ChatGPT Has Become a New Third Voice in the Exam Room
A patient with dizziness came in expecting a familiar diagnostic conversation. Jon Ebbert, MD, took the history, examined her and suspected benign paroxysmal positional vertigo. She cut him off with a line that has already become emblematic of the new visit dynamic: “That’s what ChatGPT told me.”
The exchange is brief, but the clinical meaning is large. Patients are no longer arriving only with symptoms and anxiety; they are increasingly arriving with a machine-generated diagnosis, a list of treatment options and a set of questions assembled by a chatbot. In primary care, cardiology and other specialties, that changes the first few minutes of the encounter and often reshapes the entire visit.
Dr. Ebbert, a general internist and medical director of the Systems Innovation Accelerator at Mayo Clinic in Rochester, Minn., said the key shift is not the presence of questions. It is the confidence with which AI presents them. ChatGPT, Claude and Gemini can sound polished, direct and personalized, which makes their output feel more authoritative than a standard search result even when the underlying answer is incomplete, detached from the patient’s history or simply wrong.
One sentence captures the bedside consequence: patients now walk into clinic with algorithmic certainty.
Why AI Feels More Authoritative Than Search
For years, physicians have dealt with “Dr. Google.” The new behavior is different because large language models do not just return links; they synthesize, explain and often frame the answer in a conversational tone that sounds tailored to the user’s concern. Yaron Elad, MD, chief medical informatics officer at Cedars-Sinai and a practicing cardiologist, said patients appear to view AI as more authoritative than they ever viewed search engines.
That impression is not irrational. Chatbots often draw from stronger sources and cite them, while older search patterns could send patients to unreliable or commercially driven websites. Still, the polished delivery can hide a major limitation: the model does not know the patient’s exam, lab results, imaging, medication list, insurance constraints or personal priorities unless those details are provided. Even then, the response may still be incomplete or clinically off target.
Dr. Ebbert said Mayo Clinic tends to see AI output in three forms. Some patients bring a differential diagnosis based on symptoms. Others arrive with an explanation of a known condition. A third group comes with treatment options or with questions the chatbot suggested they ask. Any of those can be useful starting points. Any of them can also send the visit spiraling into irrelevant territory if the patient and clinician spend too long debating possibilities that do not fit the actual presentation.
What Clinicians Are Doing With AI-Generated Questions
The productive path is to treat the chatbot output as a prompt, not a verdict. Dr. Elad said he does not hesitate to ask patients where an idea came from. Often, the answer is immediate and disarming: “Oh, you know, I’m not questioning you, but this is what ChatGPT said.”
That opening gives the physician a chance to do what AI cannot do well: filter the answer through the specific case. In practice, that means confirming the diagnosis when it fits, explaining why it fits and spelling out what management looks like in this individual patient. In other visits, it means pushing back. A model may suggest a treatment that does not align with the clinical picture, is not covered by insurance or is simply not the right step based on earlier testing already completed.
The value of the clinician’s role becomes most obvious in complicated illness. An AI system can generate an endless list of plausible questions without knowing which ones were already answered two visits ago or which branch of the workup has already been excluded. The result is not just more information. It is often more noise. Physicians can tell at a glance when a suggestion does not add up; patients usually cannot.
Mayo Clinic’s Four-Step Framework for Safer AI Conversations
Mayo Clinic has turned those encounters into a formal approach for clinicians. The framework is built around safety, interpretation and shared decision-making rather than confrontation.
First, clinicians normalize AI use. Patients should know that using these tools is common and that the output often reflects what is worrying them most. Second, clinicians set expectations. AI does not know the patient’s full story, physical exam findings, test results or preferences, and it may list rare or severe diagnoses without indicating what is actually likely. The guidance also reminds patients not to enter identifiable health information into public AI tools.
Third, clinicians guide the conversation. The goal is not to win a debate with the chatbot. It is to ask what concerns the patient most, separate what is possible from what is probable and explain the reasoning behind a test, referral or treatment. Fourth, the visit should end with clear next steps: what is most likely, what happens next, what warning signs matter and when the patient should contact the care team again.
In bedside terms, the framework does two things at once. It protects against misinformation and preserves the patient’s sense that the physician is taking the concern seriously.
Visit Length, Trust and the New Educational Dynamic
The effect on visit length is uneven. Some encounters move faster because patients have already thought through their symptoms and arrived prepared. Others take longer because the clinician has to correct inaccurate details or unwind a long list of machine-generated possibilities. Dr. Elad said the longer visits can still be beneficial because they often reflect a more engaged patient who has already learned enough terminology to participate meaningfully.
That shift changes the tone of the room. A patient who has read about a condition before the appointment may ask sharper questions, challenge assumptions more directly or seek a more granular explanation of why a test is needed. For clinicians, that can feel like extra work. For the encounter, it can also mean deeper understanding if the physician remains anchored to the actual case rather than the chatbot’s framing.
Patients are also beginning to bring a second layer of AI into the visit by recording the conversation and using ambient tools to generate summaries. Dr. Ebbert said Mayo is starting to see that behavior. The caution is straightforward: an AI-generated summary should not be treated as the medical record. It can misunderstand, omit context or misconstrue the exchange in ways that a physician would normally catch before anything is finalized.
Ambient AI Adds Convenience and Another Layer of Error Risk
Dr. Elad said he has not yet had a patient ask to record a visit with their own AI tool, but he would not object in principle. Legal consent still matters, and in California all parties must agree to a recording. He also emphasized a point that clinicians already know from their own use of ambient AI documentation: the technology makes mistakes that require human correction.
That matters because the patient side lacks the safety net of physician review. If a recording tool mishears a symptom, flattens nuance or turns a cautious plan into an overly definitive summary, the patient may leave with a distorted impression of what was actually said. The risk is not hypothetical. It is the same reason clinicians edit ambient AI notes before accepting them into the chart.
The broader lesson is not that patients should stop using AI. It is that clinicians now need a repeatable way to respond when AI walks into the exam room with the patient. The right posture is neither dismissal nor deference. It is disciplined interpretation. Patients will keep bringing chatbot output, and many of those questions will be useful. The clinical challenge is making sure the answer is grounded in the individual patient, not in a model’s most fluent guess.