Summary 🔗
FDA's August 2026 discussion paper on generative AI-enabled medical devices asks the public 26 questions. I grouped them by subject into eleven proposals, from the two-axis risk framework to foundation model master files, and two language models independently coded every public comment on docket FDA-2026-N-7874 against each one. As of September 20, 2026, the public has filed 98 comments, from 87 distinct submitters. A comment can take up any number of the eleven proposals; counted one submitter and one proposal at a time, the docket holds 411 positions. Of those positions, 88% support FDA's direction with requested changes, 7% support it without changes, and 2% oppose it. The same requests recur: account for reversibility of harm and speed of failure detection, evaluate the deployed system and not the model alone, and keep accountability with the sponsor. In 59 comments both coders see a push toward more oversight; in seven, toward less. The figures update nightly until the docket closes on October 19, 2026.
Materials and methods 🔗
Source material. I retrieved every comment on docket FDA-2026-N-7874 through the Regulations.gov API, with its 84 PDF and Word attachments, and extracted the text: 98 comments and about 280,000 words. One submission is a verbatim copy of FDA's own paper and is coded as taking no position. Submitters who filed more than once are counted once, which gives 87 submitters.
Coding panel. Claude Fable 5.1 and GPT-6 Astra each coded every comment independently from the same written rubric. The eleven proposals are my grouping of FDA's 26 questions; Figure 1 lists the question numbers under each. Question 26, on agentic devices, is not covered by any of the eleven. For each of the eleven proposals a coder recorded whether the comment addresses it, a position (support, support with changes, mixed, oppose), a verbatim passage, and the specific changes requested. Passages were string-matched against the source and discarded if inexact. A "yes, provided" answer and a "no, unless" answer that state equivalent conditions both count as support with changes.
Reconciliation. The models agreed on 94.1% of 1,078 coding decisions (98 comments times eleven proposals) (Cohen's kappa 0.89), and on 97.3% of whether a proposal was addressed at all. Where they differed, each re-read the comment alongside both anonymised codings and decided again, which settled 39 decisions. I adjudicated the remaining 25 (2.3%) myself.
Adjudication rules. The 25 split decisions fell into four patterns, and I decided each pattern once.
- A comment that endorses a proposal and then lists safeguards or implementation details is support with changes, not support (6 decisions). Example: endorsing third-party testing subject to independence and conflict-of-interest criteria.
- A comment that proposes a larger model containing FDA's two risk axes is support with changes, not opposition (2 decisions). One such comment calls the two axes "a good starting point" and adds model and lifecycle layers.
- A one-sentence endorsement inside a comment scoped to other questions does not count as addressing the proposal (3 decisions).
- A concrete request that lands on a proposal's subject counts as support with changes even when the comment never names the proposal (11 decisions). Example: asking for multi-turn evaluation without mentioning patient-facing devices. This is the most consequential rule, and a stricter reader would drop these 11.
Three decisions fit no pattern: a comment that favors third-party testing but doubts it is feasible is mixed; a hedged sentence in a comment that scopes third-party testing out is not addressed; and a recommendation to use synthetic test suites for drift detection counts as a requested change on synthetic data.
Recurring requests. One model named four to seven recurring requests per proposal from the coded material, and that list is frozen. A comment counts toward a request only when both models assign it, so those counts are floors.
I analyzed the paper itself in an earlier article.
Results 🔗
Figure 1 shows the positions on each proposal. Support without requested changes is rare. It is highest for clinical confirmation (6 of 41 submitters) and the generalist versus specialist question (4 of 16), and absent for synthetic data (0 of 23) and the premarket-for-postmarket trade (0 of 30). Selecting a proposal in Figure 1 opens it in Figure 2, which lists the recurring requests and links to the comments behind them.
Figure 1. Positions on the discussion paper's eleven proposals
Unique submitters who take up each proposal, ordered from least to most unqualified support. Bar length is the number of submitters (n at right). Select a row to open that proposal in Figure 2.
