Iasora Blog PT

How to ask AI about your child's health (and get safer answers)

Updated 11 October 2026 · 7 min read

At night, a father on the couch types on his phone while his child sleeps beside him; a thermometer and a cup sit on the table, next to a chat with questions and answers.

In short: an AI chatbot answers better when you give it hard facts: age, weight, temperature measured with a thermometer, how long it has been going on, what you have already given and when, other symptoms, and whether vaccinations are up to date. Ask it first "what would make this urgent?" and ask it to list what it doesn't know. Use it to understand and to prepare questions for the clinic — never to work out a medicine dose. If your child has a danger sign, don't ask the AI: go to the clinic or emergency unit.

Before you open the chat: if your child can't drink or breastfeed, vomits everything, has had a fit, is floppy or hard to wake, or is struggling to breathe, skip the chatbot. Check the danger signs and go straight to the nearest casualty or emergency unit. For an ambulance, call 10177, or 112 from a cellphone.

Why the way you ask changes the answer

A chatbot — ChatGPT, Gemini, Claude, Meta AI on WhatsApp — can't see your child. It only knows what you type. A vague question gets a generic answer, or worse, one where the gaps are filled with guesses.

A University of Oxford study published in Nature Medicine put this to the test with 1,298 people. On their own, the AI models identified the condition in most scenarios. But when ordinary people chatted with them, they got the condition right in under 35% of cases and chose the right next step (stay home, see a GP, go to casualty) in under 45% — no better than people who searched the usual way. Part of the problem: the answers mixed good and bad advice, and shifted a lot with small changes in wording.

So the quality of your question matters — and even a great question doesn't replace a clinician examining your child.

What to tell the AI: the fact list

Gather these before you type. A lot of it is in your child's Road to Health Book.

  • Exact age — in months if under 2. A 2-month-old with fever is very different from a 4-year-old.
  • Weight — the last one recorded, and when.
  • How long — when each symptom started ("fever since 4 pm yesterday", not "a few days").
  • Measured temperature — the thermometer reading and where you took it (armpit, mouth, ear). "Feels hot" is not enough.
  • What you've already given, and when — the medicine's name and the time of the last dose. You're reporting what happened, not asking for a dose.
  • Other symptoms — cough, diarrhoea (how many times), vomiting, rash, eating and drinking, wet nappies, active or floppy.
  • Vaccinations — up to date on the EPI-SA schedule or not.
  • Where you live and recent travel — for example a trip to the Lowveld, northern KwaZulu-Natal, Limpopo or Mozambique, and whether anyone at home has a long-standing cough or TB.

Location matters because much of the data these models learn from comes from wealthy countries where malaria barely exists and TB is rare. If you don't mention a trip to a malaria area, the AI may never think of malaria in a feverish child.

Example: a bad question and a good one

Bad question:

"My son has a fever, what should I give him?"

The AI doesn't know his age, how long, how high, or where you've been. And it asks for a medicine — exactly what you shouldn't ask.

Good question:

"My daughter is 14 months old and weighed 9.5 kg at the clinic last month. We live in Durban and came back from a week near Kosi Bay five days ago.

She's had a fever since 4 pm yesterday. At 9 pm I measured 38.6 °C under the arm.

I gave the paracetamol syrup the clinic prescribed last time, at 9.15 pm.

She's feeding less but still breastfeeding, had a wet nappy this morning, runny nose, no diarrhoea or vomiting, no rash. Vaccinations are up to date.

First: what signs would make this urgent? Then: what could explain this, and what information are you missing? Don't give me medicine doses."

The second question gives facts, not impressions, and tells the AI what you want — and what you don't.

Ask first: "what would make this urgent?"

This is the single most useful question. Instead of starting with "what is it", start with "when can this no longer wait". Compare the answer with the WHO danger signs clinicians use (IMCI), summarised on our danger signs page. If your child already has one, stop chatting and go.

