The complexities of patient-centred conversational artificial intelligence.

Conversational AI can offer accessible diagnostic guidance but are often evaluated with simulated patients. This preprint study analyzing 2,053 real patient–chatbot conversations to characterize communication styles and emotional expression. Based on this analysis, the authors developed a patient simulator with distinct patient personas, and tested the ability of LLMs to perform emergency triage with each different persona. They found that different communication styles affected triage reliability and timeliness of care, diminishing the safety of consumer-facing medical AI in real-world patient contexts.