Why You Shouldn’t Rely on ChatGPT Health in an Emergency (And What to Do Instead)
What a New Study Reveals About ChatGPT Health and Medical Emergencies
When you or someone you love suddenly feels very sick, it’s natural to reach for your phone and ask an AI chatbot what to do. With the launch of ChatGPT Health, OpenAI’s new health-focused assistant, many people hoped they’d have a “doctor in their pocket.” But a new study in Nature Medicine, covered by NBC News, found that ChatGPT Health frequently underestimated how serious medical emergencies were — what researchers call “under-triage.”
In simulated emergency scenarios, ChatGPT Health often recommended a lower level of care than trained clinicians did. That doesn’t mean AI is useless for health questions, but it does mean this: in an emergency, AI should never replace real-world medical judgment or emergency services.
Below, we’ll break down what the study actually found, what “under-triage” means for your safety, and how to use tools like ChatGPT Health wisely — without putting yourself at risk.
The Core Problem: Under-Triaging Medical Emergencies
Triage is the process of deciding how urgent a medical situation is and what level of care is needed. In the Nature Medicine study, researchers fed a wide range of emergency cases into ChatGPT Health — from chest pain to breathing difficulties and severe infections — and compared its advice to that of human clinicians.
- Under-triage means the AI said the situation was less urgent than experts thought.
- Over-triage means the AI suggested a higher urgency than necessary (for example, telling you to go to the ER when urgent care would do).
According to NBC News’ reporting on the study, ChatGPT Health under-triaged roughly half of the simulated emergencies. That’s concerning, because in real life, downplaying an emergency can delay life-saving treatment.
“From a patient-safety standpoint, under-triage is what scares us most. Over-triage is inconvenient; under-triage can be dangerous.”
— Emergency physician commenting on AI triage tools
Importantly, this was a controlled, simulated study — not a real-world rollout. But it highlights why AI health tools must be treated as information companions, not emergency decision-makers.
What the Nature Medicine Study Looked At
The researchers designed detailed case vignettes, similar to what you might describe when you type symptoms into a chatbot. These covered a spectrum:
- Clear emergencies (e.g., classic heart attack symptoms, stroke warning signs).
- Urgent but less obvious issues (e.g., worsening shortness of breath in someone with asthma).
- Non-urgent problems (e.g., mild rashes, minor injuries).
Clinicians independently rated how urgent each case was and what action they’d recommend: call emergency services, go to the emergency department, visit urgent care, call a primary care doctor, or manage at home with self-care.
ChatGPT Health then triaged the same cases. Compared with clinicians, it:
- Was sometimes appropriately cautious.
- Occasionally overreacted, steering cases to the ER unnecessarily.
- But too often suggested less urgent care than clinicians deemed safe.
Why Under-Triage by AI Can Be Risky
Under-triage is dangerous because many serious conditions worsen silently at first. A few examples:
- Heart attack: Not everyone has “Hollywood” chest-clutching pain. Some only have fatigue, nausea, or mild chest pressure.
- Stroke: Subtle slurred speech or slight weakness could be dismissed as stress or fatigue.
- Sepsis (severe infection): Early symptoms can look like a regular flu until blood pressure suddenly drops.
A human clinician is trained to spot the pattern and err on the side of caution when a patient’s story “doesn’t feel right.” AI systems, including ChatGPT Health, work differently: they rely on patterns in data they were trained on and the words you provide. If your description is incomplete, or if the model hasn’t seen enough similar emergency cases, it may not fully appreciate the risk.
“AI can be impressively knowledgeable, but it doesn’t see your vital signs, your skin color, or how distressed you look. Those are often the clues that push clinicians to say, ‘We need to act now.’”
— Critical care specialist, academic medical center
When AI Health Tools Like ChatGPT Health Can Still Be Helpful
None of this means you should never use ChatGPT Health or similar tools. They can be genuinely useful when used for the right purposes.
Appropriate uses for AI health assistants
- Learning about chronic conditions (e.g., diabetes, hypertension, asthma).
- Understanding test results or medical jargon you’ve already received.
- Preparing questions to ask your doctor before an appointment.
- Exploring lifestyle changes for sleep, nutrition, or stress management.
- Finding reputable sources, guidelines, or patient organizations.
Used this way, AI can help you feel more informed and confident in conversations with your health-care team. It should be a supplement to, not a replacement for, professional care.
When You Should Not Rely on ChatGPT Health
Regardless of how advanced AI becomes, there are situations where you should skip the chatbot and go straight to emergency services or in-person care. If you or someone else has any of the following, call your local emergency number (such as 911 in the U.S.) or go to the nearest emergency department:
- Chest pain or pressure, especially if it spreads to the arm, jaw, or back.
- Sudden difficulty breathing or feeling like you can’t get enough air.
- Sudden weakness, numbness, facial drooping, confusion, or trouble speaking (possible stroke).
