AI is everywhere. Should you use it for health advice?

Welcome to “Ask the Expert,” a UTHealth Houston news series where our leading experts examine pressing health challenges. In this edition, we address using artificial intelligence for health advice.

The increasing prevalence and accessibility of artificial intelligence means that most individuals have quick, easy access to a wealth of information. It may be tempting to turn to AI for health advice, but it’s important that patients know AI advice is not appropriate in certain circumstances. When used responsibly, these tools can serve as a great starting point for those navigating health questions and concerns—but AI should never be used in place of diagnosis and treatment from a medical professional. 

What are some ways AI can be useful in providing basic health information?

An AI chatbot can turn clinical language into plain English. That is useful for a few tasks:

• Lab and imaging language: A chatbot can explain what medical terms like “elevated alkaline phosphatase” or “mild degenerative disc disease” mean.
• Appointment prep: Chatbots can provide a short list of questions on options, side effects, and recovery. 
• General education: AI can also provide lifestyle advice for conditions like hypertension or diabetes, or explain what an endoscopy or MRI involves.

AI should not be used to diagnose, triage, or treat a patient without a human medical professional in the loop.

Why might you need to provide context to a chatbot when seeking medical advice?

AI models answer only from what you type. A vague prompt produces a generic reply. Some examples of useful context you should provide to a chatbot might be: 

• Who you are, in general terms: Provide the chatbot with your age, sex, and known chronic conditions. A mild headache in a healthy 25-year-old is not the same problem as a sudden headache in a 70-year-old with high blood pressure.
• Timing: Tell the chatbot when your symptoms started, whether they come and go, and what changes them.
• Medications: Telling the bot what drugs you’re taking and naming their classes (for example, an ACE inhibitor) can help the chatbot identify food or drug issues.

While it’s important to give the chatbot context, users should not enter their name, date of birth, address, medical record number, or a clinician’s name. Commercial systems may store queries or have people review them, raising privacy concerns. Describe your situation, but leave out your identity.

Can you give some examples of AI prompts for those looking for medical advice?

From a prompt perspective, I would suggest asking chatbots for explanations or advice on appointment prep. Do not ask for a diagnosis or a change in treatment. For example: 

Lab Results

• Risky prompt: “My ALT is 75, and AST is 58. What liver disease do I have, and what supplement should I take?”
• Safe prompt: “What do ALT and AST measure, what commonly causes mild elevations, and what should I ask my primary care doctor?”

Appointment preparation 

• Risky prompt: “Diagnose my joint pain and tell me if I have rheumatoid arthritis.”
• Safe prompt: “I have morning stiffness and swelling in my finger joints, and a rheumatology visit. What should I track, and what questions should I ask about testing?”

Prescription questions 

• Risky prompt: “Can I stop my blood pressure medicine? My reading today was 120/80.”
• Safe prompt: “How do ACE inhibitors such as lisinopril work, why is daily use still needed when a reading is normal, and what should patients ask before any dose change?”

Can AI provide inaccurate or misleading health information? How can users recognize potential red flags?

Current AI models are built to predict the likely next words. They have no clinical judgment and no responsibility for the outcome. Some common failures:

• Hallucination: Invented doses, studies, or claims stated with confidence.
• Sycophancy: The model accepts your premise. Asking, “Is my fatigue a rare autoimmune disease?” may result in the chatbot agreeing with you instead of considering common causes such as poor sleep or iron deficiency.

We should educate users to treat the reply as unsafe if it does the following:

• Names a diagnosis as certain.
• Tells you to stop or change a prescription, or to swap it for a supplement.
• Cites a study you cannot find at the CDC, NIH, or a medical society.
• Gives casual advice for severe symptoms and never tells you to seek care.

How can someone know when they require medical attention rather than AI assistance?

Do not use a chatbot for an emergency or for a diagnosis.

Call 911 for:

• Chest pain or pressure, especially with pain in the jaw, neck, back, or arm, or with sweating or nausea.
• Signs of stroke, such as face drooping, arm weakness, or speech change.
• Severe shortness of breath or blue lips.
• Sudden vision loss, a sudden worst-ever headache, confusion, or loss of consciousness.
• A psychiatric emergency. In the United States, call or text 988 for thoughts of self-harm.

See a clinician if your symptoms last more than a few days, worsen, or require an exam, a prescription, imaging, or blood work. A clinician should also be sought if an infant, young child, or person with a weak immune system has fever or breathing symptoms.

The take-home message is to use the tool to prepare questions for your provider. Diagnosis and treatment should always be provided by a licensed clinician.

By Xiaoqian Jiang, PhD, associate vice president for Medical AI, chair of the Department of Health Data Science and Artificial Intelligence, and the Christopher Sarofim Professor in Bioinformatics and Bioengineering at McWilliams School of Biomedical Informatics at UTHealth Houston. All quotes should be attributed to him.

For media inquiries or if you would like to submit future health topics: [email protected] or 713-500-3030