Are AI symptom checkers accurate? What studies show, and how to use one safely

An AI symptom checker can feel like a doctor in your pocket. You type what's wrong, and seconds later you get a list of possible causes and advice on what to do next. It's fast, often free, and there at 2 a.m. when nobody else is. But how accurate is it? Researchers have tested these tools for more than a decade, and the answer is more useful than a plain yes or no.

This guide covers what the studies found, where these tools help, where they go wrong, and a checklist you can use with any of them.

What an AI symptom checker actually does

There are two broad kinds:

• Symptom checker apps and websites. They ask you a series of questions, then give a list of possible conditions and a triage level: look after yourself at home, see a doctor, or get emergency care.
• General AI chatbots. You talk to them in your own words, and some now have health sections. They explain things in full sentences, but they don't always follow a fixed set of questions.

Both do two separate jobs, and it helps to keep them apart:

1. Diagnosis: guessing what the problem might be.
2. Triage: advising how soon and where to get care.

Studies show these tools are usually better at the second job than the first. For your safety, the second job matters more.

Are AI symptom checkers accurate? What the research shows

The early tools: right about a third of the time

A well-known 2015 audit in the BMJ tested 23 online symptom checkers with standard patient cases. The correct diagnosis came first in only about a third of cases, and appeared anywhere in the top 20 in just over half. Triage advice was better, but quality varied widely between tools, from about a third of cases to nearly four in five. Some checkers sent people to a doctor when home care would have been fine. The authors concluded the data did not support using these tools for diagnosis or triage.

Later reviews: still very uneven

A 2022 systematic review in npj Digital Medicine pulled together 10 studies covering 48 tools. Getting the main diagnosis right ranged from roughly 19% to 38%. Triage was better, from about 49% to 90%, but it swung widely between tools and conditions. The reviewers called overall accuracy low and raised safety concerns.

A 2025 review in the same journal compared symptom apps, AI chatbots and ordinary people deciding on their own. Apps ranged from about 12% to 90% accurate at triage. Chatbots sat in a narrower middle band, about 58% to 76%. People alone landed at about 47% to 62%. The authors' view was balanced: these tools should be neither recommended nor dismissed across the board. It depends on the tool and who is using it.

2026: the newest chatbots

Two studies published in Nature Medicine in February 2026 looked at the latest generation of AI.

The Mount Sinai study was the first independent safety test of ChatGPT Health after its January 2026 launch. Researchers wrote 60 patient scenarios across 21 specialties and ran each one under 16 variations, for 960 conversations. Three physicians set the right answer for each case, using guidance from 56 medical societies. The tool handled textbook emergencies, like stroke and severe allergic reactions, well. But in more than half of the cases doctors said needed emergency care, it advised a lower level of care. In one asthma case, its own explanation spotted early signs of breathing failure, yet it still suggested waiting. Suicide-crisis messages also appeared inconsistently.

The researchers were clear about limits: the scenarios were written, not real patients, the test was a single point in time, and children, medicines and non-English use weren't covered. They don't tell people to stop using AI health tools, only to treat them as aids rather than a stand-in for clinical judgment.

The Oxford study asked a different question: do these tools help real people decide better? Nearly 1,300 UK adults worked through doctor-written scenarios, either with an AI model or with whatever they normally use, such as a web search. Given the cases directly, the models named a relevant condition in about 95% of them. But when people used the same models, they found a relevant condition in fewer than 35% of cases and chose the right next step in fewer than 44%. That was no better than the group without AI. People often didn't know which details to give, small wording changes produced very different answers, and replies mixed good and bad advice that was hard to tell apart.

What this adds up to

• Obvious emergencies are handled best. Subtle ones are where tools slip.
• Accuracy varies a lot between tools, and the same tool can change after an update.
• Scoring well in a lab is not the same as helping you. The conversation between person and machine is where things often go wrong.
• No study supports using these tools to diagnose yourself.

Where symptom checkers can genuinely help

Used well, an AI symptom checker can still be worth your time:

• Finding words for what you feel. It can help you describe a symptom clearly before you call a clinic.
• Explaining terms. It can tell you what a word on a discharge note means, in plain language.
• Preparing for an appointment. It can suggest questions to ask and details a doctor will want.
• Calm for clearly minor problems, as long as you know the warning signs that would change the picture.
• Access. It's there at any hour and often in many languages, which matters when care is far away or hard to reach.

Where they fail

• Subtle emergencies. A heart attack without much chest pain, or a child who is unusually quiet, can sound mild in text.
• Missing context. The tool only knows what you type. If you don't mention that you're pregnant, take a blood thinner or have diabetes, it can't factor that in.
• Following your lead. If you describe things as "probably nothing", the answer may lean the same way.
• Confident tone. A smooth, well-written reply can sound more certain than it is.
• No exam, no tests. It can't listen to your chest, feel your belly or check your oxygen level.

A checklist for using any AI symptom checker safely

Step	Why it matters	What to do
Red flags first	Tools can miss urgent cases that aren't obvious	Check emergency warning signs yourself before typing anything
Give full context	Answers depend on what you share	Include age, pregnancy, long-term conditions and every medicine you take
Describe plainly	Wording changes answers	Say what happened, when it started and how it's changing, without downplaying it
Ask about urgency	Triage matters more than a label	Ask directly: "How soon should I be seen, and where?"
Treat it as a second opinion	Studies don't support self-diagnosis	Take the answer to a doctor, nurse or pharmacist instead of acting on it alone
Never delay emergency care	Minutes count in some conditions	If in doubt, call your emergency number first and ask questions later
Protect your privacy	Many health apps aren't covered by medical privacy laws	Read the privacy policy, check sharing settings and share only what's needed

Red flags that mean skip the app

Call your local emergency number or go to the nearest emergency department if you, or someone with you, has:

• Chest pain, pressure or tightness
• Trouble breathing
• Face drooping, arm weakness or slurred speech
• A severe allergic reaction, such as swelling of the lips or throat
• Sudden confusion, fainting or a seizure
• Heavy bleeding that won't stop
• Thoughts of self-harm

Think about where your words go

The US Federal Trade Commission warns that many health apps aren't covered by medical privacy laws like HIPAA, and some share data with advertisers or other companies. Before you describe symptoms in detail, check what the app collects, who it shares with, and whether you can turn sharing off. Comparing the privacy promises of two similar apps is a sensible step before you pick one.

How Saluvo handles this

Saluvo's chat answers health questions in many languages, and it checks your message for emergency red flags with fixed rules before any AI replies, so warning signs get a clear "get help now" answer. It can't diagnose you, and its answers aren't reviewed by a doctor. Use it the way this guide suggests: to understand and prepare, not to replace care.

The bottom line

AI symptom checkers are improving, but the research agrees on one point: they are weakest exactly where the stakes are highest, in emergencies that don't look like emergencies. Check the red flags yourself, give the full picture, and use the answer as a step toward talking to a professional, never as a reason to wait.