- Severe or worsening abdominal pain, especially with fever, vomiting, or faintness.
- Heavy, uncontrolled bleeding.
- Thoughts of self-harm or harming others, or a mental health crisis.
- Serious trauma (e.g., car crash, major fall, head injury).
In these moments, every minute counts. Asking a chatbot first can delay critical treatment. Use AI afterward to understand what happened — not to decide whether to get help in the first place.
Practical Steps: Using ChatGPT Health Safely and Wisely
If you choose to use ChatGPT Health or similar tools, here’s a simple, practical framework:
1. Decide your goal before you open the app
Ask yourself: “Am I curious or am I scared?” If you’re scared something might be seriously wrong right now, that’s a sign to call a clinician or emergency services instead of starting with AI.
2. Use AI for information, not for final decisions
Treat any AI recommendation as a draft. You might:
- Write down the AI’s explanation of your condition.
- Bring it to your appointment or send it through a patient portal.
- Ask your clinician, “Does this match what you’re seeing in me?”
3. Be specific and honest when you describe symptoms
If you do ask AI about symptoms, give detailed information: timing, severity, what makes it better or worse, and any relevant medical history. Vague prompts tend to produce vague (and sometimes falsely reassuring) answers.
4. Cross-check with trusted sources
After getting an AI answer, cross-check anything important with:
- Major academic medical centers (e.g., Mayo Clinic, Cleveland Clinic).
- Government health agencies (e.g., CDC, NIH, WHO).
- National specialty societies (e.g., American Heart Association).
Common Obstacles: Why People Turn to AI in a Crisis
It’s easy to say “just call your doctor,” but the reality is more complicated. Many people use AI in emergencies because:
- They don’t have easy access to a primary care clinician.
- They’re worried about the cost of an ER or urgent care visit.
- They feel embarrassed or afraid of “overreacting.”
- They’re caring for someone at home and feel alone with the decision.
These are real, valid concerns. The risk is that AI can offer a comforting, confident-sounding answer that nudges you toward waiting “just a little longer” — exactly what under-triage looks like in real life.
A Realistic Scenario: How AI Advice Can Go Wrong
Consider a composite example based on common ER cases (details changed to protect privacy):
A 54-year-old person feels vague chest discomfort and tiredness after dinner. They type into an AI chatbot: “Feeling a bit of chest tightness and heartburn after eating — maybe anxiety?” The chatbot, not wanting to alarm the user and interpreting the description as mild, lists possibilities including reflux and anxiety, and suggests trying antacids and monitoring symptoms, with a general note to seek help if things worsen.
Hours later, the pain intensifies, and they finally go to the ER. Tests show they were having a heart attack. Fortunately, they survive — but they needed more aggressive treatment than if they’d come in sooner. A human clinician hearing the same initial story might have asked more probing questions and urged them to seek urgent care right away.
This type of “seems mild at first” presentation is exactly where under-triage occurs — and where relying on an AI chatbot can delay critical care.
What This Study Means for the Future of AI in Health Care
The Nature Medicine findings and NBC News coverage don’t mean we should abandon AI in medicine. Instead, they highlight where guardrails are needed:
- Clear warnings: AI tools should explicitly state that they’re not emergency services and may under- or overestimate risk.
- Safety-first design: When in doubt, systems should lean toward caution for potentially serious symptoms.
- Integrated care: The most promising future is AI embedded within health systems, where clinicians can supervise its use.
- Ongoing auditing: Studies like this must be repeated as models evolve, to track real-world safety performance.
For now, the safest mindset is to see AI as a powerful library and translator, not as a doctor, nurse, or paramedic.
Key Takeaways to Keep You and Your Family Safe
To sum up what this new research means for your day-to-day life:
- ChatGPT Health under-triaged many simulated emergencies. It sometimes advised less urgent care than clinicians thought was safe.
- AI can help you learn, but not diagnose emergencies. Use it for education, not for deciding whether to call 911.
- When something feels seriously wrong, trust your instincts. Call emergency services or seek in-person care — even if a chatbot sounds reassuring.
- Prepare ahead of time. Know your local emergency number, nearest ER, and urgent care options, and consider nurse advice lines or telehealth services.
Moving Forward: Pair AI with Human Care, Not Instead of It
AI tools like ChatGPT Health are here to stay, and they will likely keep improving. They can make medical information easier to understand, help you prepare for appointments, and empower you to ask better questions. But this new study is a timely reminder that polished language doesn’t equal perfect judgment — especially when seconds matter.
As you navigate your health journey, consider using AI as a knowledgeable friend at the library, not the person driving the ambulance. In moments of uncertainty or fear, reach out to trained professionals who can see you, examine you, and act quickly if needed.
Your next step: Take five minutes today to save your local emergency number in your phone favorites, look up the closest emergency department and urgent care, and note any nurse advice lines on your insurance card. Those small preparations will serve you far better in a crisis than any chatbot ever could.
