What AI Can and Can’t Tell You About Your Health | Clinician’s View

 

Something has changed in the consultation room. And I think it is worth talking about it honestly, not with alarm, and not with uncritical enthusiasm, but from the place where both the benefits and the real limitations of artificial intelligence in healthcare are most visible: the clinical encounter itself.

FREE DOWNLOAD

In this article, I want to share what I actually see in my practice: three kinds of patient, a clear account of what AI can and cannot do, and five things I would tell anyone who is now using AI as part of how they navigate their health. Which, in 2026, is most of us.

How We Got Here: From Google to AI

The relationship between patients and medical information has changed more in the past thirty years than in the preceding three centuries. For most of recorded history, the doctor held the information. The patient experienced symptoms; the doctor named, explained, and treated them. Then came the internet, and then Google, and patients began arriving at appointments with printouts, symptom lists, and self-diagnoses drawn from forum threads.

Clinicians, myself included, at the time, found this frustrating. The phrase “Dr. Google” became shorthand for a phenomenon that felt like an obstacle. But with hindsight, that frustration was never entirely fair. Patients were doing exactly what made sense: trying to understand what was happening in their own bodies. The impulse was right. What was imperfect was the tool.

And now the tool has changed again. AI assistants provide medical information not as a list of links but as a structured, conversational response tailored to the specific question. The quality has improved dramatically. And critically, AI answers back. It responds to follow-up questions. For most patients, it feels less like a search and more like a consultation.

I use AI myself: for literature reviews, summarising research, maybe thinking through complex clinical questions. So when a patient arrives having already spoken to ChatGPT or Claude about their symptoms, I understand that entirely. The question was never whether patients should be informed. The question is what we do with that information when they arrive.

What I Actually See in Clinic

Over the past two years, I have observed a consistent pattern, one I hear described by colleagues across specialties. It is clearest when I describe three types of patient encounters.

The Patient Who Arrived Prepared

The first kind of patient consults AI before the appointment. They arrive with a working hypothesis, sometimes correct, sometimes partially so, and want the consultation to confirm or revise it. This patient is often easier to work with. The consultation begins at a higher level. Questions are sharper. Time is used better.

But there is a subtle difficulty. The patient who arrives with a conclusion is, by human psychology, somewhat less open to a different answer than the patient who arrives with no prior framework. Certainty in medicine is the thing you have to be most careful with. The clinician’s job in that room is both clinical and conversational, reaching the right diagnosis while genuinely shifting what the patient believes.

The Patient Who Uploaded Their Results

The second kind of patient receives test results through an online patient portal, often before their doctor has had a chance to call. They upload those results to an AI assistant, receive an interpretation, and attend the follow-up appointment already knowing, or believing they know, what the results mean.

I have seen this more and more in the past two years. And the AI interpretation is often not wrong. Reference ranges are correctly identified. Abnormal values are flagged. Plain-language explanations are frequently accurate.

“The AI read the document. It did not read the patient.”

But the interpretation lacks context. It does not know about the clinical finding noted six months ago that has not yet been acted upon. It does not know the patient mentioned something almost in passing, “actually, there’s been this one other thing”, that changes everything. The AI knew what was in the document. It did not know the patient.

The Patient Who Was Reassured Too Soon

The third kind of patient worries me most. This patient had a symptom, asked AI about it, and received a probabilistically reasonable response: “This is most likely benign,” “This probably doesn’t need urgent attention.” And so they waited. Sometimes three weeks. Sometimes longer.

The AI was not lying. It was giving an honest statistical answer. For most people with that symptom, that answer is correct. But medicine is not ultimately a statistical enterprise, it is a clinical one. The question is not what is probably true for most people with these symptoms. The question is what is true for this person, in this moment, with this history.

What AI Does Well and What It Cannot Do

This is not an anti-AI argument. AI tools offer genuine benefits to patients, and precision requires acknowledging both sides.

What AI Does Well

What AI Cannot Do: The Structural Limitations

The limitations of AI in clinical medicine are not bugs in the current technology. They are structural, they derive from what AI is.

