Can ChatGPT recognise urgent medical needs?
A study finds the health-focused chatbot often underestimates emergencies
Researchers tested a health‑trained conversational AI and found it frequently judged serious scenarios as less urgent than they were. In clinical vignettes designed to simulate emergencies, the system under‑triaged about half of cases that would typically require urgent assessment or immediate care. That means people could receive reassurance when they actually need rapid evaluation.
The scale of the problem is significant because usage is already widespread. Tens to hundreds of millions of people consult large‑language models for health questions each week, and specialised “health” versions of those models are being positioned as accessible triage or informational tools. When a system systematically underestimates severity, the most immediate consequence is delay: patients who are told their problem is minor may avoid or postpone calling emergency services, lose valuable treatment time, or attempt ineffective self‑care.
Key practical points
- Errors occurred across a range of presentations, including some that require urgent in‑person assessment. The model was more likely to miss red‑flag signs than to overcall severity.
- The study highlights the current limits of AI for safety‑critical applications: conversational fluency does not equal reliable clinical judgement.
- Developers and regulators will need to embed clear, conservative triage thresholds, explicit safety prompts, and fail‑safe routing to human clinicians for any red‑flag symptoms.
What to do now
If symptoms are sudden, severe, or suggest life‑threatening illness — chest pain, sudden severe breathlessness, altered consciousness, severe bleeding, or rapidly worsening neurological signs — seeking emergency care immediately remains the safest option. AI tools can help with basic information, but they are not a substitute for clinical assessment in situations where lives are at stake.