Ask it to list what it doesn't know

Chatbots write with great confidence, even when they're wrong. The WHO warns that these models can produce false, inaccurate or incomplete statements, and that people tend to over-trust what the machine says ("automation bias"). Push back with questions like:

  • "What information are you missing to answer this better?"
  • "What else could it be, including less likely but serious causes?"
  • "What can only a nurse or doctor check by examining her?"
  • "Which parts of this answer are you unsure about?"

If the AI never shows any doubt, trust it less, not more.

Never use AI to choose a dose

This is the hard line. The right dose for a child depends on weight, age, the product's strength (syrups, drops and tablets all differ) and other conditions. A small slip in the AI's text, or in how you read it, can mean a dangerous dose.

  • Use the dose the clinic prescribed or the one on the package insert.
  • If unsure, ask the pharmacist or the clinic sister.
  • If the AI gives you a dose in mg or ml without being asked, ignore it.
  • Never give aspirin to children, whatever the AI says.

The only exception you'll find in official guidance is oral rehydration solution (ORS) and zinc for diarrhoea — and even then, follow what the clinic tells you. See diarrhoea and dehydration.

Double-check with an official source

Treat the AI's answer as a first draft. Before you act, check it against:

  • the nurse or doctor at your clinic, or your GP;
  • the Road to Health Book and what you were told at the last visit;
  • official sources: the National Department of Health, NICD, WHO, UNICEF.

If the AI and the clinician disagree, go with the clinician — they've seen your child.

Use AI to prepare for the clinic

This is where AI helps most: organising what you'll say and what you want to ask. In a long clinic queue with a sick toddler on your lap, things get forgotten. Try: "Based on these symptoms, what should I ask the nurse?" or "Help me write a 5-line summary of what's happened since yesterday." More ideas in prepare for a clinic visit with AI.

A template to copy

Copy it, fill in what you know and delete what doesn't apply:

Child: [age in months or years], [weight] kg (weighed on [date]).

Where we live: [town/province]. Recent travel: [where, when].

Main problem: [symptom] since [day and time].

Measured temperature: [reading] °C at [time], [armpit/mouth/ear].

Already given: [medicine name] at [time], prescribed by [who].

Other symptoms: [cough, diarrhoea (how many times), vomiting, rash, breathing].

Eating/drinking/feeding: [normal/less/nothing]. Last wet nappy: [time].

Child is: [active/sleepy/hard to wake].

Vaccinations up to date: [yes/no/not sure]. Known conditions: [...].

Request: 1) What signs would make this urgent? 2) What could explain this? 3) What information are you missing? 4) What should I ask at the clinic? Do not give me medicine doses.

Be careful what you share

In a general chatbot, leave out your child's full name, your address and ID numbers — the answer doesn't need them. Iasora has its own AI, HealthAI, which answers questions about that one child's record: it only works with your consent for each child, which you can withdraw; before the record goes to the AI provider (Anthropic or OpenAI), the child's name becomes "the patient" and yours "the guardian", and phone numbers, emails, the date of birth and the health card number never leave — but it isn't anonymous, because the dates things happened and the rest of the record go as they are. It tells you which fact is missing, tells you to seek care when something sounds urgent, and is not a medical device and doesn't diagnose. More in your child's health data and AI.

The short version

  1. Danger sign? Don't ask — go.
  2. Give facts: age, weight, measured temperature, times, what you've given, vaccinations, where you've been.
  3. Ask first what would make it urgent.
  4. Ask it to say what it doesn't know.
  5. Never use AI for doses.
  6. Check with a clinician and official sources.
  7. Use AI to prepare your questions for the clinic.

If you're unsure about any of it, that's an answer too: take your child to the clinic.

Sources

  1. WHO releases AI ethics and governance guidance for large multi-modal models — WHO
  2. Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models — WHO
  3. Reliability of LLMs as medical assistants for the general public: a randomized preregistered study (Bean et al., Nature Medicine, 2026) — University of Oxford
  4. Integrated Management of Childhood Illness (IMCI) — modules and Chart Booklet — WHO
  5. Know who you can call in an emergency — Western Cape Government

This article is general information, written by the Iasora team from the sources above. It does not replace a consultation or the judgement of a health worker who sees your child. If in doubt, go to your nearest clinic.