FDA's generative AI medical device paper: what its risk grid means for chatbot makers before October 19

On August 18, FDA's Center for Devices and Radiological Health (CDRH) published a 31-page discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices, and opened docket FDA-2026-N-7874 for feedback. It ends with 26 numbered questions. Comments are due October 19, 2026.

Start with what the paper is. It is not draft guidance, and it says so in its first paragraph. It does not propose policy or say what evidence FDA will expect in future marketing submissions. It also sets aside whether FDA has the legal authority for the approaches it describes. Nothing in it is enforceable.

It still matters if you build software that uses a large language model to talk to patients or clinicians. The paper sketches how CDRH reviewers are thinking about risk and evidence, and that thinking tends to show up in pre-submission meetings well before it shows up in guidance. That last point is our read; the paper doesn't say it.

The risk grid: what the software does, and what happens if it's wrong

The paper proposes a two-axis grid as an "organizing heuristic." One axis is device activity, running from non-directive information through action-directing information to action-taking, with autonomy increasing as you go. The other is consequence: how much harm follows if someone relies on an incorrect output. Risk rises toward the upper right.

The useful part for product teams is how finely CDRH slices "informational." The paper walks through four versions of the same answer about blood pressure medication:

• General information: "lisinopril dosages are sometimes increased when blood pressure remains above the treatment goal"
• Connecting practice to the user: "in situations similar to this, clinicians often increase the lisinopril dosage"
• An endorsement: "I recommend increasing the lisinopril dosage"
• An instruction: "increase the lisinopril from 10 mg to 20 mg daily"

CDRH says the distinction "would likely involve a continuum," and it is considering treating "action-directing" outputs as higher risk. It is weighing whether directiveness depends on "the substance and context of the output, not solely on whether it uses words such as 'recommend,' 'should,' or 'consider.'" It is also considering whether a patient-facing answer becomes any less directive because it adds "talk to your doctor" or "I am not a medical professional." If your safety plan leans on a disclaimer appended to every response, that plan is the thing this paper questions.

Four other considerations move a function up the grid:

Conversations drift. A chatbot can start with general information and end up directing action several turns later. CDRH is considering assessing risk "across realistic conversational trajectories," not one function at a time.

Patients can't check the answer. Patient-facing functions might shift up the consequence axis because patients can't identify a wrong output the way a clinician can. The paper also names the other side, the benefit of putting clinical information in patients' hands, and asks where the line should sit.

Measurement outputs. GenAI-enabled in vitro diagnostic and signal processing functions may be higher risk even though they look non-directive, because the user usually can't evaluate the basis for the number.

Escalation errors run both ways. For functions that advise whether to seek emergency care, CDRH is considering counting both under-escalation and over-escalation as failures.

Premarket review modeled on how doctors are credentialed

CDRH's starting premise, carried over from its November 2024 Digital Health Advisory Committee meeting, is that it may be unreasonable to test every input an open-ended device could see. So it floats a "competency-based" approach with two parts. Device benchmarking is high-throughput, non-clinical testing against prespecified acceptance criteria. Appendix A lists ten elements in four groups: safety (escalation, staying in scope, communicating uncertainty), clinical proficiency (knowledge, information gathering, quantitative analysis, communication), generalizability (reproducibility, subgroup performance), and one element for agentic devices. Clinical confirmation follows, and it need not be a prospective trial in every case. Options run from retrospective evaluation on real patient inputs, through shadow deployment and standardized patient actors, up to randomized trials.

Three details stand out. FDA would evaluate "the final user-facing device, as configured and intended to be deployed," not the foundation model standing alone. Expert adjudicators should be independent of both the sponsor and the model developer, and the paper says that would still apply "when the expert adjudicator is itself an LLM." For open-ended outputs with no single right answer, the comparator might be a panel of clinicians or "a median clinician in practice."

How it fits with the AI guidance already on the books

The paper builds on, and doesn't replace, existing CDRH documents:

• PCCPs. The predetermined change control plan guidance for AI-enabled devices was finalized in December 2024 (current version August 2025). It assumes you can describe planned modifications in advance. The paper calls a PCCP "one mechanism" for GenAI changes, then asks in Question 23 how PCCP concepts would work "when the nature or scope of future modifications cannot be fully prespecified." Question 24 asks how a manufacturer detects and responds when the third-party model developer changes the underlying model.
• Lifecycle guidance. The January 2025 draft guidance on AI-enabled device software functions (docket FDA-2024-D-4488) was still a draft as of August 2026.
• Scope. Whether a function is a device at all still runs through the Clinical Decision Support and General Wellness guidances, both reissued as final in January 2026. The paper notes its measurement-function reasoning "seems consistent with" the independent-review criterion of the CDS exclusion in section 520(o)(1)(E).
• PCCPs in regulation. On June 17, 2026, FDA classified radiological machine learning-based quantitative imaging software with a PCCP into class II with special controls. Change control is written into the device type itself. That is how FDA handles adaptive AI today: changes described and bounded in advance.

The postmarket section raises one idea current guidance doesn't. Question 18 asks whether FDA should "accept greater premarket uncertainty" in exchange for "greater reliance on postmarket monitoring," using periodic re-benchmarking against the premarket baseline, sampled clinician review, and drift monitoring.

Foundation model master files

CDRH asks about a voluntary Foundation Model Master File under the existing Device Master File program. Model developers could file architecture, training data provenance, known failure modes, benchmark results, update notification commitments, and audit log availability, which device sponsors could then reference with permission. Filing would not authorize the model for any use. FDA concedes the obvious problem in Question 25: model developers "may have limited incentive to disclose safety-relevant information."

What to do before October 19

1. Inventory each GenAI function and place it on both axes. First decide whether it is a device function at all under the CDS and wellness guidances. Then rate it for directiveness on the lisinopril scale above, without counting the disclaimer.
2. Test conversations, not single prompts. Build multi-turn test sets where an informational exchange drifts toward advice. That is where the paper expects risk to show up.
3. Read your model vendor contract. Look for version pinning, advance notice of model changes, and audit logs. Update notification and audit logs are both on CDRH's list of master file contents, and together they are your answer to Question 24.
4. Design benchmarks you can rerun. If re-benchmarking becomes the change-control baseline, a test set you can repeat after every model update is worth more than a one-time validation report.
5. Decide whether to comment. You don't have to answer all 26. Questions 2, 5, 23, and 24 carry the most consequence for conversational products. Submit at regulations.gov, docket FDA-2026-N-7874.

If you use GenAI only for clinical documentation, patient outreach, or clinical workflow support: the paper says those uses may not be functions that are the focus of FDA's device oversight. No action required, though it is worth confirming each function stays there.

We don't know what comes after the comment period. The paper commits to no timeline and no next document.

Sources

• FDA CDRH, Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback, August 18, 2026, Docket FDA-2026-N-7874
• FDA press release, FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices, August 18, 2026
• FDA, Executive Summary, Digital Health Advisory Committee, November 2024
• Federal Register 2024-28361 (PCCP guidance for AI-enabled devices, December 4, 2024); 2026-12166 (classification order, June 17, 2026)

Policy Canary tracks every FDA action, warning letters, recalls, guidance, and rules, against the specific products you make, by name and by ingredient. See what we would flag for your products.