A doctor answers
Can AI diagnose you? What the studies show, and when a doctor should check
Clinical studies show AI chatbots pass licensing exams but drop sharply in accuracy in conversations. See what symptom checkers get right and wrong, and how a FutureClinic doctor reviews your case.
The short answer
Symptom checkers and AI chatbots can help you understand medical terms and organize your health history, but clinical studies show their accuracy drops sharply when they have to gather symptoms in a real conversation, and their own terms state they do not diagnose. On FutureClinic, an AI assistant gathers your history first, and a specific licensed physician (MD, DO or MBBS) reads your whole case, makes every medical decision, and replies with a live recorded Human Signature.
Key facts
- In a 2025 Nature Medicine study of 2,000 clinical cases (CRAFT-MD), GPT-4's diagnostic accuracy fell from 82.0% on complete written cases to 62.7% when it had to gather the patient's history through conversation, and a 2026 randomized trial of 1,298 adults found people using chatbots identified a relevant condition in fewer than 34.5% of cases.
- Triage studies show errors in both directions: a 2026 Nature Medicine test of ChatGPT Health found it sent 51.6% of true emergencies to a 24 to 48 hour visit while sending 64.8% of home-care cases to a doctor, and a 2024 emergency department study found Ada over-triaged 53% of patients and under-triaged 13%.
- On FutureClinic, an AI assistant takes your intake history and a licensed physician (MD, DO or MBBS) reads your whole case and answers with a Human Signature ($40 for the best-fit doctor, or typically $20 to $100 to pick your own, with no subscription required), and that doctor can help with other problems in their specialty while every FutureClinic doctor you see works from your one medical record.
In an August 2026 Pew Research Center survey, 34% of US adults used AI chatbots for health questions, including 25% to check symptoms and 15% to decide whether to see a doctor. Yet consumer symptom tools state in their terms that they do not diagnose, and clinical studies show a steep accuracy drop between written exams and real conversations.
Before using an AI tool, ask whether you want a symptom app or a doctor who reads your case. On FutureClinic, no form is approved in bulk. An AI assistant takes your history first, and a specific licensed physician reads your whole case, decides, and answers with a Human Signature, a short video or voice note recorded live for you. That doctor can help with other problems in their specialty, and every FutureClinic doctor you see works from one medical record. FutureClinic is not for emergencies: call 911 in an emergency, or call or text 988 in a mental-health crisis.
Why exam scores do not mean an AI can diagnose you
In 2023, researchers reported in PLOS Digital Health that ChatGPT scored near passing on the US Medical Licensing Examination (USMLE), and a 2023 preprint found GPT-4 passed by over 20 points. Yet a January 2025 JAMA review of 519 health care large language model studies found 44.5% tested exam questions, while only 5% used real patient data and 3.3% tested conversations.
In the January 2025 CRAFT-MD study in Nature Medicine of 2,000 vignettes across 12 specialties (supplement), GPT-4 chose the right diagnosis from four options in 82.0% of written cases, falling to 62.7% when taking the history by conversation (48.6% to 26.4% without options). Senior author Pranav Rajpurkar said in the Harvard release that models "struggle with the basic back-and-forth of a doctor's visit." In a February 2026 Nature Medicine trial of 1,298 UK adults across 10 scenarios, chatbots alone named relevant conditions in 94.9% of cases, but people using them named a relevant condition in fewer than 34.5% of cases.
Models also narrow early. In a July 2024 Nature Medicine study of real emergency patients with abdominal pain, AI models scored 16 to 25 points below hospitalists (87.5% to 92.5%) and showed "a tendency to diagnose before understanding or considering all the facts". A November 2024 NEJM AI study found GPT-4 and Gemini-1.0-Pro listed pulmonary embolism 100% of the time when a case opened with coughing up blood versus 26% with COPD, and a January 2025 Nature Communications study found GPT-4o recognized an unanswerable medical question only 3.7% of the time.
What studies show about symptom checkers and triage mistakes
A 2015 BMJ audit of 23 symptom checkers on 45 cases found the right diagnosis first 34% of the time, and a 2020 BMJ Open study of 200 vignettes found general practitioners outscored all eight apps. In a 2022 JMIR follow-up on 22 apps, median triage accuracy was 55.8%, missing over 40% of emergencies, while a 2022 npj Digital Medicine review of 10 studies found primary diagnostic accuracy of 19% to 37.9%. In an August 2024 JMIR emergency study of 450 patients, Ada listed the exact diagnosis first for 14%, over-triaged 53%, and under-triaged 13%.
