Skip to main content
    Path to Science

    Journal

    Can I trust what ChatGPT tells me about my health?

    Imagine it's late at night and you have a strange ache in your side. You don't want to bother anyone, so you type your symptoms into a chatbot. Within seconds you get a calm, tidy answer that lists three possible causes, explains each one in plain words, and suggests you keep an eye on it. It sounds like a kind, well-read doctor, so you feel better and go to sleep.

    But how much should that answer count for? A smooth answer and a correct answer are not the same thing. In this essay we'll look at what researchers have found when they tested chatbots on health questions, and we'll build a simple habit for checking any answer you get.

    The question

    Here's the question in plain words: when a chatbot gives you a health answer, can you trust it as it is?

    We're not asking whether chatbots are "good" or "bad." That's too broad to test, so we're asking something narrower. If you rely on the answer without checking, how likely are you to end up with the right idea and the right next step? By the end, you should know what the studies can and can't tell us, and how to check an answer yourself.

    What the evidence actually says

    Lots of people already ask chatbots about their health. In a Pew Research Center survey of 5,111 US adults, run on 20–26 October 2025 and published in April 2026, 22% said they get health information from AI chatbots at least sometimes.⁶ Among those users, 23% said the information was not too accurate or not at all accurate, while 18% said it was highly accurate. By comparison, 65% of respondents said health care providers were highly accurate.⁶

    Doubt is common too. In a KFF poll from June 2024, 56% of people who used AI chatbots were not confident the health information was accurate. Among all adults, the figure was 63%.⁷ And in an October 2025 NORC survey, 46% of US adults said clear sources, including links, would make them trust AI health information more.⁸

    Of course, what people think about accuracy isn't the same as accuracy. For that we need tests, and every test below covers a specific model at a specific time. New versions come out every few months and they can do better or worse.

    The model alone versus real people using it. In a randomized trial published in February 2026, researchers tested three models: GPT-4o, Llama 3 and Command R+.¹ First, they gave the models ten written medical scenarios, created by doctors. Working alone, the models named a relevant condition 94.9% of the time on average. They picked the right next step, such as self-care, seeing a GP or going to A&E, 56.3% of the time.

    Then came the real test: the researchers gave the same scenarios to 1,298 UK adults. Some used a chatbot, while a comparison group used whatever they normally would, such as an internet search. People using the chatbots identified a relevant condition less than 34.5% of the time. They chose the right next step less than 44.2% of the time. That was no better than the people in the comparison group.¹ The authors note that standard medical tests for these models did not predict how badly things went with real users. A model that does well on a clean, complete case can still fail when a real person describes symptoms in their own messy way.

    A stress test of emergency advice. A study published in February 2026 tested ChatGPT Health, a consumer health tool launched in January 2026.² Doctors wrote 60 short patient stories across 21 medical areas, and each story was tested under 16 conditions. That produced 960 responses in total. The study didn't state which underlying model the tool used. The tool handled classic emergencies like stroke and severe allergic reaction correctly. But in this test, 52% of true emergency cases were given less urgent advice than they needed. Some of these were told to get checked within one or two days instead of right away. The advice also shifted when a friend or relative in the story played down the symptoms.² This was a written stress test, not a count of real-world errors. A preprint, not peer reviewed, argues that the test format may have made the problem look bigger.⁹

    Wrong details in your question. In a 2025 study, researchers tested six models with 300 medical cases.³ Each case included one invented detail, such as a lab test or disease that doesn't exist. Depending on the model and setup, the models repeated or built on the false detail in 50–82% of their answers. Adding a safety instruction helped but didn't solve it: the average fell from 66% to 44%. GPT-4o, the best performer, went from 53% to 23% with the safety instruction.³ The lesson for you is simple: if your question contains a wrong assumption, the chatbot may run with it.

    Sources that don't exist or don't match. In June 2025, researchers asked GPT-4o to write six reviews on mental health topics and checked its 176 references.⁴ About one in five (19.9%) were made up, and of the real ones, 45.4% had errors, often a broken link. Made-up references were more common for less-studied conditions.⁴ A separate 2025 study tested several models, including GPT-4o with web search, on 800 medical questions.⁵ Between 50% and 90% of the answers were not fully backed up by the sources the models themselves cited. In other words, a link being there didn't mean the link supported the claim. These models are already older versions, so today's results could differ.

    Evidence label: Mixed (fast-moving). Models tested alone on tidy cases can score well, but in one large trial, real people using them did no better than people using their usual sources. Several studies document specific failures. The evidence supports checking chatbot answers. It doesn't support calling chatbots either dangerous or as good as doctors, and it goes out of date quickly.

    The thinking tools

    Credible sources

    A credible source is one you have good reasons to trust, such as a national health service or a peer-reviewed study. A chatbot is better thought of as a fluent summary than as a source. It blends patterns from many texts, and it may add references that are wrong or invented.⁴˒⁵ So treat its answer as a starting point, because the real question is always: where does this claim come from, and does that place actually say it?

    Lateral reading

    Lateral reading means leaving the page you're on to check what other trusted sites say about a claim. Fact-checkers do this instead of studying one page closely. With a chatbot, it means opening a new tab and looking up the claim on an official health site. If the chatbot gave a link, open it and see whether it really says what the chatbot claimed.⁵

    Confirmation bias

    Confirmation bias is our habit of favouring information that fits what we already believe or hope. Chatbots can feed this: if your question already assumes an answer, a chatbot may go along with it, as the planted-detail study showed.³ The Pew survey found that frequent users were much more likely to call chatbots highly accurate (45%) than occasional users (13%).⁶ That gap could have many explanations, and the survey doesn't tell us which. But it's a good reminder to notice when an answer feels right mostly because it matches what you wanted to hear.

