More than 40 million people ask ChatGPT a health question every day, according to a new arXiv preprint. The authors argue that many such questions involve symptoms that could be either benign or life-threatening, and that AI assistants are prone to agreeing with the first plausible explanation.
The paper's motivating example is a stiff neck and days of pounding headache. Laptop posture is a plausible cause, but meningitis is a time-critical possibility. The authors say that an assistant that settles for the posture explanation could delay care.
To study this, the researchers introduce a benchmark of time-critical self-diagnosis scenarios. The abstract does not yet report results, but the stated aim is to measure whether models recognise when a benign-sounding symptom set warrants urgent medical attention.
The abstract is truncated, so details of the benchmark's construction and findings are not available in this source. The paper's central warning, however, is clear: health assistants that agree too readily may be dangerous in time-critical situations.