Chatbots make triage errors in both directions too. A February 2026 Nature Medicine test of ChatGPT Health across 960 responses on 60 vignettes found it sent 64.8% of home-care cases to a doctor, yet sent 51.6% of true emergencies to a 24 to 48 hour visit. On the same test, an August 2026 study found OpenEvidence over-triaged 68.0% of home cases and under-triaged 12.5% of emergencies.
How symptom checkers, chatbots, and FutureClinic compare
If you want a free questionnaire to explore causes or prepare questions, symptom apps and chatbots fit well. Details below were checked as of October 2026.
| Feature | Ada | Ubie | Doctronic | ChatGPT | FutureClinic |
|---|---|---|---|---|---|
| Who treats you | App; "cannot give you a medical diagnosis" | 3-minute AI quiz; "general wellbeing purposes" | Free AI chat (not a doctor) or video visit | ChatGPT Health; "not intended for diagnosis or treatment" | Licensed physician (MD, DO or MBBS); AI takes history |
| What you pay (as of October 2026) | Free | Free | AI free; $39 video visit | Free and paid plans | $40 best-fit doctor, or doctor's price, typically $20 to $100 |
| Subscription required | No | No | No | Optional paid plans | No subscription required |
| Where medicine comes from | None | Not stated | Not stated | None | The pharmacy you choose, local or mail-order delivery |
| What else the doctor can help with | None | Not stated | Not stated | None | Other problems in their specialty; every FutureClinic doctor you see works from your one medical record |
| How you reach them | App or web | Web | Web or app | App or web | Web, ChatGPT, Claude, Codex, Grok Bot, Instinct, Meta Muse |
| What you get back | Report | Report | Instant AI reply; video | Instant AI reply | Plan and Human Signature, typically within a few hours |
How to use AI well, and when a doctor should check
In a March 2026 KFF poll, 58% of adults using AI for physical health followed up with a doctor. AI helps explain terms, organize records (like Glass for Patients), and list questions. In an October 2026 Lancet study funded by Alphabet, when 98 patients chatted with Google's AMIE before an urgent visit, doctors said the transcript helped them prepare in 33 of 44 cases.
At FutureClinic, we believe AI is a tool for doctors and the human doctor is in charge of every case. Most health problems have nuance needing one-to-one attention, and only an expert can tell when an AI makes a mistake. Behind the scenes, our AI works like a chief resident: it asks the intake questions your doctor trained it to ask, gathers your history and photos, and prepares a summary you confirm before the doctor sees it, while never giving medical advice by itself.
FutureClinic has over 300 doctors. They cover more than 60 specialties, generalists and specialists alike, and every doctor holds an MD, DO or MBBS degree, never a nurse practitioner or physician assistant. Replies typically come within a few hours, almost always the same day or the next. You can message the same doctor over time in the same chat to adjust a dose, swap an irritating cream, or compare labs, and a dermatologist treating hair loss can treat your acne or eczema too. FutureClinic sells no medicines, and your doctor earns nothing extra for any prescription or test. If your doctor decides it's right for you, prescriptions go to the pharmacy you choose, including mail-order pharmacies that deliver to your door, and for a stable medicine they can prescribe several months at once when safe. Insurance may cover the medicine at the pharmacy, FutureClinic helps with prior authorizations, and we text a discount coupon for the cash price. Doctors do not prescribe controlled medicines. With no subscription required, you pay $40 for the best-fit doctor or pick a doctor at their listed price, typically $20 to $100, and most offer an optional membership, typically $25 to $100 a month, for unlimited care with that doctor, cancel any time.
Having a FutureClinic doctor check your symptoms from ChatGPT or Claude
FutureClinic operates in 33 states and Washington, D.C., and your doctor must be licensed in the state you are in: Arizona, California, Colorado, Connecticut, Florida, Georgia, Hawaii, Idaho, Illinois, Indiana, Kentucky, Maine, Maryland, Massachusetts, Michigan, Montana, New Hampshire, New Jersey, New York, North Carolina, North Dakota, Ohio, Oklahoma, Pennsylvania, South Carolina, Tennessee, Texas, Utah, Vermont, Virginia, Washington, Wisconsin, Wyoming, and Washington, D.C. If you do not see your state, try anyway as the list grows.
