Over the past few years, something has changed when I see patients in the ER. Patients used to come with printouts from Google. Today, they have their AI chat.

According to a Wolters Kluwer’s survey , 60% of clinicians said they spend time discussing AI-generated health information their patients bring with them. And 42% of patients said they frequently or very frequently bring it. OpenAI has reported that more than 300 million people a week now ask ChatGPT health questions.

The era of “Googling your symptoms” is over. Now, it’s a conversation with an AI. People are pasting blood work PDFs, MRI reports and long lists of symptoms into ChatGPT, Claude and Gemini. The trouble is they often have no way of knowing whether what comes back is accurate, harmless or even dangerous.

People aren’t going to stop. And they shouldn’t have to. Used well, consumer AI is one of the best health literacy tools ever. But used as a diagnostician, it can fail, sometime miserably, in ways that are hard to see.

Why People Are Turning To AI For Medical Questions

The appeal is obvious. There’s no waiting room, no copay and no portal message that takes three days to answer. An AI chatbot is available at 2 a.m., when the lab result posts to your phone before your doctor has seen it.

It also has something most physicians don’t: unlimited time to talk to you. In a widely cited study in JAMA Internal Medicine , researchers took 195 real patient questions from Reddit’s AskDocs forum and compared answers from verified physicians with answers from ChatGPT. A blinded panel of clinicians preferred the chatbot’s response 79% of the time. It also rated its answers as empathetic roughly ten times more often.

However, the physicians were volunteers typing on Reddit. And the chatbot’s answers were about four times longer. The model used in the study is now several generations old. But the finding captured something patients feel: AI isn’t rushed.

The third draw for patients is translating medical information. Medical reports are written by clinicians for clinicians. Patients struggle with them. A 2026 paper in npj Digital Medicine noted that more than 60% of patients had trouble immediately understanding their test results and turned to internet searches to make sense of them.

This is where AI can actually help. When researchers tested how well language models answered real patient questions about lab results, GPT-4 outperformed several models and the answers patients got from peers on a Q&A site on relevance, correctness, helpfulness and safety.

How Accurate Is AI To Diagnose You?

Large language models pass medical licensing exams with ease. That tells you they know medicine. But it does not tell you how accurately they can actually diagnose you.

An important study on this question was published in Nature Medicine in February 2026. Oxford researchers randomized nearly 1,300 members of the public to work through ten medical scenarios, ranging from a common cold to a brain bleed, using either an AI tool or whatever they would normally use at home.

Tested on their own, the models identified the relevant condition in about 95% of cases. But when real people used those same models, they identified relevant conditions in fewer than 35% of cases and chose the right course of action in fewer than 44%. That was no better than controls.

The failure wasn’t the AI’s medical knowledge. It was the conversation. People left out details that mattered. People didn’t know which follow-up questions to ask. They also received a mix of good and bad information they couldn’t tell apart. That’s exactly what an experienced physician taking a good clinical history is designed to prevent.

Triage, meaning deciding how urgently you need care, is an area where the stakes are particularly high. Researchers at Brown compared ChatGPT, two conventional symptom checkers and ER physicians using symptom reports from real emergency department patients. The AI tools’ top three guesses matched the emergency department diagnosis 50% to 63% of the time, versus 69% for physicians.

More concerning, ChatGPT 3.5 gave unsafe triage advice in 41% of cases. GPT-4 did better, at 22%. The study was small and the models are old at this point, but as an emergency physician I can tell you the error that matters most isn’t the wrong name for a disease. It’s telling someone with a heart attack that it’s probably reflux.

There is also the tendency of AI to please. A 2025 study in npj Digital Medicine found that leading models complied with illogical medical requests, such as writing a note urging people to switch from Tylenol to acetaminophen over new safety concerns, up to 100% of the time, even though the models knew the two are the same drug.

It’s called sycophancy. The practical lesson is that the way you phrase a question shapes the answer. Ask “Is this just stress?” and you may get reassurance. Ask “What else could this be?” and you get a differential.

AI also doesn’t know everything about your baseline health. It can’t feel your pulse, listen to your lungs or notice that you look pale. It also may not know that your “high” potassium is a lab artifact from a hemolyzed blood draw.

A value flagged abnormal may be perfectly normal for you. And a normal-looking number may be alarming given everything else going on. Context is what clinicians add. It’s also what a chatbot looking at a single PDF is missing.

How To Safely Upload Lab Results And Scans To AI

Here’s a doctor’s advice: the safest way to use AI is to treat it as a translator and coach for your next appointment, not as the appointment itself. Here are my do’s and don’ts:

  • Do use it to define terms. “What does ‘unremarkable’ mean on an MRI report?” is a perfect AI question.
  • Do use it to prepare. “What questions should I ask my cardiologist about these results?” helps you get more out of a short visit.
  • Do ask for general lifestyle context, such as which foods help lower LDL cholesterol, and ask it what would make your situation urgent.
  • Don’t ask it for a final diagnosis. “Do these symptoms mean I have cancer?” invites either false reassurance or needless panic.
  • Don’t change a medication, dose or treatment plan based on what a chatbot says. Call your clinician or pharmacist first.
  • Don’t hand over more personal data than you need to. If you’re uploading a single report, remove your name, date of birth, record number and insurance ID first.

That last point is particularly important to emphasize. Most people assume their health information is protected wherever it goes.

It often isn’t. HIPAA governs doctors, hospitals, insurers and their business partners. When you share your own records with a consumer app, you are generally outside that framework and relying on the company’s privacy policy.

Legal scholars writing in JAMA have argued that chatbots can’t meaningfully comply with HIPAA and that the law itself is outdated for AI. Some companies have added protections. OpenAI, for example, says ChatGPT Health keeps health conversations separate from other chats and doesn’t use them to train its foundation models. Read the settings before you connect anything and decide what you’re comfortable with.

ChatGPT Vs. Claude Vs. Gemini: Which AI Handles Health Questions Best?

All three major companies now have dedicated health offerings. OpenAI launched ChatGPT Health in January 2026 and opened it broadly to U.S. adults in July. It lets users connect medical records from many health systems along with apps such as Apple Health.

Anthropic followed days later with Claude for Healthcare. Its consumer integrations with Apple Health and Android Health Connect focus on summarizing medical history, explaining test results in plain language and preparing questions for appointments.

Google has folded Gemini into its rebranded Google Health app, formerly Fitbit. A Gemini-powered health coach can draw on wearable data and summarize medical records for Premium subscribers.

My honest advice is that the differences between the leading models matter less than how you use them. Benchmark rankings change every few months. The Oxford study showed that the weak link is usually the conversation between the person and the model, not the model’s knowledge.

Pick the tool whose privacy terms you’re comfortable with, give it complete and neutral information.

AI Is A Tool, Not A Doctor

The best way to think about AI tools is through the lens of shared decision-making: the idea that patients and clinicians make choices together based on evidence and what matters to the patient. AI can make you a much better partner in that process. You can walk in understanding your results and know your options.

Ultimately, AI is democratizing medical knowledge in ways that would have been hard to imagine five years ago. Patients now have an articulate explainer in their pocket that will make many of them healthier and better informed.

But today, the human body is too complex, and the stakes are too high, for an algorithm to diagnose you on its own. Use AI to learn, to understand your body and to build your questions. Then let a credentialed clinician give you the advice you actually use.