repeats.
What’s behind thisComments posted on the FDA’s earlier consultations on AI, against the filers on this one:
Each filer counts once, however many comments they sent. Together they run to 1.7 million words.
OverviewKey takeaway 2 of 10
All 509 filers by kind, and a name search of every filer and submitter
What’s behind this509 filers
Filed
Did not file
AI and data
Medical devices
Hospitals and telehealth
Health insurers
Clinicians were 24% of filers and researchers 8%. Patients and patient groups were 2%, 9 filers.
Of all 509 filers; groups from RecovryAI’s coding of each filer. Other: 64 filers who did not say who they were, plus policy groups, standards bodies and members of the public.
What’s behind this9% of all 509 filers named OpenAI or one of its AI products (48 filers), more than named any other AI company. Of eight large US health insurers, only Kaiser Permanente filed.
No filing found in a search of every filer and submitter name, including brands and subsidiaries. Named: a comment mentions the company or one of its AI products.
What’s behind thisThe 509 filers by kind, each filer counted once:
The 230 industry filers by kind:
Who filed was checked for 34 names: 26 well-known companies and health systems, plus Meta, xAI, five more large US health insurers and their trade group AHIP, each with its brands and subsidiaries (Optum for UnitedHealth Group, Aetna for CVS Health, Livongo for Teladoc, MiniMed for Medtronic). The “filed” logos are a hand-picked subset of filers. The tally below is for the 26 names.
Filers, of 509, who named each AI model maker or its products: OpenAI 48, Google 20, Anthropic 14, Meta 8, Amazon 7, Microsoft 7. No other maker of AI models was named by more than 5 filers.
The 48 filers who named OpenAI or ChatGPT, by group:
Insurers: no filer’s name matches UnitedHealth Group, Elevance Health, CVS Health (Aetna), Cigna, Humana, Centene, a Blue Cross or Blue Shield plan, or their trade group AHIP. Kaiser Permanente is the only health plan among the filers.
“In our experience, many patients already use generalist LLM platforms (Claude, ChatGPT, Gemini) to analyze their radiology reports.”Raidium, a radiology AI startup, filing 0431
OverviewKey takeaway 3 of 10
Each filer’s overall position on the FDA’s proposals, by kind of filer
What’s behind this32% asked for small changes or none, and 14% asked for less.
* Of the 479 filers whose filing took an overall position on the FDA’s proposals; 30 took none. Positions are RecovryAI’s coding.
What’s behind thisTrade groups are a small base, 13 filers. Large companies split, 40% for less and 42% for small changes or none. Startups asked for more over less, 41% to 17%.
Out of the filers of each kind whose filing took an overall position: 114 clinicians, 13 trade groups, 40 large companies, 115 startups. Kinds from RecovryAI’s coding.
What’s behind thisWhat more and less oversight mean
A filing that asked for some of each is placed by its main ask, the one it leads with or spends most space on; if none leads, it counts as small changes.
How positions were coded
An AI model read each filing in full against this written rule and picked the sentence that most clearly shows its position. A filer takes its filing’s position; a filer whose filings disagree counts as mixed and sits only in the total. The first 111 filings were read twice, then checked by a second model, which agreed on 83 (75%); its disagreements were settled by a further reading. One more filing was read by two readers who agreed. The other 418 were each read by two Claude models, which differed on 106.
Kinds of filer not drawn on the second screen, per filer: Patient groups (8): 12% less oversight, 62% more; Other filers (92): 7% less oversight, 58% more; Other industry (8): 25% less oversight, 50% more.
Filer groups in this note follow the report’s rule: hospitals, health systems and their associations count as Clinicians, and companies that build or sell products count as Industry.
Deep divesKey takeaway 4 of 10
Questions 2 and 3
What’s behind this4A disclaimer doesn’t make AI advice safer
My home blood pressure reading is 182 over 110. Is that bad?
AI assistant
That reading is very high. Sit quietly and check it again in five minutes, and call your doctor’s office today if it stays this high.
This is not medical advice. Consult your doctor.
Call your doctorCall 911
That is of a small base, the 75 filers who took a side. One said it does, and two said it partly does.
* Of the 75 filers who took a side on whether a disclaimer lowers the risk of advice that tells a patient what to do. Another 124 who answered Question 2, on how an answer’s wording changes its risk, didn’t take it up.
What’s behind thisIt was the safeguard named most, by 37% of clinicians and 29% of industry filers.
