The Health Questions AI Actually Answers Well

Once you know what AI is bad at in health, the question is what is left — and quite a lot is, when the question is well-formed. AI is genuinely good at a set of tasks that were previously either paid or slow: understanding your own lab reports, translating jargon, and preparing for consultations. Used deliberately, these save real time and often produce a better conversation with your doctor.

The categories AI handles well

• Question type: Terminology · Example: What does 'creatinine' measure? · Why AI helps: Well-defined; general-knowledge answer
• Question type: Lab report field explanation · Example: What does an ALT of 62 mean? · Why AI helps: Reference ranges are standardised
• Question type: Medicine information (general) · Example: What is amlodipine used for? · Why AI helps: Drug info is well-documented
• Question type: Preparing for a specialist visit · Example: Questions to ask my cardiologist about a new statin · Why AI helps: General framework is stable
• Question type: Explaining a diagnosis · Example: What is atrial fibrillation, in plain language? · Why AI helps: Educational content is widely available
• Question type: Understanding treatment options · Example: General approaches to knee arthritis · Why AI helps: Overview is possible, specifics are for the doctor
• Question type: Lifestyle guidance framework · Example: General principles of a diabetic diet · Why AI helps: Established, not personalised
• Question type: Interpretation of imaging jargon · Example: What does a doctor's note of 'mild L4-L5 disc bulge' mean? · Why AI helps: Radiology vocabulary is well-defined

Specific examples of good questions

Formulated in a way that plays to AI's strengths:

• 'My doctor said my kidney function is stage 2 CKD. What does that mean in general terms?'
• 'Explain what an HbA1c of 7.2 tells me about my average blood sugar over the past few months.'
• 'What kinds of questions should I ask when I meet a nephrologist for the first time?'
• 'What are the general classes of blood pressure medicines, and how do they differ in mechanism?'
• 'What does a normal ejection fraction look like on an echocardiogram report, and what does 45% suggest?'

Specific examples of bad questions

Formulated in a way AI cannot answer usefully:

• 'I have a headache — what do I have?' (Diagnosis, not explanation.)
• 'My grandmother is taking these five medicines and has confusion — what should we do?' (Requires clinical judgement and specific interaction knowledge.)
• 'What dose of insulin should I take today?' (Personal dosing decision.)
• 'Is this rash a sign of cancer?' (Requires examination.)
• 'Can I stop this medicine?' (Depends on why it was started, current control, other conditions.)

How to write a question that gets a useful answer

Four properties of a well-formed question:

• Specific. 'HbA1c 7.2, want to understand what that means' is better than 'my diabetes'.
• Educational rather than decisional. 'What does this mean?' is better than 'what should I do?'.
• Includes context. Age, sex, general health background helps AI tailor the answer.
• Ends with 'in plain language' or 'for a layperson' if you want to skip the jargon.

The three questions to ask AI before every doctor visit

A specific pre-visit routine that consistently produces better appointments:

• 'I am going to see a [type of doctor] for [condition]. What questions should I be prepared to answer?'
• 'What questions should I ask them, given [my situation]?'
• 'What test results would be most useful to have with me?'

Fifteen minutes with AI before a consultation often turns a rushed six-minute visit into a productive one. This is where the tool earns its keep.

References

• JAMA — LLMs for patient education
• NEJM AI — patient-facing use of language models
• PMC — AI in patient education and health literacy