You’ve had a twinge in your side for three days. It’s 10pm, the medical centre’s closed and your phone is right there. Five minutes later you’ve either talked yourself into something terrible or been told it’s nothing, and neither answer came from anyone who’s met you.
Plenty of us do it. OpenAI says more than 300 million people a week ask ChatGPT health-related questions. A University of Michigan poll of 2,883 Americans aged 50 and over, run in February 2025, found 14% had used AI for health information, and 46% had very little or no trust in AI-generated health information.
That caution looks well placed.
What the research says
In a randomised trial led by the University of Oxford and published in Nature Medicine in February 2026, 1,298 UK adults worked through ten medical scenarios written by doctors. Some had help from an AI chatbot (GPT-4o, Llama 3 or Command R+). The rest used whatever they’d normally use, such as a web search or their own judgement.
Tested on their own, the chatbots identified the relevant conditions in 94.9% of cases. People using those same chatbots identified a relevant condition in fewer than 34.5% of cases, and chose the right course of action, such as seeing a GP or going to the emergency department, in fewer than 44.2%. That was no better than the group without AI.
The weak point was the conversation. The researchers found people often gave the chatbot only part of the story, and the chatbot sometimes suggested the right answer without getting it across. Small changes in wording also produced very different answers.
“Patients need to be aware that asking a large language model about their symptoms can be dangerous, giving wrong diagnoses and failing to recognise when urgent help is needed,” says Dr Rebecca Payne, a GP and the study’s lead medical practitioner at the University of Oxford.
A second Nature Medicine study, from Mount Sinai in New York, tested ChatGPT Health, OpenAI’s health feature, on 60 scenarios written by clinicians. Of the cases that were genuine emergencies, it undertriaged 52%, sending some people with conditions such as diabetic ketoacidosis to be seen within 24 to 48 hours instead of to the emergency department. It handled classic emergencies such as stroke and anaphylaxis correctly.
In borderline cases where family or friends played the symptoms down, its advice shifted, mostly towards less urgent care. It’s worth remembering next time someone tells you it’s probably nothing.
Old-school Dr Google leans the other way. A 2015 Harvard audit of 23 online symptom checkers, published in the BMJ, found they put the correct diagnosis first in 34% of cases. Their triage advice was generally risk averse, sending people to seek care when looking after themselves at home was reasonable.
That can feed worry. Australian researchers who pooled 20 studies with 7,373 participants found health anxiety tends to go with cyberchondria, which they define as excessive or repeated searching for health information that leaves people distressed or anxious. It’s a correlation, so it can’t say which comes first.
How to use it well
Healthify He Puna Waiora, the national health information site, says AI can be a useful tool but isn’t a replacement for professional medical advice. Its advice boils down to a few habits.
Use it for words, not verdicts. Healthify says AI can explain medical terms, give general information about conditions, help organise your symptoms into a timeline and help you prepare questions for an appointment. It can’t safely or accurately diagnose a medical condition.
Ask open questions. Healthify warns that AI can agree with your assumptions instead of challenging them, which it calls AI sycophancy.
“Could this be a kidney stone?” invites a yes. “What are the possible causes of pain in my side?” is more likely to get a balanced answer.
Remember it hasn’t met you. AI doesn’t know your full medical history and can’t examine you or run tests, Healthify says, and it may sound confident even when it’s wrong. Information may also not work equally well for different ages and health situations.
Watch for red flags. Healthify says to be wary of anything that promises quick or guaranteed cures, encourages you to stop prescribed treatment or creates fear or pressure.
Think before you share. Health New Zealand’s guidance for the health sector says the provider usually keeps whatever is typed into publicly available AI tools, and may use it to train future models. In ChatGPT, OpenAI says you can change whether your chats are used for training in Settings, under Data Controls.
When to put the phone down
Chest pain, severe breathing difficulties, signs of stroke and thoughts of self-harm need urgent medical care, Healthify says. In an emergency, call 111.
For everything in between, there’s Healthline. It’s free on 0800 611 116, 24 hours a day, and you talk to nurses and paramedics.
Health New Zealand says to call if you’re worried or unsure, can’t get to a GP or need advice about your medicine, and there’s no need to wait until things get very bad. You can ask for a Māori clinician between 8am and 8pm, and interpreters are available for up to 150 languages.
If a chatbot has turned up something that worries you, Healthify’s advice is to discuss it with your healthcare provider. Bring the timeline it helped you write. That twinge will be much easier to explain.
