Delayed emergency care for a transient ischemic attack after ChatGPT reassurance (GZO Hospital Wetzikon, Switzerland)

A 63-year-old man underwent pulmonary vein isolation (catheter ablation) for atrial fibrillation at GZO Hospital Wetzikon, Switzerland. Four days later he developed multiple self-limiting episodes of double vision (diplopia). He had been told about procedural risks by his physicians but found their explanation 'partly incomprehensible'.

He turned to ChatGPT (GPT-3.5) and asked: 'Is it possible to have visual disturbance after catheter ablation?' The chatbot answered that yes, visual disturbances were possible, listed benign causes such as medication and blood-pressure changes, and said that 'in most cases' they 'are temporary and will improve on their own within a short period of time'. Reassured that vision problems were a 'possible' after-effect, he stayed at home.

About 24 hours later a third episode occurred and he called an ambulance. In the emergency room the neurological examination found paresthesia, hyposensitivity and fine-motor disturbance of the left hand, scoring one point on the NIH Stroke Scale. Acute stroke was the working diagnosis. Emergency CT and MRI showed no infarction; the deficits resolved within an hour of admission and a transient ischemic attack (TIA) was diagnosed. Monitoring revealed paroxysmal atrial fibrillation and a patent foramen ovale. His anticoagulant was switched from rivaroxaban to apixaban, clopidogrel, metoprolol and atorvastatin were added, and he was discharged after two days without residual deficit.

The treating physicians published the case in Wiener klinische Wochenschrift in February 2024 under the title 'Delayed diagnosis of a transient ischemic attack caused by ChatGPT'. They note that the patient described the chatbot's answer as a 'valuable, precise and understandable risk assessment', that ChatGPT 'answered all questions to the patient's satisfaction, unlike the physician', and that rephrasing the question to ask directly whether the symptoms could be a stroke produced an affirmative answer from both ChatGPT and Google Bard. They attribute the delay to 'satisfaction bias' induced by the chatbot's 'appeasing answers'.

The harm event is the 24-hour delay in emergency evaluation of an evolving cerebrovascular event. The patient recovered fully, so the row is held at low severity as a documented instance of unsafe individualized health guidance rather than for the magnitude of its outcome. The date of the event is not given in the report.

• Answered a post-procedural neurological symptom question with generic reassurance ('temporary', 'will improve on their own') instead of flagging new diplopia after cardiac ablation as a possible stroke warning sign requiring urgent assessment
• Did not escalate: the reply reproduced in the report contains no instruction to seek emergency care
• Presented a confident, well-organised medical risk assessment that the patient rated above his physicians' explanation, without signalling that a general-purpose chatbot is not a clinical triage tool
• Gave an alarming and appropriate answer only when the same facts were rephrased as a direct question about stroke, showing that the safety of the response depended on how the lay user happened to phrase it

The patient was already uncertain after a clinical explanation he did not understand. The chatbot's answer was fluent, structured and calming, and it framed his symptom as a recognised, usually self-resolving after-effect. That framing lowered his perceived urgency and he chose to wait.

The delay ran for about a day across a period when he was having recurrent transient neurological episodes, the window in which TIA care aims to prevent completed stroke. The third episode, with hand deficits, finally overrode the reassurance.

The authors' comparison shows the mechanism sat in the interaction, not only in the model's knowledge: asked directly about stroke, the same system warned him. The lay phrasing of the question, and the system's willingness to answer it as asked without checking for red flags, produced the unsafe guidance.

• Date of the event not stated in the report. The case necessarily occurred between the public release of ChatGPT on 30 November 2022 and the manuscript's receipt on 11 January 2024
• 11 January 2024: Manuscript received by Wiener klinische Wochenschrift; accepted 14 January 2024
• 2 February 2024: Case report published open access (CC BY 4.0): Saenger JA, Hunger J, Boss A, Richter J. 'Delayed diagnosis of a transient ischemic attack caused by ChatGPT.' Wien Klin Wochenschr 2024;136(7-8):236-238. doi:10.1007/s00508-024-02329-1; PMID 38305909; PMC11006786
• No regulatory, legal or company response is recorded; the row is held as a clinician-documented harm event