    Multiple lines of evidence

    Multiple lines of evidence means trusting a conclusion more when separate, independent sources point the same way. One chatbot answer is one line, while an official health page and your own doctor are others. When they agree, you can be more confident. When they disagree, trust the official source and your clinician, and ask more questions.

    So what should I do?

    Here are the facts: chatbots can explain medical ideas clearly, and alone they often name the right condition in tidy test cases. But in one large trial with real people, that didn't lead to better choices. They can also miss urgency, repeat wrong details and cite sources that don't support them.

    How much you use them is partly a question of values, such as convenience versus caution. (See facts vs values.) This essay doesn't recommend or warn against any particular product. But some rules hold whatever you decide:

    • Emergencies: if you think it might be an emergency, call emergency services (such as 112, 999 or 911). Don't ask a chatbot first.
    • Medicines and treatment: don't start, stop or change any medicine or treatment based on a chatbot answer. Check with a doctor or pharmacist.
    • Any health decision: use the chatbot to prepare questions, then talk to a doctor before you act.
    • Privacy (general advice): think before sharing details that could identify you, such as your name, date of birth or full medical history. Read how the service stores your data first.

    And that late-night ache from the start? A new pain that worries you is a reason to see a doctor, not to rely on a chatbot's answer. If it's severe, call emergency services.

    Try this week

    The three-question check

    • What to do: Ask a chatbot three general health questions you're curious about, such as what a common condition is or how a common medicine works. Don't use it for a symptom you're worried about right now.
    • How to check: For each answer, open a new tab and look up the same question on an official site such as the NHS, CDC or WHO. If the chatbot gave links, open them and check that they say what it claimed.
    • What to record: For each question, mark "agrees," "partly agrees" or "disagrees." Note any links that were broken or didn't support the claim.
    • How to read it: Three questions won't tell you how accurate chatbots are in general, because that sample is far too small. But it can show you how often you'd have missed something without checking.
    • When to get help: If any question involves your own symptoms or medicines, take what you found to a doctor or pharmacist.

    Key takeaways

    • A confident, fluent answer is not the same as a correct one. In a 2026 trial, people using three chatbots did no better than people using their usual sources.¹
    • Tests have found specific failures: under-triaged emergencies in a written stress test, repeated false details, and invented or unsupported references.²˒³˒⁴˒⁵
    • Check any health answer against an official source, and talk to a doctor before making a health decision. In an emergency, call emergency services.

    Sources

    1. Bean, A. M., Payne, R. E., Parsons, G., Kirk, H. R., Ciro, J., Mosquera-Gómez, R., Hincapié M, S., Ekanayaka, A. S., et al. (2026). Reliability of LLMs as medical assistants for the general public: A randomized preregistered study. Nature Medicine, 32(2), 609–615. https://doi.org/10.1038/s41591-025-04074-y
    2. Ramaswamy, A., Tyagi, A., Hugo, H., Jiang, J., Jayaraman, P., Jangda, M., Te, A. E., Kaplan, S. A., et al. (2026). ChatGPT Health performance in a structured test of triage recommendations. Nature Medicine, 32(5), 1671–1675. https://doi.org/10.1038/s41591-026-04297-7
    3. Omar, M., Sorin, V., Collins, J. D., Reich, D., Freeman, R., Gavin, N., Charney, A., Stump, L., et al. (2025). Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support. Communications Medicine, 5. https://doi.org/10.1038/s43856-025-01021-3
    4. Linardon, J., Jarman, H. K., McClure, Z., Anderson, C., Liu, C., & Messer, M. (2025). Influence of topic familiarity and prompt specificity on citation fabrication in mental health research using large language models: Experimental study. JMIR Mental Health, 12, e80371. https://doi.org/10.2196/80371
    5. Wu, K., Wu, E., Wei, K., Zhang, A., Casasola, A., Nguyen, T., Riantawan, S., Shi, P., et al. (2025). An automated framework for assessing how well LLMs cite relevant medical references. Nature Communications, 16. https://doi.org/10.1038/s41467-025-58551-6
    6. Pasquini, G., Stocking, G., Kikuchi, E., Pula, I., & Yam, E. (2026, April 7). Users of social media and AI chatbots for health information are more likely to say they are convenient than accurate. Pew Research Center. https://www.pewresearch.org/science/2026/04/07/users-of-social-media-and-ai-chatbots-for-health-information-are-more-likely-to-say-they-are-convenient-than-accurate/
    7. KFF. (2024, August 15). KFF Health Misinformation Tracking Poll: Artificial intelligence and health information. https://www.kff.org/public-opinion/kff-health-misinformation-tracking-poll-artificial-intelligence-and-health-information/
    8. NORC at the University of Chicago. (2025). AI is changing how Americans access health information—but many remain skeptical (AmeriSpeak Omnibus, 9–13 Oct 2025). https://www.norc.org/research/library/ai-changing-how-americans-access-health-information-many-remain-skeptical.html
    9. Fraile Navarro, D., Sheng, J., Magrabi, F., & Coiera, E. (2026). Evaluation format, not model capability, drives measured triage failure in the assessment of consumer health AI. arXiv preprint 2603.11413 (not peer reviewed). https://arxiv.org/abs/2603.11413