As AI becomes a larger part of endocrine care, endocrinologists need to be aware of the legal considerations that may shape its use in clinical practice. This includes evaluating not only whether an AI tool is available or widely used, but also whether evidence supports its use for a particular patient—and how using or not using the tool may factor into the standard of care, professional liability, and disclosure to patients. 

In part two of his conversation with Johnson Thomas, MD, FSCE, FEAA, Editor-in-Chief of AACE Endocrine AI, Steven Petak, MD, JD, MACE, FACP, examines these issues and how evolving legal standards apply to the use of AI in patient care.  

(Note: The information provided in this interview is for general informational and educational purposes only. It is not intended to constitute and should not be relied upon as legal advice. The following transcript has been edited for clarity and length.)   

Dr. Thomas: We'll talk a little bit more about the legal aspects now and legal terminologies. Standard of care—basically what a court expects a competent physician to do—is shifting from what most physicians do towards what a reasonable physician would do, given the evidence. In plain terms, what does this mean for practicing endocrinologists? What should we look into?  

Dr. Petak: Well, there are two familiar defenses, and they're both becoming weaker and weaker over time. One is that nobody around here uses that tool, and the other one is everybody uses that tool. In 2024 the American Law Institute restatement pushed toward reasonable care by evidence, not just by custom alone. States are going to adopt this differently, so these standards will vary by state.

I would go back to this 1932 case that I talk about in my lectures, involving tugboats without radios. In 1932, radios were not standard. These tugboats lost their barges and had a major liability associated with that. But the courts reasoned that although radios weren’t standard, a prudent operator would have had one. Applying that to medicine today, a validated tool can become part of a reasonable care before it's universally adopted.  

The reverse, though, is also true: Widespread use does not necessarily rescue a weak or poorly validated tool. So, from a practical standpoint, the answer is not to chase every product. You want to follow the quality of the evidence, have specialty guidance, see which patients actually fit that particular intervention, and know the consequences of both using and not using a tool. So, the bottom line is that “everybody does it” and “nobody does it” are not really good defenses now.

Dr. Thomas: If there is a well-validated tool in the guidelines and we are not using it, and something happens, we might become liable or be considered negligent. Is that a correct interpretation?

Dr. Petak: It's legally plausible, but we don't know where that threshold is yet. The fact that a product is called AI does not mean that it is mandatory. So, you want to look for evidence that the tool actually improves meaningful outcomes, that it’s validated in patients like those you’re actually seeing in your clinic, and that the patient is a candidate. You want to have specialty guidance that's credible. And you also want to know that the tool is available and that the workflow can be safe for that patient. If those factors are all present, then not using the tool becomes harder to defend. For example, autonomous diabetic retinopathy systems or a program that identifies vertebral fractures on CT scans illustrate where we’re heading with this. If it consistently finds a preventable disease that clinicians might otherwise miss, then a future plaintiff may ask why it wasn't used. But promising performance alone isn't enough. You have to show clinical benefit, and that the patient specifically can benefit from that tool. AI is going to become part of the standard of care once it’s validated—one case at a time, not one headline at a time. 

Dr. Thomas: As a practicing physician, how does one evaluate an AI model before it's used for clinical practice? Is FDA clearance enough? Should we look for our society's guidelines? How do you recommend that a physician who is planning to implement this evaluate these models?  

Dr. Petak: FDA authorization is an important starting point, but it's not the finish line. You want to determine what applies. Premarket approval is the strongest FDA approval pathway. 510(k) [a premarket notification sent to the FDA by a device maker], which is basically a substantial equivalence, is a less defensible position. You want to ask what the clinical problem is, what the software version was, because software changes over time, so validating it on an initial software version doesn't mean subsequent versions are going to be valid for that patient. 

You want to know what the population was, whether updates are being controlled, whether you're notified of what those updates are. The consequence of error is really critically important. If you have an imperfect administrative tool, that's going to create a problem, but a relatively modest one. If you have a problem that recommends insulin dosing changes that are harmful, the potential harm is going to be much, much greater, and the evidence and monitoring is going to have to be greater as well. So, FDA clearance doesn't mean universally appropriate, bias-free, or safe. You have to correlate it clinically. 

Dr. Thomas: Many times, when we do procedures, we get informed consent from the patient. How do we do that for an AI product, which might be making a decision, or a decision support for the patient.  And then we have AI scribes, which listen in and then transcribe this. I understand there might be a difference in laws depending on which state you are in, but what do you think about getting informed consent for these tools. 

Dr. Petak: The issue really is whether it's material to that patient's decision making. If the AI output significantly can affect the patient's decision, you should document it. I think that's really critically important. You don't want to turn every low-risk administrative use into a lengthy consent ceremony. If the AI is helping to schedule appointments or doing workflow, that's one issue. But if you're using an AI tool that is material to the patient's decision-making, you need to have a brief conversation. And that conversation looks something like this. You could say, “We use an AI-assisted tool to help estimate your risk. I reviewed the result, rather than accepting it automatically. It was validated in the kind of patient that you fall into, but it does have limitations, and we do have alternatives. In Texas, there's particularly stringent criteria now that was just initiated as of September of 2025 for diagnostic AI, and that is that there has to be a disclosure statement that says that the patient decision-making was based on an AI decision, and that you have to disclose that.