Modern medical AI can answer board-exam questions at or above the level of newly certified clinicians, and several large language models now produce empathetic, well-organized answers to health questions. That's real. But medicine is not a question-answering contest — it's a decision-making process under uncertainty, with a body attached. Knowing which task you're handing to a chatbot is the whole game.
Where AI Genuinely Shines
- Health literacy — explaining a diagnosis, a lab report, or a medication's mechanism in plain language, at 2 a.m., for free. This is arguably its best use.
- Visit preparation — organizing symptoms into a timeline, generating questions to ask, decoding jargon your doctor used in ten seconds you didn't get to clarify.
- Pattern surfacing for discussion — "these symptoms together are sometimes associated with X; mention it to your doctor" is a legitimate framing — as a prompt for a visit, not a verdict.
- Administrative navigation — insurance questions, appointment scripts, understanding consent forms and discharge instructions.
- Second opinion on reasoning — cross-checking whether an offered workup makes sense for your situation, to raise in consultation.
The Silent Failure Modes
AI failures in health contexts aren't usually dramatic hallucinations. They're quieter:
- Plausible wrongness — a fluent, confident differential that misses the actual diagnosis. Models optimize for likely answers; medicine's dangerous cases are often the unlikely ones.
- No physical exam — a huge share of diagnosis comes from touching, looking, listening, and smelling. A chatbot sees only what you type, and patients don't know which details matter.
- Missing the sick patient — mild-sounding descriptions can mask serious illness ("just tired and a bit short of breath"), and a model's reassurance can delay care.
- Stale or skewed knowledge — training cutoffs and population-skewed data mean guidance may lag current guidelines or fit you poorly.
- Privacy exposure — pasting detailed health data into general-purpose chat services may mean data you can't retrieve is used in ways you didn't intend.
Always See a Human When...
- Any red-flag symptom — chest pain, breathlessness at rest, one-sided weakness or facial droop, sudden severe headache, uncontrolled bleeding, high fever with confusion, sudden vision loss, thoughts of self-harm. Emergency services first, chatbot never.
- A new symptom that persists or worsens beyond a reasonable watchful-waiting window — a chatbot can't examine the lump, hear the murmur, or feel the mass.
- Medication decisions — starting, stopping, or dosing anything, especially with other drugs, pregnancy, kidney or liver disease in the picture. Interactions are exactly where generalized knowledge is riskiest.
- Children, pregnancy, older adults, chronic disease — populations where presentation is atypical and stakes are high.
- Mental health crises — AI can support mild, everyday struggles, but crisis lines and human clinicians are the right tool for emergencies.
A Practical Division of Labor
Think of it as a two-stage system. Stage one — orient: use AI to understand your body's vocabulary, structure what you're experiencing, and decide whether this is a "wait a week," a "book an appointment," or a "go now." Stage two — decide and treat: a licensed clinician, who can examine you, order tests, prescribe, and be accountable. AI is excellent at getting you to stage two better prepared. It is a poor substitute for stage two.
One rule of thumb: if you find yourself trying to get the AI to reassure you that you don't need to see a doctor, that itself is a signal to book the appointment.
The Bottom Line
The question isn't "AI or doctor" — it's "which task for which tool." Use AI for explanation, preparation, and navigation; use humans for examination, diagnosis, prescriptions, and anything with red flags. The people who get the best care in this decade will be the ones fluent in both — and honest about the boundary.