It is 2 a.m. Your chest feels odd, or your child has a fever, or a rash showed up and it looks worse than it did at dinner. The doctor’s office is closed. Your phone is right there.
So you type your symptoms into an AI chatbot. You are far from alone. A KFF poll fielded in February and March 2026 found that 32% of US adults had used an AI chatbot for health information in the past year. Among them, 65% said they wanted quick, immediate information, and 41% wanted to look something up before deciding whether to see a provider.
That is a sensible instinct. The question is how much to trust the answer. This post is general information, not medical advice.
Rule Number One: Red Flags Come Before Any App
Before anything else, know when not to type at all.
Call 911 (in the US) or go to the nearest emergency room if you or someone with you has any of these. This is a common-sense list and not complete.
- Chest pain or pressure
- Trouble breathing
- Signs of a stroke: face drooping, arm weakness, trouble speaking, sudden confusion
- A sudden, severe headache unlike any you have had
- Heavy bleeding that will not stop
- Fainting, or a seizure
- Thoughts of harming yourself. In the US, call or text 988.
If something feels like an emergency to you, trust that feeling. Do not let a chatbot talk you out of it, and do not wait for a chatbot to talk you into it.
What the Studies Actually Found
Chatbots can fail at the most important job: urgency. A Mount Sinai study, reported in early 2026, found that ChatGPT under-triaged about 52% of emergency cases, meaning it rated them as less serious than they were. The weak spot was time-sensitive cases, where the bot often misjudged how long a person could safely wait. OpenAI responded that the study did not reflect how people actually use ChatGPT and that the version tested was older and has since been improved.
People struggle to use them well. An Oxford study published in Nature Medicine gave people medical scenarios to work through with chatbots. Afterward, participants named the correct condition about a third of the time and chose the correct next step, such as going to the ER or staying home, 43% of the time. Researcher Andrew Bean said people “don’t know what they are supposed to be telling the model.”
Wording changes everything. In the Oxford work, two people described the same situation differently. The one who wrote “the worst headache I’ve ever had” was told to go to the ER. The other was told to take aspirin and stay home, and that second condition was life-threatening.
A different kind of tool did better. Mount Sinai’s own symptom checker, built on structured clinical protocols, asks about 10 to 14 questions and then points users to a type of care. In a published review of its first year, its recommendations matched a clinician panel 88% to 96% of the time, and no safety incidents were reported. Its developers say tools run by health systems are more likely to be clinical grade than general chatbots. Keep in mind that this evidence measures agreement with clinicians and user satisfaction, not patient health outcomes.
Two Very Different Kinds of Tools
This is the most useful distinction in the whole topic.
A general chatbot (an all-purpose AI assistant) answers from patterns in text. It sounds confident whether or not it is right. It does not have your history, and it asks few follow-up questions unless you tell it to.
A health-system triage tool (often on a hospital or insurer website or app) follows a structured set of clinical questions. It is built to route you to emergency care, urgent care, a doctor, or home care, and it often lets you book an appointment. It is not perfect, but it is designed for the job.
When you have a choice, start with the second kind. Check your own health system’s website or app to see if it offers one.
How to Describe Symptoms So the Answer Is Better
Since wording matters so much, give a chatbot what a doctor would ask for. Copy this format:
- Who: age, sex, and any major conditions or medicines.
- What: the main symptom, in plain words.
- When: how long, and whether it is getting worse.
- How bad: a 1 to 10 scale, and what makes it better or worse.
- What else: fever, nausea, dizziness, anything that feels different, even if it seems minor.
- Your question: “What are the possible causes, and how urgently should I be seen? What would make this an emergency?”
Then add one more line: “Ask me any follow-up questions a doctor would ask before you answer.” This works because the Oxford researchers noted that doctors are trained to ask about symptoms you may not realize you should mention. Prompting the chatbot to do the same helps close that gap. For more on writing better prompts, see why most people fail with AI prompts.
Where AI Is Genuinely Useful for Health
Used as a helper, not a judge, it can do a lot.
- Prepare for an appointment. Ask it to turn your symptoms into a short timeline and a list of questions for the doctor.
- Understand what you were told. Paste in a term from your visit summary (without your name or ID) and ask for a plain-language explanation.
- Learn about a condition you have already been diagnosed with, and check what you learn with your care team.
An illustrative example: someone with two weeks of headaches asks AI to organize when they happened, what was eaten and how they slept, then to draft three questions. The chatbot does not diagnose. It helps the patient communicate.
Experts quoted in coverage of the studies broadly agree on this use. Dr. Robert Wachter of UCSF said he encourages patients to use these tools, and Harvard’s Dr. Adam Rodman said there are no downsides to better understanding your health.
Common Problems and Practical Solutions
Problem: The chatbot said it was probably nothing, but I am still worried. Solution: Trust your worry. A chatbot cannot examine you. If symptoms are new, worsening or unusual for you, call your doctor or a nurse line.
Problem: I cannot afford to see a doctor. Solution: This is a real reason people turn to AI. The KFF poll found 19% of AI health users cited cost. Look into a community health center, a nurse advice line through your insurer, or telehealth, and use AI to prepare for those visits. Do not let cost push you to skip emergency care.
Problem: I want to upload my lab results. Solution: Think twice. KFF found 41% of AI health users had uploaded personal medical information, and 77% of adults worry about the privacy of that data. Remove names and IDs first, and read the tool’s privacy policy. Our guide to AI wearables also covers health data, which is worth knowing before you share it.
Problem: I used the chatbot and never followed up with a doctor. Solution: You are in good company. KFF found 42% of those who asked about physical health did not follow up with a clinician. If it is worth asking about, it is often worth a quick call.
Mental Health Questions Need Extra Care
Many people, especially younger adults, now ask chatbots about mental health. The KFF poll found 28% of adults under 30 had done so, compared with 8% of those 50 and older. A chatbot can offer general information, but it is not a therapist or a crisis service. Our look at AI for therapy and mental health support explains what it can and cannot do. If you are in crisis, call or text 988.
The Bottom Line
A good rule of thumb is to use AI to prepare, not to decide. Let it help you describe symptoms, understand terms and write questions. Let a clinician, a health-system triage tool, or your own judgment in an emergency make the call about urgency.
For related reading, see how AI is changing medical education, how predictive AI works in everyday apps, and the wider picture in AI in daily life. To keep reading plain-language AI guides, you can learn more about us on the AI Overview Search about page.
Frequently Asked Questions
Are AI symptom checkers accurate? It depends on the tool. Studies in 2026 found general chatbots often misjudged urgency, while structured, health-system triage tools matched clinicians much more closely. Neither replaces an exam.
How many people use AI for health information? A KFF poll from February and March 2026 found 32% of US adults had used an AI chatbot for health information in the past year.
Can AI diagnose me? No. It can suggest possibilities, but it cannot examine you or order tests.
Is it safe to use AI in a medical emergency? No. If you think you are having an emergency, call 911 and do not wait for a chatbot.
Is it safe to upload my medical records? Be cautious. Remove personal identifiers and review the tool’s privacy policy first.
