When a patient came to Jon Ebbert, MD, with dizziness, he took her history, examined her and told her he suspected benign paroxysmal positional vertigo.
“She looked at me and said, ‘That’s what ChatGPT told me,'” said Dr. Ebbert, a general internist and medical director of the Systems Innovation Accelerator at Rochester, Minn.-based Mayo Clinic.
He asked whether the chatbot had told her what to do about it. It had not. “We could move pretty quickly into the part where I added the most value: confirming that the diagnosis fit her presentation, explaining what it meant in her specific situation and talking about how to manage it,” Dr. Ebbert said.
Scenes like that are becoming routine. Patients are arriving with diagnoses, treatment options and question lists from ChatGPT, Claude and Gemini. The habit itself is not new; clinicians are very familiar with “Dr. Google,” but AI changes the conversation enough that physician leaders at Mayo Clinic and Los Angeles-based Cedars-Sinai told Becker’s their health systems are figuring out how physicians should respond.
“AI feels much more personalized and conversational, and it can present information confidently,” Dr. Ebbert said. “That can make it feel more authoritative than a traditional Internet search, even when the answer may be incomplete or lack clinical context.”
Yaron Elad, MD, chief medical informatics officer at Cedars-Sinai and a practicing cardiologist, has noticed the same shift. “I think patients probably view AI as a little bit more authoritative than they ever felt that their Google searches were before,” he said.
In his view, that confidence is often earned. Chatbots tend to draw on stronger sources and cite them, while search engines could send patients to sites that were not evidence-based.
At Mayo Clinic, AI output tends to arrive in three forms, Dr. Ebbert said: a list of possible diagnoses based on symptoms, an explanation of a condition the patient already has or treatment options and questions the AI suggested they raise.
Sometimes the source is obvious, Dr. Elad said, because patients bring a printout of their query. Other times it surfaces as a list of questions. He does not hesitate to ask where an idea came from.
“Usually, they’re apologetic about it, like, ‘Oh, you know, I’m not questioning you, but this is what ChatGPT said,'” he said.
For patients with complicated conditions, those lists can run long, Dr. Ebbert said, citing cases his colleagues have described.
“The issue isn’t necessarily that patients are asking questions, which is a good thing,” Dr. Ebbert said. “It’s that an AI system can generate an almost unlimited number of plausible questions without knowing which avenues have already been evaluated or which are actually relevant to that particular patient.”
For Dr. Elad, the bigger concern is accuracy. He said physicians are trained to spot AI output that does not add up, he said.
“I worry that patients may not be able to always discern when the advice they’re getting from AI is either inappropriate for their clinical situation, or worse, actually erroneous or a hallucination,” he said.
That is why he encourages patients to bring what they find to him. “Please bring it here so I can filter it for you,” he said he tells them. “I don’t know of any clinician here who is offended by a patient doing their own research and coming in.”
Often, he can confirm the chatbot’s answer. Other times, he pushes back. “In your particular situation, because I know more about your case than the AI knows about your case, this is why I’m doing this, or this is why I disagree with what ChatGPT told you,” he said. Sometimes the gap is practical, such as when AI recommends a treatment the patient’s insurance does not cover.
Mayo Clinic has turned those conversations into a formal framework. “The emphasis is not on dismissing AI-generated information, but on helping patients interpret it safely and appropriately,” Dr. Ebbert said.
The guide gives clinicians a four-step framework:
- Normalize it. Acknowledge that patients use these tools and that their AI output shows what worries them.
- Set expectations. Explain that AI may not know the patient’s history, exam findings, test results or preferences and can list rare or serious diagnoses without saying what is likely. The guide also reminds clinicians to tell patients not to enter identifiable health information into public AI tools.
- Guide the conversation. Do not debate whether the AI was right or wrong. Ask what concerns the patient most, “separate what is possible from what is probable” and explain the reasoning behind a test, referral or treatment.
- Agree on next steps. Close with what is most likely, what happens next, what to watch for and when to contact the care team again.
The effect on visit length is mixed for now. For Dr. Ebbert, some visits run faster because the patient is prepared. Others run longer as physicians correct misinformation or work through long lists of AI-generated possibilities.
Dr. Elad said AI-informed visits often do take longer, “but I think in a good way.”
“They’ve come in having thought about their condition or whatever their issue is, and they’ve formulated questions based on what they’ve learned,” he said. “You’re actually talking to an educated audience who’s at least familiar with some of the terminology you might use.”
Some patients are going a step further and recording their visits, essentially bringing their own ambient AI into the exam room. Dr. Ebbert called it a newer issue Mayo Clinic is starting to see and said clinicians should still address the patient’s underlying questions, but the AI-generated summary should not be treated as the authoritative medical record.
Dr. Elad said no patient has asked him to record a visit with their own AI tool yet. In California, all parties must consent to a recording, so patients would need physicians’ consent, just as physicians ask patients before using ambient AI.
He said he personally would not object, though physicians’ comfort levels vary. Physicians edit their own ambient AI notes before accepting them, he said, “and I can tell you that ambient AI certainly makes mistakes that we need to correct.”
“I would be careful if I was a patient using ambient AI to record a medical visit because it might misunderstand something or misconstrue something, and in that situation, you don’t have the benefit of your physician reviewing it,” he said.
https://www.beckershospitalreview.com/quality/hospital-physician-relationships/thats-what-chatgpt-told-me-how-ai-is-entering-the-exam-room/
