Here is a question nobody asked you when you started using that AI tool. Is it a medical device?
Not the software suite your EHR calls an upgrade. Not the ambient scribing product the practice manager piloted last quarter. The specific algorithm that processes clinical data and returns something that influences a clinical decision. That thing.
If it meets the FDA's statutory definition of a medical device, which a significant and growing class of clinical AI products does, the rules that govern it are categorically different from the rules that govern most software you work with. What the vendor can claim. What evidence they had to produce before it reached your workflow. What your liability looks like when the output is wrong.
Most physicians cannot answer that question about the tools they use daily. That is not a knowledge failure. The product probably did not tell you. The vendor's pitch almost certainly did not lead with it. The procurement committee may not have known to ask.
Here is what changes when the answer is yes.
The FDA's definition of a medical device has been federal law since 1976. It covers any instrument, apparatus, or article intended for use in the diagnosis, cure, mitigation, treatment, or prevention of disease, or intended to affect the structure or function of the body.
In 2016, Congress drew a clearer line for software specifically. The 21st Century Cures Act excluded certain clinical decision support software from device regulation: tools that display information for a clinician to independently review and are not intended to replace clinical judgment. That exclusion was designed for reference tools. Look up a drug interaction. Pull a clinical guideline. The physician reads it and decides.
The exclusion does not cover software that drives the clinical decision itself. An algorithm that flags a radiology image for cancer, triages a patient for sepsis, or recommends a medication based on clinical parameters, and where the clinician is expected to act on that output, is not a reference tool. It is a decision-support system making a clinical recommendation.
"That is a different regulatory category, and it matters. When a vendor says 'FDA cleared,' ask which pathway. The answer is not confidential."
When an AI tool is a medical device, it goes through one of three FDA approval pathways. Each tells you something specific about how much evidence the manufacturer had to produce before that product reached your workflow.
| Pathway | What it means | Evidence required |
|---|---|---|
| 510(k) Most Common |
Substantially equivalent to a predicate device already on the market | No clinical trial required. Requires a predicate. If the prior device was cleared, and the new one is substantially similar, it clears. |
| De Novo | Novel device with no clear predicate; FDA creates a new risk classification | More rigorous than 510(k). FDA evaluates safety and effectiveness for the novel risk class. Still not a clinical trial. |
| PMA Gold Standard |
Premarket Approval; required for highest-risk devices | Valid scientific evidence, typically clinical data, demonstrating safety and effectiveness. Very few AI products clear this bar. If yours did, the manufacturer will tell you. |
510(k) and "FDA approved via PMA" are not the same sentence. Many clinical AI products go through 510(k). That is not a scandal. It is also not a clinical trial. The distinction matters when you are deciding how much weight to put on the output.
Here is the part the pitch deck does not cover.
If an FDA-cleared AI tool recommends a course of action and a clinician follows that recommendation, and the patient is harmed, the liability question becomes a clinical negligence analysis rooted in professional standard of care. The FDA cleared the tool. That does not protect the clinician who applied it to a specific patient in a specific context without independent verification.
The argument that the AI said so has never been a successful defense in a medical malpractice case. It is unlikely to become one now.
What does change is the upstream accountability question. If the tool was FDA cleared under false or incomplete evidence, if it performed differently in your patient population than in the study population it was validated on, if the vendor misrepresented its intended use, those are product liability and regulatory enforcement questions. They are not the question being asked in the exam room when something goes wrong.
"The professional liability question is still yours. The FDA clearance does not reassign it."
Three questions worth running down on the clinical AI products in your current stack.
The answers are not complicated to obtain. The reason most organizations do not have them is that procurement happened at the vendor-relations layer, not the clinical-governance layer. That is the gap this week's episode of The Clinical Realist names directly.
Episode 20, "When the Algorithm Is a Device: AI, the FDA, and the Question Almost Nobody Is Asking," walks through all three questions in full, with the liability framework that makes the regulatory question actionable. Link below.
If you are a health system leader who wants to audit your clinical AI stack against FDA regulatory status, intended use, and liability exposure, the structural diagnostic starts with a one-hour conversation.
Book a Discovery and Clarity Session →Dr. Sarah Matt, MD, MBA, is a surgery-trained physician-executive currently practicing internal medicine (charity care). She advises health systems and health technology organizations on clinical AI governance, vendor selection, and physician leadership strategy.