Online symptom checkers are much worse than doctors at naming what is wrong with you. In the best-known head-to-head study, physicians listed the correct diagnosis first 72% of the time and symptom checkers 34%. They do better at the other job, telling you how urgently to seek care, but even there accuracy hovers around 55–60%, so treat their advice as a prompt, not a verdict.
Most people who open a symptom checker want an answer to the wrong question.
They want to know what they have. Is it a migraine or something worse? Is this rash eczema? The tools oblige with a ranked list of possible conditions, and that list is the part people screenshot and send to their family.
But the question a symptom checker is built to answer, and the only one where it has any real track record, is a different one: where should you go, and how soon? Home, a pharmacist, a GP next week, urgent care today, an ambulance now. That is triage, and it is a separate skill from diagnosis.
Here is the non-obvious part. For years the standard criticism of symptom checkers was that they were too cautious, sending everyone with a cough to a doctor. The more recent evidence says some of them have drifted the other way and now miss a large share of real emergencies. Knowing which job you are asking a tool to do, and how well it does that job, is most of what a patient needs to know.
How accurate are symptom checkers at diagnosis?
Not very, and the evidence has been consistent for a decade.
The landmark study came from Harvard Medical School researchers led by Hannah Semigran, published in The BMJ in 2015. They ran 45 standardised patient cases through 23 English-language symptom checkers. The cases ranged from emergencies such as a pulmonary embolism to things that are fine to manage at home, like a common cold. The tools listed the correct diagnosis first in about a third of cases and somewhere in their top 20 suggestions in just over half.
A year later the same group did something unusual: they gave the same cases to doctors. In a 2016 research letter in JAMA Internal Medicine, 234 physicians working through the cases on an online platform named the correct diagnosis first 72.1% of the time, compared with 34.0% for the symptom checkers. Within the top three suggestions, doctors reached 84.3% and the tools 51.2%. The doctors had only the written case, with no examination and no test results, which is roughly the same information the apps get.
Later reviews have not changed the picture much. A systematic review in npj Digital Medicine in 2022, by Wallace and colleagues, pooled studies across many conditions and found diagnostic accuracy that varied enormously but was often poor, including primary diagnosis accuracy as low as single digits for some infectious diseases. The honest summary is that a symptom checker’s top suggestion is closer to a guess than a diagnosis.
How accurate are symptom checkers at triage?
Better than at diagnosis, but not as good as many users assume.
In the 2015 BMJ study, the tools gave appropriate triage advice 57% of the time overall. They were best with true emergencies, where they recommended urgent care in about 80% of cases, and worst with minor problems, where they advised self-care in only a third of cases that did not need a doctor. The authors described the tools as generally risk averse. They erred towards sending people for care.
That reputation stuck. It was also, for a while, a reasonable trade-off. An app that sends too many people to the GP wastes appointments. An app that sends too few people to the emergency department can kill someone.
Then researchers in Berlin repeated the test. In a 2022 study in the Journal of Medical Internet Research, Schmieding and colleagues ran the same 45 cases through 22 symptom checkers available in 2020 and compared the results with 2015. Median triage accuracy had barely moved, from 59.1% to 55.8%. But the pattern of errors had shifted. The 2015 tools made nearly three over-triage errors for every under-triage error. By 2020 the ratio was close to one to one, and the apps were missing more than 40% of emergencies. Few of the apps did better than ordinary laypeople at deciding whether a case needed emergency care, and none beat laypeople on both the emergency and the self-care decision.
That is the finding worth remembering. The common assumption that a symptom checker will at least err on the side of caution is not safe.
Are AI chatbots better than older symptom checkers?
Large language models are far more fluent than the decision-tree apps of 2015, and on written cases they can name the right condition very often. But the evidence from real users is sobering.
A randomized study from the University of Oxford, published in Nature Medicine in February 2026, gave 1,298 members of the UK public medical scenarios to work through with a chatbot or with their usual sources. The models alone chose the right course of action in 56.3% of scenarios, strikingly close to the triage accuracy of the old symptom checkers. When people used the models themselves, they chose the right action in fewer than 44.2% of cases, no better than a control group using ordinary search.
So the newest tools share the old problem. Knowing a lot of medicine does not automatically translate into telling a real person, with an incomplete story, whether to stay home or go to hospital. The researchers traced much of the gap to the conversation itself: people left out details the model needed, and the answers mixed good advice with poor advice. A decision tree at least asks every user the same questions. A chatbot asks whatever it thinks of, and the user decides what to volunteer.
A worked example: what the numbers mean for one person
This example is hypothetical, using the published accuracy figures to show their practical meaning.
