A patient arrives with a diagnosis from ChatGPT. How is this different from “Dr Google”?

A search engine shows results; a chatbot says “in your case”

A patient sits down in the consulting room and says: “ChatGPT says I have an autoimmune disease. Do you think so too?” Before the appointment, she entered a few symptoms and one lab result into the chat. She received a coherent answer listing possible causes, then kept asking follow-up questions until one of the hypotheses began to dominate. A traditional search engine led to a range of pages that did not always agree with one another. The chatbot combined the information into a single answer that feels personal. Today’s search engines also use AI, so the technological boundary is no longer sharp. What matters is that the patient receives a single, coherent-sounding story about their health. A doctor who opens with “Please don’t read that rubbish” not only dismisses the hypothesis but may also belittle the patient’s effort to find answers. A better opening is: “Could you show me which symptoms you described and what exactly the model replied?” That way you know which of the patient’s beliefs needs to be clarified.

A coherent-sounding answer is no substitute for history-taking and examination

A language model can come up with reasonable hypotheses, but it does not know what the patient has not told it: how long the symptoms have lasted, comorbidities, current medication, physical examination findings or details of previous investigations. It may also weigh probabilities wrongly, especially when the description is incomplete or the question suggests the answer. Research on AI in medicine has to be interpreted with the way the tool was used in mind. A randomised trial published in 2026 assessed how doctors worked with a specific AI system on diagnostic tasks. The results of this study do not mean that a patient using a public chatbot will receive an equally reliable diagnosis. A doctor with the full clinical context works one way, an algorithm assessing a task description another, and a person worried about a single symptom yet another. So there is no need to prove to the patient that AI is “always wrong”. It is enough to show which questions remain open and which tests can help answer them.

First check what the patient actually heard from the model

“ChatGPT diagnosed me” is often shorthand. The model may have written “one possibility is…”, and after several conversations the patient remembered it as a certainty. Ask: “May I see the answer? Which part worried you most?” Then check what information the patient gave the chatbot. Perhaps there was no mention of a recent infection, a change of medication or an earlier test. Don’t demand access to their entire history of private conversations; the material relating to the medical question is enough. This lets you say: “This answer took three symptoms into account, but it didn’t have your other results, so the condition it mentions is just one of several hypotheses.” The patient hears a reasoned argument rather than a contest of authority. If the text contains good questions, you can make use of them. This approach shows that the doctor is not afraid of the tool, but is putting its use in its proper place.

Uncertainty can be explained without false reassurance

You cannot always say straight away: “It’s definitely not that disease.” The doctor may need an examination, observation or more information. A patient who has been shown an alarming suggested diagnosis often perceives uncertainty as a lack of competence. It is worth naming it directly: “Based on what we know today, this possibility hasn’t been confirmed. We’ll now look at the symptoms that support it or argue against it, and agree on a plan.” It also helps to separate three statements: what we know, what we don’t know and what we’ll do about what we don’t know. This is different from listing ten differential diagnoses without explaining what they mean. Don’t use overly categorical assurances just to calm a stressed person down quickly. If a particular symptom or result requires urgent action, apply the appropriate clinical assessment and procedure; the way you frame the conversation does not change how urgently you need to act.

How to respond to “The AI says something different from you”

The hardest moment comes when the patient compares the two sources as if they were two equal voices in a debate. “Please trust your doctor” may not be enough at this point. It is better to show the basis for the decision: “The model doesn’t know the findings of your physical examination, or that your symptoms started after your treatment was changed. That affects how we assess the possible causes.” Then check that the patient understands the difference: “Is my explanation clear? What are you still unsure about?” If they have brought a publication or a good question, take it seriously. If they want a second opinion, you can discuss the appropriate referral route. The goal is not to win against the chatbot, but to establish together what data justify the decision and what still needs to be clarified. The doctor can acknowledge the patient’s engagement without uncritically accepting generated hypotheses.

How to close the appointment so the patient leaves with a clear plan for what happens next

At the end, the patient needs a concrete plan more than a general “try not to worry”. Say what conclusions have come out of the appointment, which tests or consultations are needed, who will pass on the results and by when the patient should get in touch. Ask what the patient will do if the symptoms change or if signs appear that call for a quicker assessment. A brief check of understanding is useful: “Just to make sure I’ve explained everything clearly, how would you sum up the next step?” This does not sound like a test when the doctor takes responsibility for the clarity of the message. It is also worth saying what questions the patient can bring to the next appointment. The patient does not have to stop using technology. But they should leave with a clear distinction: a chatbot can help them put their questions into words, while diagnosis and decisions about their health require the right clinical context.

A patient with an AI answer usually needs two things: respect for their own search for answers and a clear explanation of what that answer still leaves unknown. The doctor can treat the generated hypothesis as material for the conversation, check the missing information and plan further assessment. Trust grows when uncertainty is named and the patient understands what will happen after the appointment.

How Em helps you prepare for a conversation with a patient convinced by an AI diagnosis

A doctor is about to see a patient who is coming in with a screenshot of a chatbot’s answer and says that doctors have ignored him in the past. In Em, she describes the situation without the patient’s name, test results or any other identifying details. She asks to rehearse the first two minutes of the appointment. Em helps her find an opening: “I can see you’ve done a lot of reading and want to understand your symptoms. Could you show me what information you gave the model?” She then practises how to respond to the challenge “The AI says it’s a definite diagnosis”, as well as how to name uncertainty calmly. She can prepare questions about the patient’s biggest worries and a short close with a plan for next steps. Em supports her choice of language, the structure of the conversation and how to respond to emotions; Em does not interpret medical records or make a diagnosis. With this kind of preparation, the doctor can focus the appointment on the person and the clinically relevant facts, instead of arguing about how reliable the chatbot is.