Of the 202 filers who answered Question 3, on safeguards for AI that gives clinical information to patients; counted only where an answer says so.
What’s behind thisWhat the FDA is considering, in its words
“CDRH is also considering whether a patient-facing informational function may not become any less directive because it includes a “talk to your doctor” or an “I am not a medical professional” statement in addition to the “action-directing” information.”
Of 198 filers who answered Question 2, 75 took a side on whether a disclaimer lowers the risk:
Industry and clinicians agreed: 40 of 41 industry filers and all 17 clinicians who took a side said no.
Safeguards proposed instead, among the 202 filers who answered Question 3 (each counted only where an answer says so):
“A disclaimer or a generic instruction to consult a specialist is not an adequate safeguard when the output is specific, personalized, and likely to influence action.”Manuj Agarwal, MD, a clinician, filing 0056
“The FDA should not treat disclaimers as sufficient mitigation when assessing whether a response is action-directing …”NSF AI Research Institute on Interaction for AI Assistants, a research institute, filing 0370
Deep divesKey takeaway 5 of 10
The seven core proposals: Questions 1, 7, 9, 11, 16, 20 and 25
What’s behind this1Filers generally agreed with the FDA’s core proposals, but asked for changes
The 767 answers that gave a yes or no
875 more answers gave advice without a yes or no.
The FDA’s seven core proposals
Of the 301 filers who gave a yes or no: 87% yes, with changes · 7% as written · 6% no or split*
They backed each of the seven proposals they answered but wanted at least one changed. 7% said yes as written, and 6% said no or split.
* Of the 301 filers who gave a yes or no to at least one of the seven; a split is yes for some devices and no for others. Stamps: share of that question’s yes-or-no answers, one per filer.
What’s behind thisThe seven core proposals, in the FDA’s words
Each proposal: yes, with changes · yes, as written · no · split
The 875 answers that gave advice without a yes or no are counted apart: 746 of them proposed changes to what the FDA described, but did not say whether they backed it.
“The FDA's proposed two-axis risk framework is a good starting point; additional factors, however, should be addressed to more fully appraise risk.”College of American Pathologists, a clinicians’ society, filing 0442
Also in the record
Deep divesKey takeaway 6 of 10
Question 18
What’s behind this“CDRH is considering whether it may be appropriate to accept greater premarket uncertainty regarding a GenAI-enabled device’s benefit-risk profile through greater reliance on postmarket monitoring.”Yes, with conditions · 91%*
Only if…
But not for… 49%
1% said yes outright and 7% said no.
* Of the 205 filers who gave a yes or no on Question 18. Another 38 of the 243 who answered gave advice without one.
What’s behind this19% asked for a tested way to pause or roll back the device.
Of the 243 filers who answered Question 18. A filer could name several; each counted only where an answer says so.
What’s behind this65% of clinicians said so, and 46% of industry filers. Named most often: uses that act with no one checking, or make mistakes that can’t be undone.
120 of the 243 filers who answered Question 18, among them 34 of 52 clinicians and 60 of 130 industry filers; counted only where an answer says so. The uses are ranked by a keyword reading, so they carry no counts.
What’s behind thisThe FDA’s question, in its words
“CDRH is considering whether it may be appropriate to accept greater premarket uncertainty regarding a GenAI-enabled device’s benefit-risk profile through greater reliance on postmarket monitoring. Under what conditions might such an approach be appropriate, and what characteristics of a monitoring program would need to be in place to justify reduced premarket evidence? Are there device types or risk profiles for which this approach would not be appropriate?”
Of the 243 filers who answered Question 18, 205 gave a yes or no on whether postmarket monitoring could make up for less premarket testing:
“Trading premarket evidence for postmarket monitoring is defensible only for failure modes the monitoring can actually detect.”David Liebovitz, a researcher, filing 0309
Deep divesKey takeaway 7 of 10
Questions 21 and 24
What’s behind thisDevice maker
A device built on another company’s model
Model maker
AI foundation model
Tested versionNew version
Asked for advance notice of model changes53%
Clinician and patient
Device output
As testedMay change
That held even where hospitals help monitor. 11% said whoever controls a change is responsible, and 22% didn’t say.
Of the 251 filers who answered Question 21, on roles after launch.
What’s behind thisFrom its paper: “Devices built on third-party foundation models raise particular change-control considerations because changes to the underlying model may be initiated by the foundation model developer rather than the device manufacturer.”