Figure 2. What submitters ask FDA to change, by proposal
| Submitter | Position | Summary of position |
|---|
Scope and risk (Questions 1 to 6) 🔗
The two-axis risk framework draws 52 submitters, and 47 want it changed. Twenty-three ask FDA to add reversibility of harm and time to harm as risk modifiers, 15 ask for traceability of an output to its source, and 10 ask for credit when the user can detect an error. On patient-facing devices, 14 submitters ask that conversational devices be evaluated across whole multi-turn exchanges, 10 ask that under-escalation and over-escalation be reported separately, and eight ask FDA not to presume that patient-facing functions are higher risk. On the generalist versus specialist question, the leading request is to base risk on the gap between the task and the user's competence, not on professional title.
Premarket evidence (Questions 7 to 17) 🔗
Competency-based benchmarking draws 49 submitters, 45 with changes. Ten ask for sequestered, contamination-controlled test sets, 10 for evidence that benchmark scores predict clinical performance, and nine for benchmarking of the final deployed configuration and not the foundation model alone. On clinical confirmation, 14 submitters ask FDA to evaluate the human-AI team in realistic workflows, and 10 would use care without the device as the comparator. On synthetic data, the requests are disclosure of the generation method and lineage (7), separate reporting of synthetic and real results (6), and a generator independent of the device's model family (6). On third-party testing, submitters ask for published accreditation criteria (9), conflict-of-interest rules (9), and more than one qualified provider (9).
Postmarket and change control (Questions 18 to 25) 🔗
No submitter accepts the trade of premarket evidence for postmarket monitoring without conditions. Of 30, 28 accept it conditionally and two reject it. Twenty would exclude high-risk functions, 18 require that failure be detectable before harm accumulates, 16 require that harm be reversible, and 13 want monitoring triggers fixed before authorization. Postmarket monitoring is the most discussed proposal (59 submitters). Seventeen caution that shared responsibility across clinicians, institutions, and societies must not dilute the manufacturer's accountability, and 14 want periodic reassessment combined with prespecified event triggers. On change control, 14 submitters ask for pinned, identifiable third-party model versions, 11 would define predetermined change control plans by change envelopes and acceptance criteria instead of enumerated changes, and 11 ask for rollback capability. On foundation model master files, 13 ask FDA to specify minimum content.
My position 🔗
I have argued since the November 2024 Digital Health Advisory Committee meeting that FDA should accept lighter premarket evidence for generative AI devices in exchange for strong postmarket evidence (my remarks are here). Premarket clinical evidence is the largest source of delay for the sponsors I work with, and it is a capital expense paid before a product earns anything. Postmarket surveillance is an operating cost that a prudent manufacturer carries anyway once version 1.0 ships.
The submitters are right that the trade depends on detection latency: if the first observable sign of failure is patient harm, monitoring is a record of injury and not a control. I disagree with excluding high-risk functions as a category. Risk category is a proxy for reversibility and detection latency, and some high-risk functions score well on both. I also agree with the submitters who would define a change control plan by its acceptance criteria. A sponsor cannot predict what a foundation model vendor will change, but it can fix the tests every change must pass.
Limitations 🔗
The sample is self-selected. Trade associations, foundation model developers, and health systems have not yet filed. The coding was done by language models under rules I set, and I have not read every comment myself. Agreement was lower on the overall direction of a comment (kappa 0.70) than on positions, so that count includes only comments where both models agreed.
Updates 🔗
The figures are regenerated nightly. I will revise the text if the distribution changes materially, and once after the docket closes.
About the author 🔗
Yujan Shrestha, MD
Partner
I have led or supported FDA strategy and submissions for 51+ medical devices, most AI-enabled, since co-founding Innolitics in 2012, across 510(k), De Novo, Breakthrough, Q-Sub, and deficiency responses. My work spans radiology AI (CADe, CADx, CADt, quantitative imaging), cardiac and neuro imaging software, dental AI, LLM and foundation model devices, and I have managed or contributed to 65 Innolitics projects. FDA's public clearance letters for AI Metrics (K202229, 2020) and Medweb's Lightning Viewer (K242362, 2024) name me as the sponsor's correspondent. I presented at FDA's first Digital Health Advisory Committee meeting on generative AI in November 2024, I have published 78 articles on innolitics.com, and I built fda.innolitics.com, our public FDA device search tool.
