Nurses Have Until October 19 to Weigh In on the FDA’s Generative AI Device Framework

Part of Nurse.org’s Nursing AI Watch, our ongoing investigation into how artificial intelligence is reshaping nursing practice. In this contributor feature, nurse educator Susan Deane reads the FDA’s open docket on generative and agentic AI-enabled devices and finds nursing’s voice nearly missing from the record, with weeks left to change that. Disclosure: Susan Deane, EdD, MSN, CNE, RN, is a nurse educator whose consulting practice focuses on AI education for nurses.

On August 18, the FDA’s Digital Health Center of Excellence released a discussion paper called Considerations for the Regulation of Generative AI-Enabled Medical Devices and opened it for public comment. The paper covers large language models, foundation models, and agentic AI systems that plan and act across several steps. It asks the public 26 questions about how these tools should be evaluated before and after they reach patients.

Comments close October 19 under docket FDA-2026-N-7874 on Regulations.gov.

This is not a rule. The FDA says plainly that the paper does not set policy. But discussion papers like this often come before guidance, and guidance shapes what manufacturers must show before a product reaches a hospital. What goes into this docket now can influence the tools nurses are handed in the years ahead.

Nurses will explain these tools’ outputs to patients. We will notice when they are wrong. We will decide what to do next. That view belongs in the record, and right now it is nearly absent.

Want to see more Nurse.org articles in your Google results? Add us as a preferred source.

The FDA’s paper is about 30 pages of regulatory language, but the questions at its core touch everyday nursing work. They fall into four areas: how risky a tool is, how it should prove it works before reaching patients, how it should be watched once it is in use, and what extra issues come with foundation models and agentic systems. Here is what each area asks and where nurses have something to add.

1. How risky is this tool? The FDA proposes sorting generative AI devices along two axes: how much the tool does on its own, and how bad it is when the tool is wrong. A tool that offers general information sits low; one that directs or takes a clinical action sits higher. The paper’s own examples make the point: bad advice about hydrocortisone cream for poison ivy is very different from bad advice about a basal insulin dose. The paper also asks whether tools that talk directly to patients pose different or higher risks than tools used by clinicians, and what safeguards should apply. Nurses work where those lines blur, such as a patient portal message a nurse reviews after the patient has already read it.
2. How should a tool prove it works before it reaches patients? The paper floats a “competency-based” approach inspired by how human clinicians are evaluated and credentialed: benchmark testing first, then confirmation in real clinical use, then ongoing assessment. Every concrete training and credentialing example it gives is physician-oriented: medical licensing, board exams, residency, and fellowship. When it asks who an AI tool’s performance should be compared against, it asks whether that should be generalist or specialist physicians. The paper does not mention nurses. Nursing has its own long tradition of competency validation, and it belongs in this conversation.
3. How do we watch it after it is in use? Generative tools can give different answers to the same question and can change after deployment. The FDA is asking how to catch problems in real-world use, through periodic re-testing, clinician review of real cases, and monitoring for performance drift. Question 21 asks directly what role clinicians, healthcare institutions, and professional societies should play in that monitoring. Nurses are often the first to see when a tool starts getting things wrong, which makes this a question nursing should answer.
4. What about foundation models and agentic systems? Many products are built on a general-purpose model made by another company. Agentic systems go further: they plan, use other software tools, and take actions in sequence. The paper asks how to evaluate these, including whether a system keeps a human checkpoint before an action that cannot be undone. In most clinical workflows, that human checkpoint will be a nurse.

The FDA’s public list of authorized AI-enabled medical devices now has more than 1,500 entries. Most of them do one defined job: flag a possible abnormality on an image, classify a finding, predict a single risk score. They are tested against a fixed task, and they give the same output for the same input.

Generative and agentic tools work differently. They take open-ended input, such as a patient’s question, a chart, or a conversation. They can do many tasks instead of one. Their answers vary from one run to the next. Agentic systems go a step further: they break a goal into steps, call other tools, and act on the results.

That is a hard thing to regulate with rules built for single-purpose software. You cannot fully test an open-ended tool against every question it might be asked. A fixed pre-market test cannot fully predict how a tool will behave once it meets a real unit on a real night shift. The FDA’s own paper names these gaps, and that is why it is asking for help.

For nurses, it comes down to one question. If an AI tool works through several steps to reach a recommendation, who checks its reasoning along the way? Nurses should help answer that before the rules are written, not after.

AI informs. The nurse decides. That only holds if the systems are built, tested, and monitored with the nurse’s decision in mind.

Who Has Filed So Far, and Who Hasn’t

As of October 5, 129 comments were publicly posted to the docket, with Regulations.gov reporting 146 submissions received; seventeen had not yet been published. I read the posted comments to see whose voices are in the record.

Many comments are anonymous or don’t state the author’s role. Among those that do, most come from physicians, from people who have built AI tools, and from technology companies. Professional societies are starting to file: the Congress of Neurological Surgeons and the American College of Allergy, Asthma & Immunology have both submitted society-level responses, and Stanford’s Brainstorm lab for mental health innovation has weighed in.

I have not yet found a submission from a national nursing organization among the posted comments. (Submissions can take days to appear publicly, so late or pending filings would not show yet.) Only two comments clearly identify their authors as registered nurses, and both are thoughtful and grounded in practice. Other nurses may have commented without saying so. Even so, nursing’s voice is hard to find in this record, and that does not reflect the profession that will do much of the day-to-day work with these tools.

Organizations often file in the final week, and I hope nursing groups do. But the record should not depend on a few late filings. Individual nurses can speak from their own practice, and the FDA wants to hear it. The agency’s call for feedback lists clinicians and the public alongside manufacturers and researchers.

>>Listen to The Latest Nurse News Podcast

You do not need to be a regulatory expert, and you do not need to answer all 26 questions. The FDA explicitly says partial responses focused on your area of experience are welcome.

1. Read the paper’s overview. Start with the FDA’s discussion paper page; the full paper is there as a PDF. The introduction and the question lists are the parts that matter most.
2. Pick one to three questions you know firsthand. Good starting points for nurses:
   • Tools that talk directly to patients, and what happens when a nurse has to correct them
   • What “competency” should mean for an AI tool, drawing on how nurses are validated
   • How problems should be reported once a tool is in use, and who should report them
   • Where a human checkpoint belongs before an agentic system acts
3. Write from your practice. Name your role and setting. Give a concrete example. One clear paragraph grounded in real care is worth more than a page of general concern.
4. Add a one-line disclaimer if you’re concerned about your employer. You can comment as an individual, not on behalf of your hospital, school, or professional society. A single sentence at the top makes that clear, for example: “These comments reflect my own views as a registered nurse and do not represent the views of my employer or any organization I am affiliated with.” You can describe your role and setting without naming your employer.
5. Submit on Regulations.gov. Go directly to docket FDA-2026-N-7874, select “Comment,” and paste or upload your response. You can comment as an individual or on behalf of an organization.
6. Do it before October 19. Comments are public, so leave out any patient information.

If you teach, consider making this a class or unit activity. Students can read one section, discuss it, and draft a response together. It is a real-world lesson in how nursing shapes policy.

The tools nurses will be asked to trust are being defined right now. Once guidance is written, changing it is slow. Right now, one well-grounded paragraph from a practicing nurse can become part of the record the FDA works from.

There are a few weeks left. Pick one question, write what you know, and submit it.

Related Nursing AI Watch Analysis:

More from Nurse.org’s reimbursement series:

🤔 Which of the FDA’s questions matters most in your practice, and have you submitted a comment yet? Share your experience in the comments below.

Nurse.org Analysis