A third of American adults asked an AI chatbot a health question in the past year, according to KFF polling from March. Among those users, 41 percent uploaded personal medical information to do it. If that’s you, I’d like to share a few pitfalls and pointers. I’m a practicing gastroenterologist who also works in health tech.

Let’s start with the pitfalls. There are two that most people haven’t been told about.

The first is privacy: Your medical information can outlive the conversation. The free versions of ChatGPT, Gemini, and Claude all use your conversations to train future models unless you find the setting that turns that off, and most people never do. The symptom you describe, the lab report you paste, the colonoscopy result you photograph, all by default become training data. Medical details identify you more than you’d guess. Privacy researchers demonstrated decades ago that 87 percent of Americans can be identified from just ZIP code, sex, and date of birth. The privacy rules that govern your medical chart stop at the chatbot’s front door.

The second is guidance: The answer itself can hurt you. A chatbot sounds equally confident when it’s right and when it’s wrong, a trait you don’t want in a medical adviser. Sometimes the answer is incorrect. Sometimes it’s incomplete in a way that matters, because the chatbot doesn’t know your medications, your family history, or the detail you didn’t think to mention. This isn’t hypothetical. In a published case report, a man trying to cut salt from his diet asked ChatGPT about replacing chloride and came away substituting sodium bromide, a chemical cousin of table salt that’s toxic to eat. Three months later he spent three weeks in the hospital with paranoia and hallucinations from bromide poisoning. It was a technically true answer that never asked the question that mattered: what he planned to do with it. Sometimes it’s a bad call on the question that matters most, whether you need to be seen by a real clinician, and how fast. Gallup found that one in seven recent AI health users skipped a doctor’s visit based on what a chatbot said, and 11 percent reported getting advice they later learned was unsafe. The deepest harm here isn’t usually a wrong fact. It’s an answer built on an incomplete picture of you.

Now to pointers. The first two protect your privacy, while the third protects you from bad guidance.

Pointer one: Flip the privacy switch

Every major chatbot has a setting that keeps your conversations out of training. In ChatGPT it’s under Data Controls. Gemini and Claude have equivalents. Several also offer a temporary mode that keeps nothing at all. Turn the training setting off once, and use temporary mode for anything medical. It takes a minute, and the answers are identical.

Pointer two: Strip the identifiers before you paste

Before you paste a lab report, a visit summary, or a question about your symptoms, delete your name, date of birth, address, phone number, record numbers, and the name of your clinic or hospital. The answer does not get worse. A chatbot explaining your iron studies needs the numbers, not your name. The medical content of your question is what’s useful. The identity attached to it is what’s risky.

Pointer three: Use chatbots to prepare, never to decide

Used well, AI is a translator and a preparation tool. Paste a de-identified test result and ask what the terms mean. Describe your symptoms and ask what questions a doctor would want answered, so you walk into your appointment organized.

AI chatbots should never decide whether you need care. They can’t examine you, can’t order the test that settles the question, and don’t notice what you left out. In my corner of medicine, some symptoms need expedited evaluation:

• Black, tarry stools
• Trouble swallowing
• Unintentional weight loss
• New yellowing of your skin or eyes

None of these are ask-the-internet problems, and neither is anything that’s getting worse while you research it. If you find yourself asking a chatbot whether something is serious enough to see a doctor about, treat the question itself as the answer.

Your questions about your body are good questions. Keep asking them. Just ask them with the identifiers stripped, the training switch off, and the one decision that matters, whether to be seen, reserved for someone who can actually see you.

This essay is cited in the KevinMD record on artificial intelligence.

Simon Mathews is a gastroenterologist.