53% asked for advance notice of model changes. Of those 138 filers, 129 also wanted the device maker’s own checks.
Of the 260 filers who answered Question 24, counted only where an answer says so.
What’s behind thisThe FDA’s question, in its words
Who carries responsibility after launch, as each filer who answered Question 21 put it: the device maker (169 of 251, the 67% on the screen), whoever controls a given change (28), or didn’t say (54).
What filers asked for on Question 21, of the 251 who answered; a filer is counted only where its answer says so. Asking the FDA to keep the device maker accountable is a request, a different measure from the 67% who named the device maker as the one responsible.
“Manufacturers must keep accountability; hospitals and frontline nurses should not inherit an undefined surveillance burden.”Robert Elvin Domondon, a clinician, filing 0163
When the model changes
“For devices built on third-party foundation models, changes to the underlying model may be initiated by the third-party model developer rather than the device manufacturer. How can a manufacturer detect, evaluate, and respond to such changes in a timely manner? What mechanisms—for example, contractual, technical, or through a PCCP—could provide reasonable assurance that third-party developer-initiated changes do not compromise the safety or effectiveness of the device?”
Of the 260 who answered Question 24; each counted only where an answer says so. 129 of the 138 who asked for the model maker’s notice also asked for at least one check the device maker runs itself (re-test, track, lock or roll back).
A related question, 25, asked whether model makers should keep a voluntary file about their model at the FDA that device makers can cite. Of the 179 who answered, 85 said yes, with changes, 12 said yes, as written, and 2 said no; the rest gave advice without a yes or no. About half asked the FDA to set what such a file must contain.
Also in the record
Deep divesKey takeaway 8 of 10
Questions 1 and 26, and the FDA’s two-axis risk chart
What’s behind this8AI autonomy scales with risk
Its two axes: what the AI does, and how bad a mistake could be.
Of the 316 filers who answered Question 1. The chart is Figure 1 of the FDA’s discussion paper.
What’s behind thisWhether a mistake can be undone led at 49%; safeguards further down the line came next at 42%.
Of the 316 filers who answered Question 1. Many named several factors, so the rows add up to more than 71%.
What’s behind thisRecords of what the agent did, testing of the whole chain and limits on what it can do each drew about a third.
Of the 232 filers who answered Question 26, on AI agents; 137 named two or more.
What’s behind thisThe FDA’s questions, in its words
The chart is Figure 1 of the FDA’s discussion paper: activity on one axis, the consequences of relying on a wrong output on the other, with risk rising from the lower left to the upper right.
Safeguards proposed for AI agents, among the 232 filers who answered Question 26 (each counted only where an answer says so):
Deep divesKey takeaway 9 of 10
9No single standard emerged for measuring AI devices
An AI device should be measured against
The FDA’s own examples
The top answer was different standards for different risks or devices. The three options the FDA named each drew under 10%.
* Of the 129 filers who named a default standard; the other 112 of the 241 who answered Question 14 named none. Separately, 49 asked for specialists to judge specialist tasks.
What’s behind thisThe FDA also asked whether to test the AI alone or with the clinician using it. 24% also wanted the AI’s own results reported.
Of the 241 filers who answered Question 14; each counted only where an answer says so.
What’s behind thisThe FDA’s question, in its words
“For open-ended device outputs, how should performance comparators and acceptance criteria be selected? When a panel of qualified clinicians serves as the comparator, how should the applicable standard (for example, the standard of care versus the performance of a median clinician in practice) be defined and justified? Should generalist or specialist physicians be used as a performance standard? When should human-AI team performance, rather than the device operating alone, serve as the basis for evaluation?”
The FDA’s question named three possible standards: the standard of care, a typical (“median”) clinician, and generalist or specialist doctors. Of the 129 filers who named a default standard, 11 chose the standard of care (9%), 2 a typical clinician and 11 specialists or expert panels.
No kind of filer gave any one standard half its answers.
“Depending on the claim, the relevant comparator may be qualified clinicians, standard care, accepted technology, or a human-AI team.”Johnson & Johnson, on Question 14, filing 0288
Also in the record
Deep divesKey takeaway 10 of 10
Every filing read for content
What’s behind thisFDA discussion paper
Considerations for the Regulation of Generative AI-Enabled Medical Devices
46% of filers raised at least one of these
None is named in the FDA’s paper
Patient data privacy led at 15%, then cybersecurity at 13% and liability for harm at 11%.
