November 2026

When AI Listens to Your Doctor’s Appointment

November 6, 2026   |   Charles Schelle

 Imagine you're sitting at your doctor's office talking about something personal. Maybe it's a new symptom. Maybe it's your mental health. Maybe it's something you haven't told anyone else.

Healthcare providers recording and transcribing between themselves and the patient is a common practice.

Now that artificial intelligence tools are commonplace to improve transcription accuracy and advance medicine, where does that information go? How is it helping medicine, and when does review by a doctor happen during the process?

Katherine E.  Goodman, PhD, JD, is an associate professor of epidemiology and public health at the University of Maryland School of Medicine who studies artificial intelligence in healthcare.  Goodman also serves on the faculty of the University of Maryland Institute for Health Computing. She brings a unique experience in public health and patient rights, practicing FDA law in Washington, D.C., prior to pursuing her doctorate in epidemiology. She frequently collaborates with colleagues at the University of Maryland Francis King Carey School of Law.

In a new episode of “The UMB Pulse” podcast, Goodman talks about what patients should know as AI increasingly enters the exam room.

Doctors are increasingly using AI scribes and tools as part of what Goodman sees as a “triple win” satisfying interests of patients, doctors, and the healthcare system.

By eliminating more electronics in the exam room, Goodman said doctors can have more eye contact with patients and focus on what they’re saying, less time is spent afterhours by doctors correcting clinical notes, and healthcare systems are seeing happier clinicians with the reduced burden, along with have better billing procedures.

What Goodman sees as a problem that needs to be addressed is the risk of clinically significant details being omitted by an AI scribe. For example, if a patient shares that they are experiencing leg pain and adds they were on an overseas flight last week, the scribe might omit the flight information that could signal concerns for possible blood clot.

“These tools have not been sufficiently studied, and we need that evaluation to be going because the FDA’s not reviewing them before they’re released to clinics, and many places afterwards are not reviewing them on an ongoing basis to see how well they’re working,” Goodman said.

Here are five questions to take away from the conversations. Listen to new episodes of “The UMB Pulse” wherever you get your podcasts or watch on YouTube. 

Questions and answers have been edited for brevity and clarity.

What is an AI scribe, and how does it work during a doctor’s appointment?

Katherine E. Goodman: When a doctor is using an AI scribe, actually a lot of what’s happening initially is not even AI at all. At base, what happens is it’s usually an application or an app, a piece of software, often on the physician’s phone, in fact. You hit a record button, and that software then records the conversation that occurs during that clinical visit. When the visit is over, you stop the recording, and the backend of that software uses certain statistical tools to convert basically that audio into written text.

At that point, we have a transcript of that clinical conversation, and then that’s the piece really where the generative AI comes into the picture. On the backend, it takes that transcript and it converts it into a summary of the clinical visit, and then sends that draft summary to the physician who will review it, maybe edit it, and then they sign off on it as the final note, the documentation for the visit that is then sent to the patient’s permanent medical record.

Is an AI scribe recording me, and could my health information be used to train AI?

Goodman: It’s one of the most important questions. And I would say that unfortunately, there is no clear short answer. The policies as to what is being recorded and how it’s being stored and preserved or not by AI scribes varies widely by healthcare system, by the actual tool or the company or what we call the vendor that’s making the tool, of which there’s dozens of different AI scribe companies.

Some products will create an actual audio file of the conversation and then convert that into a written transcript. Other products, if it’s captured, it’s not preserved. It’s basically deleted in real time. Audio might be purged very quickly while transcripts go to live on, perhaps not at the healthcare system level, but on the back end with the vendor or with the company.

What I do want to underscore though is that I think patients shouldn’t be overly fearful of the privacy risks of this technology. By and large right now, anecdotally, many healthcare systems are not preserving these data at all. And when vendors or companies on the backend may keep them, basically universally they are de-identifying them. So, what that means is they’re stripping out provider names, patient names, other identifying information.

Companies won’t come out and say this publicly necessarily, but a lot of them are using these data to train their models. Now, that isn’t necessarily a bad thing. That helps make the models be safer. But I do believe it’s important to understand how are these data entering into a training pool for future AI, for what purposes.

Can I say no to an AI scribe at my doctor’s appointment?

Goodman: Really universally, yes, absolutely, patients can opt out of this. The question I think that’s more material is what are the costs of that opt-out and how easy is it?

Where you give that consent and how transparent it is can vary widely. You go to your visit, they give you an iPad perhaps, or it’s on your phone ahead, and you’re clicking through a million forms. You don’t even know, unfortunately, what you’re signing. So, you may have consented to that without even realizing it. Or maybe you get into the visit and your physician or provider asks, “Is it okay if I record this?”

You can certainly say no. But what then if your provider is so used to using these tools that de facto they aren’t prepared then to pull up your chart in the computer that’s collecting dust in the corner of the room and to actually take their notes the semi-old-fashioned way?

But I would again say as well, opt-out isn’t necessarily universally a good thing. There’s a lot of data to show that human-drafted notes very frequently contain errors. And those things also live on in the medical record. And so it may be that an AI scribe will end up more accurate than a rushed physician.

What happens if an AI scribe gets something wrong in my medical record?

Goodman: Right now, these tools cannot actually put something incorrect in your medical record on their own. What they’re producing is a draft summary of the visit, but that must always be signed off by the human provider, by some licensed human before it enters your medical record. That said, there are data and just general experience to suggest that someone who is rushing or it’s been a few days later, the fact that they signed off on a note, it doesn’t mean that they read it carefully.

The emerging empirical evidence suggests that the tools that are out there vary widely on this. Some are more prone to hallucinations, so basically making up information and putting something into the record that was never spoken or said.

But others are actually more prone to a different type of risk that I consider in fact more concerning, which is clinically significant omissions. There was something that was said in that clinical visit that is clinically relevant to that patient’s care, and it has not made its way into that visit summary.

Once that ends up in the medical record, forever after that, that important piece of information, it’s just absent. It’s gone, and we don’t know that it’s gone. And secondly, the data suggests that omissions are much harder for clinicians to detect and fix.

A lot of medicine, diagnosis happens in the details. A patient may mention calf pain or leg pain, but then right after that say, “Oh yeah, I was on a flight to Singapore last week.” Immediately for that clinician, that’s going to be bringing up concerns for a clot in the leg. But the summary may just say leg pain or calf pain without that other detail about the long-haul flight even though that’s very clinically relevant.

These tools have not been sufficiently studied, and we need that evaluation to be going because the FDA’s not reviewing them before they’re released to clinics, and many places afterwards are not reviewing them on an ongoing basis to see how well they’re working.

What should I ask before agreeing to let an AI scribe listen to my appointment?

Goodman: If I was in your shoes, I’d be asking, “Okay, what data are stored or captured? Is it audio and transcript, or is it just transcript?” And then the question is, how long are these stored, by whom, and for what purposes?

If I know that perhaps audio is deleted but a transcript lives on, and that transcript lives on by the healthcare system and it’s solely used for quality improvement to help study care and these tools to make them safer, then that’s probably good enough for me. But if there’s no answers to that question or no one’s really sure, then maybe that gives me pause and I say, “I’ll consider it in the future, but I’m not going to authorize that today.”

My hope is that we’re at the point where these tools are becoming so normalized that these are not questions that patients should have to be thinking to ask at the last second when they’re already in that visit room.

You get, what, an average of 13 minutes now for a primary care visit, and you’re going to use up two of them just having this conversation? No. I want us to be moving towards a place where this information is disclosed in plain language for patients ahead, there’s FAQs, and so they’re not caught off guard.