Who actually files — the room where it is decided, and how few people are in it
A short file about a large number and a small one. Three hundred and forty million people live in the United States. Fifty-one of them filed on the federal docket that will shape how artificial intelligence may be used in medical devices. Twenty-one of the fifty-one were industry.
Companion to the reading notes, which name every filer and read what they said, and to the field guide, which explains the mechanics — including the fact that inverts most people’s civic instinct: the process is not a vote.
1. The count
Docket FDA-2024-D-4488. Fifty-one comments. By what the filer is:
| Who filed | n |
|---|---|
| Industry trade associations and coalitions | 11 |
| Companies | 10 |
| Clinician and professional bodies | 10 |
| Named private individuals | 13 |
| Patient and public-interest organizations | 4 |
| Anonymous | 3 |
Twenty-one of fifty-one — a clear plurality — are the regulated industry or its associations. Four are patient-side. The judgment calls behind those categories are stated and disputable in the reading notes, which also records that the substance of 29 of the 51 has not been read by this project.
What this file does not claim. That any filer acted improperly — filing on a public docket is exactly what a docket is for, and the industry submissions are frequently the most detailed and technically serious documents in the file. That the agency counts comments like votes; it does not, and should not. That fifty-one is unusually low; it is entirely ordinary, which is the point.
2. What was being decided in that room
Not a technicality. The guidance concerns how AI-enabled devices are described, validated, labeled, monitored after release, and how much anyone downstream is entitled to know about the model inside.
Set that beside who is on the other end. From a Pew Research Center survey of 5,119 US adults, 17–23 February 2026:
- 49% of American adults use AI chatbots — up from 33% in 2024
- 20% — one in five — use one for medical advice
- 10% use one for emotional support or advice
- 59% are not confident that US companies will develop and use these tools responsibly
Source: Pew Research Center, 17 June 2026 — ⚠ R under the confidence rubric.
One in ten American adults brings emotional distress to one of these systems. Six in ten do not trust the companies that make them. People are not using them out of confidence. They are using them because they are there, and free, and answer at three in the morning.
These are decisions about how such systems may be used on children, on minority groups, and on all of us in our worst moments — in a clinic, in a benefits assessment, in a classroom, at 3 a.m. And the room where they are settled held fifty-one people, of whom twenty-one sell the products.
3. The demand exists. The channel does not carry it.
The obvious reading is that the public did not turn up, and it is wrong. People are asking for this constantly — 59% of American adults tell Pew they are not confident these companies will develop and use these tools responsibly, and that is a settled majority position rather than a mood. Demand for accountability is not scarce.
What is scarce is any route from that demand to a statute. A poll is not a comment. A majority is not a filing. The docket is the mechanism by which a view becomes part of an administrative record, and it is the one place the majority is absent — not because it has no view, but because nothing carries it there.
That distinction matters because it points at a different fix.
Nobody knows the docket exists. Notice is given in the Federal Register. There is no mechanism by which a person using a chatbot for medical advice would ever learn that the rules for such tools were open for comment for ninety days.
The format excludes. The field guide sets out what a useful comment requires: the docket number, the guidance line numbers, the relevant standards, and a proposed alternative. A trade association has staff who do this professionally. A patient organization has one policy person. A member of the public has an afternoon and no idea that line numbers are the currency.
And the incentive is asymmetric by construction. A rule that costs an industry ten million dollars is worth a specialist submission. The same rule spread across three hundred million people costs each of them a few pennies and an outcome they will never trace to a decision. This is ordinary collective-action arithmetic and it would operate identically if every party were acting in perfect good faith. No villain is required, and this file does not supply one.
The consequence is nonetheless real: the people who get heard are the people who file, and the people who file are, overwhelmingly, the people with a commercial reason to.
4. The shape underneath, which is the finding
Follow a harm and follow a duty and watch them travel in opposite directions.
The harm moves outward and downward. A model is trained. It is released. It is built into a product, deployed by an institution, operated by a professional, and finally used by a person at their most vulnerable. At each step it reaches more people, is understood by fewer, and lands further from anyone who chose anything.
The accountability moves upward and outward. The user accepted the terms. The professional exercised judgment. The institution validated the deployment. The vendor integrated a component. The developer disclosed what it chose to disclose. At every step the answer to who is responsible is somebody further up — until you reach the top, where no statute is waiting.
Harms go down. Accountability goes up. They pass each other.
And this project can put a number on the far end of that journey rather than asserting it. Across six frontier regimes read one at a time — four enacted, two introduced — not one places a duty on a named natural person, and the fifty-one commenters on this docket, in a file whose whole subject is safety, named no upstream person either. Sources: the bill census; the reading notes, finding F1.
5. What follows from it, stated modestly
This is not an argument that dockets are a sham, and the corrective is not cynicism. Comments from patient bodies on this docket asked for algorithmic impact assessments and a public registry; clinician bodies asked for mandatory model cards and pediatric statements. Those are serious asks made through the proper channel by people doing the work.
It is an argument about where a duty should sit. A process that depends on the affected public showing up will systematically under-represent them, for reasons of notice, format and arithmetic rather than indifference. A duty attached to a named person upstream does not depend on anyone showing up. It is there whether or not the fifty-second comment is ever filed.
That is the whole claim: not that the public should participate more, but that their protection should not be conditional on it.
Corrections to the project contact; they enter the errata register with the fix attached and permanent credit. The composition table is this project’s categorization of a public file and the judgment calls are listed where they are made — anyone who would draw the lines differently has improved this page.