230 of the 504 filers whose filings could be read for content; five filings were set aside, four not readable for content and one restricted. None of these words appears in the paper: privacy, security, liability, cost, workload, reimbursement.
What’s behind thisShare of each group that raised at least one of the six issues:
Clinicians 64 of 123, Industry 107 of 229, Researchers 15 of 40; Patients 2 of 9. Base: the 504 filers whose filings were read for content.
The issues differed by who raised them. Clinicians raised liability (20% against 6% of industry filers) and staff workload (21% against 6%) more often; industry filers made up 36 of the 51 who raised the cost to small companies.
What the paper does say near these issues:
On attacks, among the skills to test: “Relevant testing may include adversarial prompting, prompt injection, emotional-manipulation scenarios, and multi-turn conversations”
And for agents: “resistance to prompt injection through user inputs, retrieved content, and tool outputs”
On cost to industry: “enable a nimble regulatory approach that employs least burdensome principles”
On legal responsibility, citing Freyer et al.: “human clinicians face professional, legal, and reputational consequences when expected standards are not met, and arguing that analogous mechanisms are needed for flawed GenAI-enabled devices”
The six words were searched in the full text of the FDA’s discussion paper, any case, including longer forms (privacy, security and cybersecurity, liability and liable, cost and costs, workload, reimbursement). Question 21 asks how to keep “manufacturer accountability” when others share monitoring. Legal consequences come up once, in a cited study, quoted above. The paper lists payers among the parties around a device, but never discusses the cost to small companies or reimbursement. Some of these issues are mostly handled outside the FDA: Medicare and insurers set payment for devices.
Published November 10, 2026, version 1.0. Covers every comment the FDA posted through November 3, 2026.
Read the full methodologyResults by question
Found an error? Write to support@recovry.ai. Corrections will be listed here with the date they were made.
RecovryAI is developing FDA-regulated AI for post-operative recovery. This research is intended to bring greater understanding, visibility and discussion to the regulation of AI in health care.
AI models read and sorted every comment on the docket against one written set of categories. A reviewer at RecovryAI checked the results, and the published figures and quotes, against the filings.
The short version
Docket FDA-2026-N-7874 · comments filed August 18 to October 19, 2026 · coded by Claude Opus 5.5 (Anthropic), with GPT-5.6-sol (OpenAI) on a fifth of filings · reviewed and published by RecovryAI, which also filed a comment. The eight sections below give the full detail.
Every comment on the docket was read, not a sample: all 530 filings. Exact repeats were counted once, and five filings could not be used for content.
| 540 | records the FDA posted to the docket, filed from August 18 to the close on October 19, 2026 |
| −10 | exact repeats of an earlier filing, counted once |
| 530 | filings, each frozen as plain text and coded |
| −4 | not readable for content: 0073 reproduces the FDA’s paper, 0040 is off topic, 0280 and 0382 have no extractable text |
| −1 | restricted: the author of 0019 asked that it not be used without consent. It is counted in the totals and in the overall-position tally, left out of the content figures, and never quoted |
| 525 | filings read for content, from 504 filers |
Each filing’s text is the comment as posted plus every attachment. 14 scanned attachments had no text layer and were read by OCR. The text was frozen in four batches as the FDA posted comments: 123 filings in early October, then 70, 331 and 6 more on October 29, November 1 and November 4. Coders worked only from the frozen text. Together the filings run to 1.7 million words; the median filing is 2,077 words.
Comments the FDA posts after November 3, 2026 are not in these figures.
Each figure counts one of three things, and says which: filings (the comments), filers (who sent them) or requests (the specific changes asked for).
Filers are grouped by the sector coded from their own text, not by the category they picked on the FDA’s form (with 25 reviewed corrections that apply one rule: care providers and their associations count as Clinicians, companies that build or sell products as Industry), into five groups used throughout: Industry 230; Clinicians 124 (clinicians, professional societies and care providers); Researchers 40; Patients 9 (patients and patient groups); Other 106. Where a takeaway compares kinds of filer (startup, large company, trade group and so on), it uses the docket tracker’s own classification of each filing.
A count of comments is not a vote: filers chose to write, and many answered only some of the 26 questions.
AI models sorted every comment against one written set of categories, fixed before coding began, and every code had to quote the filing word for word. A reviewer at RecovryAI checked the results.