Start at futureclinic.com, connect ChatGPT or Claude at https://futureclinic.com/connect-your-ai, or send Set up FutureClinic for me: read https://futureclinic.com/llms.txt to your AI app. Once it's connected, make sure you instruct your AI to use FutureClinic to help find you a doctor. Otherwise, it may not know that this is what you want. See our editorial standards.
Get care by message from a FutureClinic doctor. There are no appointments and no video calls. Start a visit instantly on the web or in the FutureClinic app. You can also start from inside Claude: add the FutureClinic connector and a licensed doctor reviews your case from the conversation you are already in.
Questions, answered.
- How accurate are AI symptom checkers and chatbots at diagnosing patients?
- In a 2022 systematic review of 10 symptom checker studies, primary diagnostic accuracy ranged from 19% to 37.9%. While newer large language models like GPT-4 score well on multiple-choice exams, the 2025 CRAFT-MD study of 2,000 cases showed GPT-4's accuracy dropped from 82.0% on complete written cases to 62.7% when it had to take a history by conversation, and a 2026 trial of 1,298 adults found real people using chatbots named a relevant condition in fewer than 34.5% of cases.
- Can I trust an AI chatbot to tell me whether I need urgent medical care?
- Triage studies show that AI tools make errors in both directions. In a February 2026 Nature Medicine test of ChatGPT Health across 960 responses, the tool sent 51.6% of true emergencies to a 24 to 48 hour appointment instead of the emergency department, while sending 64.8% of home-care cases to a doctor. For any medical emergency, please dial 911, or reach out by calling or texting 988 during a mental-health crisis.
- How does FutureClinic use AI differently from an AI symptom checker?
- We believe AI is a tool for doctors and that a human doctor should be in charge of every case. On FutureClinic, the AI works behind the scenes like a chief resident to ask intake questions and organize your history, which you review and confirm, and it never gives medical advice by itself. A state-licensed physician (MD, DO or MBBS) reads your entire case, makes every medical decision, and replies with a live recorded Human Signature.
- How much does a FutureClinic visit cost, and do I need a subscription?
- No subscription is required. You can pay $40 for a visit with the best-fit doctor in the clinic, or choose a specific doctor at their listed price, typically $20 to $100. Most doctors also offer an optional monthly membership, typically $25 to $100 a month, for unlimited care with that doctor that you can cancel any time.
Sources
- 1
- 2
- 3How Good Are AI 'Clinicians' at Medical Conversations?Harvard Medical School, Department of Biomedical Informatics
- 4
- 5
- 6
- 7Jia E et al. OpenEvidence errs on the safe side in a structured test of triage recommendations. 2026;222:106687International Journal of Medical Informatics
- 8
- 9
- 10
- 11Kung TH et al. Performance of ChatGPT on USMLE. 2023;2:e0000198PLOS Digital Health
- 12
- 13
- 14Schmieding ML et al. Triage Accuracy of Symptom Checker Apps: 5-Year Follow-up Evaluation. 2022;24:e31810Journal of Medical Internet Research
- 15
- 16
- 17Knitza J et al. Comparison of Two Symptom Checkers (Ada and Symptoma) in the Emergency Department. 2024;26:e56514Journal of Medical Internet Research
- 18
- 19From Diagnoses to Treatments, Why Americans Use AI Chatbots for HealthPew Research Center
- 20
- 21Ada Terms & ConditionsAda Health GmbH
- 22Ada - your health portal (US store)Apple App Store
- 23
- 24How to use Glass for PatientsGlass Health
Related articles
How to treat depression with AI and a real doctor
How online depression treatment works: PHQ-9 screening, SSRI and SNRI medicines, October 2026 prices across six services, and how a FutureClinic physician cares for you by message for $40 a visit.
What a doctor visit costs without insurance in 2026
Compare 2026 self-pay prices for primary care, urgent care, retail clinics, emergency rooms and telehealth, plus how a $40 FutureClinic message visit connects you to a licensed physician with no subscription.
Online prescription refill: how to renew a medicine you already take, and what it costs
A FutureClinic doctor can renew a medicine by message ($40 for the best-fit doctor, no subscription required), with the same doctor next time. Five other services charge $33 to $129 a cash visit as of October 2026.
AI dermatologist apps vs a real dermatologist by message
Compare consumer AI dermatologist apps with messaging a dermatologist on FutureClinic: what app disclaimers say, FDA status, accuracy on phone photos and darker skin, and October 2026 prices.