Is It Safe to Ask AI About Your Symptoms?

We make an AI health information app, so you should read this page knowing that. We think the honest answer is more useful to you than a sales pitch: AI chat tools can be genuinely helpful for some health tasks, are unreliable for others, and should never be the thing that decides whether you get care.

What the evidence says

The most useful test of AI for health questions is not whether it can pass a medical exam — large language models can — but whether ordinary people get better answers when they use it. A randomized study of 1,298 people published in Nature Medicine in 2026 tested exactly that.^[2] Given medical scenarios on their own, the AI models identified relevant conditions in 94.9% of cases. But when members of the public used the same models to work through the scenarios, they identified relevant conditions in fewer than 34.5% of cases and chose the right course of action in fewer than 44.2% — no better than people using whatever sources they normally would.

The researchers found the breakdown happened in the conversation: people left out information the model needed, and models gave correct answers in ways people didn’t pick up on.^[2] That matches what anyone who has used these tools has seen. The answer depends on the question, and the question depends on what you thought to mention.

The World Health Organization’s warning was about exactly this: large language models generate responses that can appear authoritative and plausible, whether or not they are right.^[3]

What AI tools are good for

• Explaining words and documents. What a term on your after-visit summary means, or what a test generally measures.
• Getting ready for an appointment. Turning a jumble of symptoms into a clear timeline, and drafting questions to ask. Our question list is a good starting point too.
• Understanding a plan you have already been given. Going over what your clinician said, so you can follow it — and noticing what to ask them about.
• Keeping track. Some tools, including ours, keep a record of what you report so that patterns over time are visible.

Where it fails

• It cannot examine you. It cannot look at your throat, listen to your chest, or see how unwell you look. Clinicians rely on all of that.
• It only knows what you tell it. Leave out one detail and the answer can change completely.
• It can be confidently wrong. A wrong answer reads exactly like a right one.
• Small changes in wording can change the answer. Ask twice, slightly differently, and you may get two different responses.
• It can underplay urgency. Or overplay it. Neither is a substitute for the emergency warning signs.

How to use it more safely

• Use it to prepare for care, not to replace it. Take what you learn to a clinician as questions, not conclusions.
• Give it the facts a clinician would want: when it started, how it has changed, how bad it is, what you have tried, your conditions and medicines. The guide to describing symptoms lists them.
• Check anything important against a reliable source such as a government health agency or your clinician.
• Never let a reassuring answer talk you out of care you think you need.
• Know what happens to what you type: whether the service keeps your conversations, for how long, who it shares them with, and whether they are used to train its models.

How PocketHealth approaches this

Our assistant says it is AI. It gives general health information, not a diagnosis, and it does not prescribe. When someone describes signs of an emergency, it tells them to call 911 straight away, and for thoughts of suicide it points to 988. It keeps what you report in a timeline you can see, correct and delete, and it turns that history into a summary you can bring to a clinician. How we handle your information is in our consumer health data policy and on our security page. None of that removes the limits above; it is designed around them.

If you have used a general chatbot for health questions, you may also have noticed it doesn’t keep your history the way a record does. Why that happens is explained in a separate guide.