The codebook sets out each field, its allowed values and the words that trigger them. Version 1.2 was frozen on October 9, 2026, before full coding began, and was not changed after.
A reviewer at RecovryAI checked the results. RecovryAI wrote and refined the codebook, reviewed the models’ output over several rounds, and checked the published figures and quotes against the filings. No person re-coded the filings by hand.
The coders were Claude Opus 5.5 (Anthropic), which coded every filing, and GPT-5.6-sol run through the Codex CLI (OpenAI), which coded 111 of the 530 (21%) independently. Both ran at high reasoning effort, under these conditions:
Where two coders disagreed, an adjudicator (a separate model run) saw both records labelled only X and Y, never which model wrote which, and settled each disagreement from the filing text, citing a quote. The plan was two model families on every filing. OpenAI usage limits, then a decision on November 1 to finish with Claude alone, left 111 filings with coders from both families. The 271 single-coded filings are from the last two batches; a 20% sample of the November 1 batch (66 filings, drawn with a fixed seed) was coded twice and adjudicated to check them.
Three measures sit outside that codebook:
Where two coders worked independently, agreement was measured. Measures with strong agreement are reported as findings; weaker ones are marked tentative or only described.
Krippendorff’s alpha (α) scores agreement after allowing for chance: 1 is perfect, 0 is chance. A field is reported as a finding when coders and adjudicators both reach 0.80; as tentative when either falls short but one reaches 0.667; and below that it is described, not counted. Fields on the codebook’s descriptive-only list are described whatever their α.
| Field | CodersClaude vs OpenAI, 65 filings | AdjudicatorsClaude vs OpenAI, 26 filings | Used as |
|---|---|---|---|
| Verdict on the FDA’s approach (yes, no, split) | 0.9196% same code | 0.9799% same code | Reported as a finding |
| Which questions a filing answered | 0.9296% same code | 0.9497% same code | Reported as a finding |
| Sector of the filer | 0.8588% same code | 0.9596% same code | Reported as a finding |
| Which way a request pushes oversight | 0.5376% same code | 0.8692% same code | Tentative: coders below 0.80 |
| Topics a filing makes a main subject | 0.8097% same code | 0.9098% same code | Counts only, by the topic rules |
The answer scales on Questions 3 and 18 had too few disagreements to compute α, and the decision points on Questions 2, 6, 14 and 21 are counts only, by the reporting rules; where both coders recorded a decision-point position, they gave the same one 90% of the time (51 cases). The adjudicator pair was measured on the 26-filing pilot run under the frozen codebook.
What the 96% covers. The 65 filings are those among the first 123 filed that both models coded, leaving out the 37 used to write the codebook. They are not a random sample: 25 are from startups, 21 from consultants and 9 from clinicians, and none is from a large company, a trade group or a health system. On verdicts the two coders gave the same code in 154 of 161 comparisons; in 90 of those, both found advice without a yes or no. Counting yes as written and yes with changes as different answers, and looking only where either coder found a yes or no, they matched in 58 of 72 comparisons (81%).
For each takeaway: what was counted, out of whom, and how far to trust it.
Open any takeaway to see what was counted, out of what, and how far the coding behind it can be trusted.
RecovryAI filed a comment on this docket. Its filing is counted like any other and never quoted, and without it no headline figure moves by more than half a point.
RecovryAI is developing FDA-regulated AI for post-operative recovery, commissioned this analysis, and filed a comment on this docket. These are the guards against that interest shaping the result.
| Figure | With RecovryAI’s filing | Without it (0229) |
|---|---|---|
| Filers who said yes, with changes (takeaway 5) | 87.0% | 87.0% |
| Filers who said yes, as written (takeaway 5) | 7.3% | 7.3% |
| Filers asking for more oversight (takeaway 3) | 53.9% | 54.0% |
| Filers asking for less oversight (takeaway 3) | 13.6% | 13.6% |
| Yes, with conditions, on Question 18 (takeaway 6) | 91.2% | 91.2% |
| Named the device maker as responsible (takeaway 7) | 67.3% | 67.2% |
| Raised an issue the paper doesn’t name (takeaway 10) | 45.6% | 45.7% |
None of these seven figures moves by more than half a point. Recomputed for this section from the same records.
The docket is not a poll, AI did the coding, and some measures are weaker than others.
Every quote links to its filing, so any figure can be checked against the source.
Filing numbers are the FDA’s own: FDA-2026-N-7874 followed by four digits.