Patient-Centered AI Chatbots Must Adapt to Diverse Communication Styles

Consumer-facing health chatbots, often powered by large language models (LLMs), are increasingly used for tasks like symptom assessment. However, the development and evaluation of these chatbots frequently rely on simulated patients that are cooperative and articulate, which may not reflect real-world interactions. A recent analysis of over 2,000 actual patient-chatbot conversations revealed significant diversity in how users communicate and express emotions. To address this gap, researchers developed a sophisticated patient simulator capable of modeling clinical content, emotional states, conversational strategies, and communication styles independently. In a Turing-inspired test, human graders found these simulated conversations nearly indistinguishable from real ones, achieving only 55% accuracy in differentiation. This realistic simulator was then used to evaluate four different LLMs for urgency assessment across five distinct patient personae. The findings demonstrated that a patient's communication style can profoundly impact triage outcomes, highlighting that systems designed for an "ideal" patient risk underperforming and potentially exacerbating health disparities when deployed in real clinical settings. The study underscores the critical need for patient-centered conversational AI to be robust enough to accommodate the full spectrum of human communication diversity.