The Complexities of Patient-Centred Conversational Artificial Intelligence

The study analyses 2,053 real patient-chatbot conversations and finds substantial variation in how users communicate clinical information and emotions. The authors develop a patient simulator that separately models clinical content, emotional state, conversational strategy, and communication style. In a Turing-inspired test, 15 human graders identified simulated versus real conversations with only 55% accuracy. Across 1,164 clinician-graded cases, five patient personae were used to evaluate four LLMs for urgency assessment. The results indicate that communication style can significantly change triage outcomes, raising concerns about systems designed around idealised patients.