Limitation	Why it matters clinically
Cannot examine you	A substantial proportion of clinical diagnoses depends on observation: the way you hold your head when describing dizziness, the asymmetry noticed before you finish speaking, the finding that wasn’t in the referral letter.
Cannot ask the right follow-up question	Clinicians are trained to probe, to sense when the stated complaint is not the whole story, to redirect, to pursue the thread the patient dropped without realising. Research confirms AI chatbots rarely do this.
Cannot build a proper differential diagnosis	A 2025 Harvard/Mass General study tested 21 AI models specifically on replicating a clinician’s differential diagnosis process — the disciplined list of what cannot be missed. The failure rate was 80%. [3]
Cannot carry your history	Each AI conversation begins fresh. It has no access to findings from previous appointments, patterns observed over time, or the relationship that makes you willing to mention the thing you’ve been reluctant to raise. Your doctor holds this.

The Question Underneath: Information vs. Understanding

There is a philosophical dimension to this shift that clinical analysis alone does not capture.

There is a difference between information and understanding. AI gives information, often accurate, sometimes excellent. But understanding is something different. Understanding requires context, relationship, time, and the willingness to remain uncertain rather than resolve ambiguity too quickly.

When a patient arrives having already been answered, even a good answer, something changes before the consultation begins. The space for not-knowing has been filled. The questions they might have asked, if they hadn’t already been told, are less accessible. The possibility that what they came in for isn’t actually the most important thing to discuss has been foreclosed.

I have sat with patients in the moment of receiving difficult news more times than I can count. What I have learned is not what to say. It is how to be present. How to let a silence have its full weight. How to notice when a patient is not actually asking what they appear to be asking. That is not information. It is something that develops through years of being in the room with people at their most vulnerable time. It cannot be trained on text. It cannot be prompted.

The patients who use AI best, I’ve noticed, are not the ones who arrive with answers. They are the ones who arrive with better questions. AI as preparation. Not as a conclusion.

“You came here to be seen, not just read. And there is a difference.”

Five Things to Do Differently from Today

If you are using AI for health information, as most people do, here is how to use it in a way that actually serves your health rather than creating new risks.

Frequently Asked Questions

Is it safe to use AI like ChatGPT for health advice?

AI tools can be genuinely useful for understanding your symptoms, translating medical language, and preparing questions before a doctor’s appointment. However, they cannot examine you, do not hold your medical history, and have significant limitations when it comes to the differential diagnosis process that underpins safe clinical reasoning. Using AI to prepare for a consultation is sensible. Using it as a replacement for clinical assessment carries real risks, particularly if it reassures you about a symptom that warrants investigation.

What can AI not do that a doctor can?

AI cannot physically examine a patient, which means it misses the observational data that always shapes a diagnosis. It cannot ask the follow-up question that changes the picture. And it cannot build a differential diagnosis, the disciplined listing of what cannot be missed, with clinical reliability. A 2025 Harvard Medical School and Mass General Brigham study found an 80% failure rate when AI was tested specifically on replicating a clinician’s differential diagnosis process. These are not minor gaps.

Should I tell my doctor I looked up my symptoms on AI?

Yes, and bring what you found into the conversation. The most productive consultations I conduct often begin with a patient saying, “The AI suggested this, can we talk about that?” Your doctor can confirm what the AI got right, correct what it missed, and add the clinical context that only an examination and a shared history can provide. Arriving with a hidden AI diagnosis and testing it silently helps neither of you.

Why do doctors use AI if they say patients shouldn’t rely on it?

Clinicians use AI as a tool within a workflow that includes examination, history-taking, and professional accountability, not as a standalone decision-maker. The concern is not AI use in itself but AI use without a clinical context. When a patient uses AI without examination findings, without continuity of care, and without the clinical training to interpret what the AI produces, the information carries a different and higher risk of misapplication.

How is the doctor-patient relationship changing because of AI?

The relationship is shifting from a model in which the doctor held all the information to one in which patients often arrive with a prior interpretation already formed. This can be positive, with better-prepared patients, more targeted questions, and more collaborative consultations. But it also introduces a new dynamic: patients who arrive certain rather than curious. The doctors who navigate this best are those who treat the patient’s AI research as the starting point of a conversation, not a challenge to be overridden. The technology is changing the shape of the consultation. The need for clinical presence, examination, and continuity of care is not.

References

1. West Health–Gallup. AI in Health Survey. West Health Institute; April 2026. Available at: westhealth.org
2. American Medical Association. Physician Adoption of Digital Health Technologies. AMA Survey; 2024. Available at: ama-assn.org
3. Succi MD, Rao VN, et al. Differential diagnosis generation by large language models: a standardised clinical vignette study. Mass General Brigham / Harvard Medical School; 2025.
4. Zocdoc. What Patients Want Report. Zocdoc; December 2025. Available at: zocdoc.com