Tom, 34, wakes up with a headache, a stiff neck and a temperature of 38.6°C. He opens a symptom checker. Its top suggestion is a viral infection, with meningitis further down the list. Its triage advice is to see a GP within a day or two.
If Tom treats the tool as a diagnostic engine, he reads “viral infection,” takes paracetamol and goes back to bed. With first-place accuracy around a third in the studies above, that top suggestion is a weak basis for anything.
If Tom treats the tool as a triage engine, he looks at the advice to be seen. But the 2022 data suggests that triage advice can be too relaxed as often as it is too cautious. Headache, fever and a stiff neck together are a well-known warning combination that warrants urgent medical assessment. A clinician on a phone line would very likely tell him to be seen urgently.
The lesson is not that the app was useless. It correctly put a serious condition on the list and told him to see someone. The lesson is that the user must know which output to trust least, and when their own sense that something is badly wrong should override the screen.
The strongest case for symptom checkers
The counterargument deserves a fair statement. Symptom checkers are not competing with a doctor who has read the case carefully. For most people at 11pm, they are competing with a search engine, a forum or nothing.
Compared with that, a tool that says “this needs a doctor” in about four out of five true emergencies, as the 2015 tools did, has real value. Some national health services have built triage tools into their phone and online services precisely because they can handle huge volumes at low cost. And the vignette studies have limits: written cases are not real patients, and some developers argue their tools perform better in real use, with follow-up questions, than in fixed test scripts.
Those points are fair. They also narrow the claim to what the evidence supports. Symptom checkers can be a reasonable first filter for people deciding whether to seek care. They are not a way to find out what you have, and their triage should never override red-flag symptoms or your own strong sense that you are seriously unwell. Primary care that is easy to reach remains the better filter, which is part of our argument that primary care is the cheapest hospital bed.
What this means for patients, health services and developers
For patients. Use a symptom checker, if at all, for the question “how soon should I be seen?”, and treat the diagnosis list as background reading. If a tool tells you to stay home but you have warning signs such as chest pain, difficulty breathing, sudden weakness, confusion, a severe headache with a stiff neck, or a feeling that something is very wrong, seek urgent care anyway. No app is a substitute for a clinician, and none can examine you.
For health services. If you deploy a triage tool, publish how it performs on under-triage, not just average accuracy. The error that matters most is the emergency sent home. Clinical safety review should look specifically at that number and how it changes with every update.
For developers and regulators. The 2022 follow-up showed that tools can change behaviour between versions without users noticing. Accuracy reported at launch says little about accuracy two years later. Independent retesting on standard case sets, repeated over time, would tell the public far more than marketing claims. That point matches a theme we have made before: AI evidence is strongest where the question is narrowest, and “where should I go?” is a narrower question than “what do I have?”
Frequently asked questions
Are online symptom checkers accurate?
For diagnosis, not very. In the 2015 BMJ study of 23 tools, the correct diagnosis came first in about a third of cases. For triage, meaning how urgently to seek care, accuracy was higher but still modest, around 55–60% in studies from 2015 and 2020. They can be a rough guide but should never replace a clinician’s assessment.
Are doctors more accurate than symptom checkers?
Yes, by a wide margin in head-to-head tests. In a 2016 JAMA Internal Medicine study using the same written cases, physicians named the correct diagnosis first 72.1% of the time, compared with 34.0% for symptom checkers. The doctors had no examination or test results, only the written description, yet still roughly doubled the tools’ accuracy.
Do symptom checkers always err on the side of caution?
Not any more. Tools tested in 2015 tended to over-triage, sending people to care they did not need. A 2022 follow-up in the Journal of Medical Internet Research found 2020 versions were less cautious and missed more than 40% of emergencies. Never rely on an app’s reassurance if you have warning signs of a serious illness.
Are AI chatbots better than symptom checkers?
Chatbots know far more medicine and write more naturally, but real-world results are not clearly better. A 2026 Oxford study in Nature Medicine found people using chatbots chose the right course of action in under 44% of scenarios, no better than people using ordinary search. Fluency is not the same as safe triage advice.
When should I ignore a symptom checker?
Ignore reassuring advice whenever you have red-flag symptoms: chest pain, difficulty breathing, sudden weakness or numbness, confusion, fainting, a severe headache with fever or a stiff neck, heavy bleeding, or thoughts of self-harm. In those cases contact emergency services or seek urgent care directly rather than relying on any app.
A symptom checker is a triage nurse with a weak memory and no eyes. It is most useful when it tells you to get help, and least trustworthy when it tells you not to.
Has a symptom checker ever sent you the right way, or the wrong way? We would like to hear which.
