
By POUYAN GOLSHANI
The Food and Drug Administration is now deciding how to regulate generative AI-enabled medical devices, tools that accept open-ended inputs, produce variable outputs and are often built on third-party foundation models. Its discussion paper asks the right questions. The best evidence for answering them comes from the specialty that has lived with regulated AI the longest. Of 1,430 AI-enabled device authorizations on the FDA’s public list from 1995 through 2025, 1,094 (76.5 percent) were reviewed by the radiology panel, and radiology, cardiovascular and neurology together accounted for 90.6 percent, according to a peer-reviewed analysis I co-authored. Annual authorizations rose from about 1.8 a year to 264 a year. Not one was recorded under a psychiatry panel.
Radiology’s experience points to one lesson above the others: requirements beyond what safety demands limit access, growth and innovation. The technology moves fast. A tool that is only somewhat useful today may improve access and care tomorrow, but only if it is allowed into the real world to prove itself. Too often we lead with product safety at the expense of innovation. Safety matters, but for tools that a clinician reviews before acting, and that do not run autonomously or change what the clinician can see in the underlying study, the responsibility should rest with the practices that use them, not with a premarket review modeled on drugs.
Why radiology went first
Radiology had the conditions that made AI buildable: images that were already digital, a common file standard and systems that route each study to the physician who interprets it. The authorization data describe this pattern; they do not establish its cause, and they measure authorizations, not adoption or patient benefit. The FDA also notes that its list is not comprehensive.
The data also show who brings these products to market. Of 740 companies, 502 (67.8 percent) had a single authorized device, while 13 companies accounted for 17.3 percent of devices. Many are single-product innovators building one software tool or one AI-enabled imaging device. Requirements that a multinational corporation absorbs easily often cannot be absorbed by them, and good ideas do not come only from incumbents.
Physicians are the real-world test
This is the point I most want regulators to hear. Reviewing an AI tool like a drug misreads who uses it. Drugs are consumed by patients, who may never report back. AI tools used in clinical practice are consumed by physicians, who notice quickly when a tool does not help and stop using it. I have watched FDA-cleared products abandoned within weeks because they were not useful in practice. Developers and their founders have their reputations at stake, and clinicians already monitor new products during rollout, authorized or not. Poor performers are noticed and dropped.
Congress has already accepted this logic once. The 21st Century Cures Act removed from FDA device oversight clinical decision-support software whose basis a clinician can independently review. The same reasoning should extend to computer-aided detection that a physician reviews and can override. A radiologist who sees a flag and disagrees with it should be free to override it, and should not be held to the high sensitivity, and sometimes low specificity, of a detection model. The standard of care should remain what other radiologists, or the clinicians who first see a scan, such as emergency physicians, would do, not what an overly sensitive algorithm flagged. The comparison is starker still because most studies in urgent care and emergency departments are first read by nonradiologists, and many outpatient scans wait days for a routine read.
The physical analogy is useful. An acoustic stethoscope does not go through premarket review; an ultrasound unit does, even though the physicians who use it can tell quickly whether it works for their purpose. Neither changes the patient. That distinction matters. Tools that are left inside patients, such as hip replacements and inferior vena cava filters, or that act on tissue, such as electrocautery, warrant real premarket scrutiny. Tools that only inform a physician’s judgment, including AI and imaging, are different, yet they are often regulated as if they were the same.
Judge tools against the care patients actually get
Radiologists know that the alternative to a tool is rarely a subspecialist. For radiographs in urgent care, the first reader is often the urgent care clinician, not a musculoskeletal radiologist. Many smaller hospitals, often serving underserved communities, have no on-site radiology coverage after hours; emergency physicians review the images and request a formal read only in the most serious cases. Routine outpatient studies are another case: by design, some wait days for a first read, and some contain fractures or findings that are life-threatening. Images do not talk, but AI can triage them. The FDA’s discussion paper asks whether performance might be judged against the care likely to occur without the device. It should be. A tool that is imperfect but available immediately may serve patients better than a specialist opinion that arrives days later.
Make pilots easy, not heavy
If government wants to help, it should make it easier for hospitals and developers to find each other and introduce tools, rather than build a heavy pilot apparatus. The first clinical use of a product is the pilot. A decision-support tool that does not match what clinicians are seeing gets dropped in the first weeks. Hospitals can help set endpoints, but even those should stay flexible, because clinicians learn in the field what a tool is good for. Triage tools that flag suspected intracranial hemorrhage are a familiar example: a model not good enough to read a study through can still keep an urgent finding from sitting unaddressed in a worklist.
Generative tools that summarize need a few simple design safeguards, because they can add a finding that is not in the source, drop one that is, or lose a qualifier such as the degree of certainty. Keeping the original report one click from the summary, linking statements to their sources and never smoothing away critical findings or stated uncertainty address most of that risk. These are design features, not grounds for heavier review.
If the FDA wants a role, it could offer a simple, voluntary system for practices to log issues during these clinical introductions, so a practice can look back across its clinicians and see whether a tool is working. That would do more for patients than a new layer of approvals. State rules should follow the same principle. California’s SB 503, signed September 30, 2026, will require the physician offices and facilities that use clinical decision-support tools to monitor them for biased impacts starting January 1, 2027. Duties like that add documentation and medicolegal friction that does little for patient care and may keep physicians from using helpful tools. Auditing can be internal and much simpler, and state and federal expectations should be aligned so that one set of records serves both.
Updates should not restart the clock
Generative AI devices change often. Prompts, retrieval methods, interfaces and underlying models are updated, sometimes by a third-party model developer, and even a small change can make the whole system behave differently. In that sense, every update is a new product. But sending each one back through full review would be cumbersome and would not make anyone safer. A product that has been validated should be able to go through a smaller revalidation step or a short pilot after a major change, and routine changes belong in the manufacturer’s quality system. The FDA’s predetermined change control plans are a start. Sound engineering already provides the tools: record the model version behind each output, rerun the manufacturer’s test set when the model changes and pin a validated version where possible.
The cost of waiting is highest in mental health
Policy should expand AI into fields it has barely reached, not tighten its grip where it is already in use. Mental health is the clearest case. No device has been authorized under a psychiatry panel, access to clinicians is limited and the patients who wait longest often have the fewest options. A tool that routes a patient to the right clinician or the right number in time is low risk and useful, yet the prospect of regulation makes developers afraid to build it. Simple, hard-coded rules can provide strong guardrails, such as routing any mention of self-harm straight to the 988 Suicide and Crisis Lifeline. Mistakes are possible, but patients already search for their symptoms online, ask chatbots and self-medicate, sometimes with dire consequences. And even a tool that clears review today will run on a different underlying model within months.
If patient safety is the concern, we should also count the patients harmed while useful tools wait for approval. Radiologists, as the most experienced users of regulated AI, should help build a framework that judges tools against the care patients actually receive and keeps the door open to the innovators who will build what comes next.
Pouyan Golshani, MD, is an interventional radiologist and the founder of GigHz and Guide.MD. He submitted comments to the FDA on its generative AI discussion paper. Disclosure: The author receives consulting fees, travel support and stock options from SinglePass, Inc., a surgical device company.





