When a patient has already asked ChatGPT

You have just finished examining a patient. Thirty years of clinical training, a physical examination, and the patient’s history have led you to a diagnosis you are confident in. You explain it. There is a pause. Then: “But I looked it up on ChatGPT, and it said it could be something else.”

This moment is happening in consultation rooms across the country. Sometimes the patient shows you the screen. Sometimes they simply hold the information in reserve. Either way, something has changed. The patient sitting across from you now has a starting reference against which they are evaluating your diagnosis.

To see why this matters, consider how patients have traditionally evaluated doctors. In business, a medical consultation is classified as a “credence service,” one whose quality cannot be evaluated even after consumption because the consumer lacks the tools to judge. Your patient leaves still unable to answer the most important question: was that diagnosis right? Was that treatment plan the best available? Even recovery offers no clean verdict. They might have gotten better anyway. Without a way to assess the quality of care they received, patients fall back on what they can assess.

In a large-scale text analysis of outpatient physician reviews in India, higher-rated doctors were those whose reviews featured more frequent mentions of empathy and empowerment. Phrases such as “the doctor was caring, supportive, and understanding” consistently predicted better ratings. Patients rewarded empathy and empowerment far more reliably than technical indicators such as diagnostic ability, diagnostic accuracy, or recovery speed. For generations, patients had no reliable way to assess the quality of care they received, so how a doctor made them feel became the most available measure. It was the only instrument they had.

Until now. 

Large language models are beginning to change how your patients evaluate you. Before the appointment, they had a conversation with an AI, received a shortlist of probable diagnoses with varying likelihoods, and developed a working hypothesis about what was wrong with them. By the time they sit across from you, they are no longer waiting for a verdict. They are benchmarking. Does your diagnosis appear on their list? Does your reasoning follow the same symptom logic as the algorithm outlined?

A study makes the shift measurable. Patients who consulted an LLM before their appointment placed significantly greater weight on diagnostic accuracy and treatment quality when evaluating a doctor than patients who had not. A single interaction with an LLM appeared to change what patients considered important in evaluating a doctor. Your patient no longer just asks, “Was my doctor kind?” They are beginning to ask, “Was my doctor right?” Whether their answer is accurate is a separate question. What matters is that they now believe they can form one. And most will say nothing about it. Research consistently finds that patients are reluctant to disclose what they have read, not because they distrust it, but because they fear challenging your authority. The benchmarking may remain unspoken, but it can still shape how the patient interprets the consultation.

While you may know that AI-based self-diagnosis is often unreliable, symptoms entered into a chatbox are incomplete, non-specific, and untethered from clinical examination. Your patient does not arrive with that knowledge. They arrive with a concern shaped by what they read, and that concern is genuine. More importantly, it is the lens through which they are evaluating the entire interaction. Dismissing it without acknowledgement does not make it disappear. It simply drives it underground, where it quietly shapes how they interpret everything you say next.

This is precisely why the burden falls on the doctor to ask. A simple question like, “Have you already described your symptoms to an AI, and what did it tell you?” does more than surface hidden information. It signals an expert’s diagnostic discipline. It tells the patient that the doctor is secure enough to examine a competing explanation, thorough enough to gather every relevant input, and attentive enough to treat the patient’s concern as real. Research on advice-seeking and question-asking suggests that seeking an opinion or asking a non-specialist relevant questions can increase perceptions of competence when the questions signal information-gathering rather than ignorance. In that moment, asking is not a sign of uncertainty. It is a signal of clinical competence. It also signals empathy, because it shows the patient that their concern is being heard rather than dismissed.

Once competence and empathy are both established, you have the standing to do what the consultation actually requires. You can explain why the AI output may have been unreliable: how symptom-prompting shapes results, how the same condition described differently produces a different differential, where hallucination risk is high. This part of the conversation will take time. It may extend the visit. But a patient who understands why your diagnosis diverges from what they read is far more likely to follow the treatment plan, return for follow-ups, and not quietly abandon care the moment doubt resurfaces. 

Hospitals and clinics can make this easier by adding a simple pre-consultation tab or intake bot where patients can enter the symptoms they described to an LLM, the questions they still have, and the AI-generated possibilities they are worried about. When the doctor sees this before or during the consultation, the conversation can begin from the patient’s actual starting point, allowing concerns that might otherwise remain hidden to surface without forcing either side into an awkward exchange about LLM interactions.

The investment in that explanation is not a concession to the algorithm. It is the most clinical thing you can do.