5 min read ยท 1,284 words
This article was written with AI. It was drafted from the sources it cites and checked against the full text of those sources before publishing. How we make articles
Medical AI Regulation: FDA’s 26 Questions on a Doctor-Style Exam
Medical AI regulation now turns on one question: can a generative-AI medical device be evaluated the way medicine evaluates a clinician, and how much proof must arrive before launch rather than after it?
On August 18, 2026, FDA’s Digital Health Center of Excellence released “Considerations for the Regulation of Generative AI-Enabled Medical Devices” under docket FDA-2026-N-7874, with comments due October 19, 2026 (JD Supra). It puts 26 numbered discussion questions to the industry (JD Supra). The document insists it changes nothing: “This paper is intended for discussion purposes only and does not represent draft or final guidance,” and it does not address whether new legal authorities would be necessary (FDA). A disclaimer with a deadline.
Public framing sounds less tentative than the disclaimer. “Generative AI-enabled medical devices are poised to reshape the health technology landscape, and the FDA has an important responsibility to provide thoughtful leadership for this new era,” Rick Abramson, MD, the radiologist who took over as head of the Digital Health Center in February, said in a statement. “This discussion paper advances the frontiers of regulatory science and propels a critical conversation about how to enable beneficial innovation, protect public health and preserve trust” (Radiology Business).
Inside sit two ideas, as Cooley summarizes them: a two-axis framework for assessing risk, and a “competency-based” model for premarket evaluation “inspired by how human clinicians are evaluated and credentialed” (Cooley). CDRH further asks whether it should accept “greater premarket uncertainty,” clearing a device on less evidence up front, when the company agrees to keep testing it in real use (AI Risk Today). Which half is the real proposal?
Both halves, one trade. On this article’s reading, the clinician-style exam and the lighter front door are not rival options but a single package: benchmarking redefines what premarket proof means, and accepted uncertainty moves part of that proof past launch day, into monitoring.
Why the Old Test Breaks
Generative AI, as Radiology Business defines it, is “a type of AI that creates new content such as text, images and audio” (Radiology Business).
Analysis: that is why the doctor metaphor fits. Traditional device review asks whether a tool gives the right output; a licensing board asks whether the person producing it can be trusted on cases never seen before. The docket turns on that shift, from verifying outputs to certifying capability.
What Medical AI Would Need to Prove Before Launch
JD Supra’s summary describes premarket benchmarking as “high-throughput, non-clinical testing of the device in its deployed configuration” across up to ten elements, among them “adversarial prompting and prompt injection, scope maintenance and boundary adherence, calibration and clinical deferral” (JD Supra). The same summary says FDA is weighing “shared, standardized benchmarks rather than bespoke test sets per sponsor,” so that reviewers could compare scores from one submission to the next (JD Supra).
FDA’s model borrows from physician credentialing: set a performance bar, test against real clinical cases, and compare results to a panel of qualified doctors or the median standard of care (Daily Living AI).
On this article’s reading, a licensing exam certifies that competence exists, not every decision a doctor will go on to make. The open question is what competence means when the examiner is a benchmark and the examinee keeps updating itself.
Grading is the unsettled part. For devices whose open-ended outputs have no single correct answer, FDA suggests the comparator could be “a panel of qualified clinicians whose consensus reflects the applicable standard of care, or a median clinician in practice” (JD Supra). A device clearing only the second option “would match the typical practicing doctor rather than best practice,” as AI Risk Today puts it (AI Risk Today). CDRH never defines the term, which returns in a question asking how the standard “should be defined and justified” (AI Risk Today).
The Half That Arrives After Launch
After launch, the paper shifts weight the other way. CDRH signals it may require greater reliance on postmarket monitoring, with the premarket assessment as the baseline: a modified device would be “re-benchmarked against the same capabilities” (JD Supra). The paper also introduces a voluntary Foundation Model Device Master File (JD Supra).
As Linda Malek, a partner in Crowell & Moring’s Health Care and Privacy and Cybersecurity groups, told Corporate Counsel: “You have what is essentially a mechanism by which the product continues to evolve and change over time. That in itself is new and poses risks that they haven’t really seen before” (Corporate Counsel).
The Case for Deeper Scrutiny
Scrutiny’s strongest case appears in a Nature editorial arguing that on whether generative-AI medical products should face deeper scrutiny, “The answer in most cases should be ‘yes'” (Nature).
Analysis: the “yes” side deserves its full weight. A patient cannot cross-examine a model the way they can question a clinician, and a device that changes over time can drift away from whatever was cleared. The paper’s answer is its after-launch half, re-benchmarking, because scrutiny cannot stop at clearance. The disagreement that remains is whether a test written before launch can carry a burden that never ends. Does the metaphor settle the timing of proof? It reads better as an argument about when proof is owed, not only about how it is tested: scores at the door, confirmation during practice.
Not a Rule, and Says So
FDA’s own text fences the paper away from policy. It “is not intended to propose or implement policy changes regarding how CDRH intends to regulate generative AI-enabled devices,” is “not intended to communicate CDRH’s proposed (or final) regulatory expectations, including its expectations for supporting evidence in future marketing submissions,” and is instead “meant to seek early input” (FDA).
Cooley’s summary matches: “The paper is not guidance, but it offers insight into how the agency is thinking about the unique risks and evidentiary challenges posed by genAI-enabled devices” (Cooley).
What comes next is another draft: comments filed through the docket “would inform a draft guidance, which would carry its own comment period before it bound anyone” (AI Risk Today).
What to Watch, and What to File
Advice in this section is this article’s view, built from the facts above. If you hold regulatory or clinical responsibility for a generative-AI health product, four items fit inside an afternoon:
- Draft positions on the premarket-evidence questions first. That block is where the comparator standard lives. Decide what your product could prove today.
- Answer the trade in writing. Less evidence at clearance in exchange for postmarket testing, yes or no. Name your terms: what triggers re-benchmarking, and who owns the shared benchmark your scores sit on.
- Define “median clinician in practice” before someone else does. The docket’s own question is how that standard “should be defined and justified” (AI Risk Today). A sponsor with a defensible definition writes its own grading key.
- Run the security elements now. Point a prompt-injection and adversarial-prompting test set at the deployed configuration and record what happens.
Binding expectations would come through guidance, and CDRH’s fiscal 2027 agenda carries the thread, with finalizing guidance on AI-enabled devices throughout their product lifecycle and on pre-determined change control plans topping the to-do list (RAPS; FDA).
Comments on docket FDA-2026-N-7874 close October 19, 2026 (JD Supra). CDRH plans to publish 11 final guidance documents and three drafts in fiscal 2027 (RAPS). How much proof arrives before launch, and how much after, begins with what reaches the docket before it closes.
