Is it safe to use ChatGPT for medical advice?

What does OpenAI itself say about medical advice?

OpenAI's Terms of Use, effective January 1, 2026, say that output may not always be accurate. They tell you not to rely on it as a sole source of truth or factual information, or as a substitute for professional advice. They also say you must evaluate output for accuracy before you use or share it.

OpenAI's Usage Policies, in the version effective November 20, 2026, bar anyone from building or offering AI tools that give others tailored advice requiring a license without appropriate involvement by a licensed professional. They also bar stating or implying that the AI itself holds professional credentials.

That rule is aimed at people building products on OpenAI's models. It is not a ban on you asking ChatGPT a health question. But read together, the two documents are clear: ChatGPT is not presented as a doctor, and you are expected to check what it tells you.

What can go wrong when you ask a chatbot about your health?

The risks below come from WHO and peer-reviewed research. None of them means every answer is wrong. They mean you often cannot tell from the answer alone whether it is right.

• False or incomplete answers. WHO's 2024 guidance on large multi-modal models notes documented risks of producing false, inaccurate, biased, or incomplete statements, which could harm people who rely on them for health decisions.
• Going along with a false premise. A 2025 study in npj Digital Medicine asked models to write a persuasive letter saying a brand-name drug had new side effects and people should take the generic instead, even though the two are the same drug. At baseline, before any prompt changes, GPT-4o, GPT-4, and GPT-4o-mini wrote the letter in 100% of 50 tests.
• No exam and no tests. A chatbot only knows what you type. In the Nature Medicine study, the authors saw cases of people giving incomplete information and of models misreading what people asked.
• Leaning on it too much. WHO warns that these tools can encourage automation bias, where patients or clinicians miss errors or hand off hard choices to the model.

What happened when ordinary people used chatbots for medical decisions?

Researchers at Oxford published a randomized, preregistered study in Nature Medicine in February 2026. They gave 1,298 UK adults medical scenarios written by doctors and asked them to name likely conditions and pick a next step, from self-care up to calling an ambulance. Some used GPT-4o, Llama 3, or Command R+. A control group used whatever they would normally use at home, mostly internet search.

On their own, the models did well. Given the full scenario, GPT-4o named at least one relevant condition in 94.7% of cases. But people using the chatbots were less likely than the control group to identify a relevant condition, and they did no better at choosing the right next step. Across all groups, only 43% picked the correct course of action, and people tended to underestimate how serious their scenario was.

The models in this study are older than what ChatGPT runs today, so newer versions may do better. The main lesson still applies: a chatbot that knows the medicine can still steer you wrong when the conversation leaves out key details, or when you cannot tell its best suggestion from its worst.

How can you use ChatGPT more safely for health questions?

These habits follow from the risks above. They lower the odds of a bad answer. They do not make a chatbot a replacement for a clinician.

• Use it to understand, not to decide. Explaining a diagnosis you already have, a lab term, or how a medicine works is lower risk than deciding whether to go to the ER.
• Give the details a clinician would ask for. Your age, when symptoms started, how bad they are, what makes them better or worse, your conditions, and every medicine and supplement you take.
• Ask neutral questions. "What is the difference between these 2 drugs?" is safer than "Why is this drug safer than that one?" because leading questions can pull the model into agreeing with a false premise.
• Ask what would make it urgent. Ask which symptoms would mean you need same-day care or 911, then watch for them.
• Check it against a trusted source. Compare key facts with pages from MedlinePlus, CDC, or the drug label, and ask the chatbot where its answer comes from.
• Never start, stop, or change a medicine based on a chatbot alone. Check with your prescriber or pharmacist first.
• Bring the answer to your visit. A chatbot is useful for writing down your questions so the appointment is spent on decisions.

When is a general chatbot the wrong tool?

Skip the chatbot and go straight to care for anything that could be an emergency, for symptoms that are new, severe, or getting worse, and for any decision about starting or stopping a medicine. If you are in a mental health crisis, call or text 988.

A general chatbot also answers whatever you ask in the order you ask it. A tool built for health questions can be designed to ask follow-up questions and flag red flags first. Our comparison of Pymander vs ChatGPT for health covers those design differences, and our guide on whether ChatGPT can replace a doctor covers the bigger question. Our AI doctor is free to text, and it does not replace a clinician either.