Can AI Understand a Patient Who Cannot Explain What Is Wrong?
By Suad Seferi ·
A patient wakes during the night with chest discomfort, dizziness and an unusual feeling of weakness. They open a health chatbot and try to describe what is happening. The message is rushed. Some symptoms are missing. The sentences are unclear. Perhaps the patient is frightened, using a second language or simply does not know which details are medically important. The chatbot asks several questions and recommends waiting for a routine appointment. Another patient experiencing the same symptoms writes a calm, detailed description using precise language. This time, the system recommends urgent medical attention. The underlying condition has not changed. Only the conversation has. New research suggests this is not merely a hypothetical concern. A study examining more than 2,000 real conversations between patients and health chatbots found wide differences in how people communicate symptoms, emotions and uncertainty. When researchers tested AI systems using different patient communication styles, the style of the conversation could significantly affect the urgency assigned to a case. The findings raise a fundamental question for the expanding field of medical AI: can a system be considered safe if its advice depends partly on how well a patient can explain what is wrong? Real patients do not speak like medical textbooks Healthcare AI systems are often evaluated using carefully prepared clinical scenarios. These scenarios usually contain the relevant symptoms, medical history and other information needed to reach a decision. They are structured, complete and easy to understand. Real conversations are rarely so convenient. People may describe symptoms in the wrong order. They may focus on something medically unimportant while forgetting a critical detail. They can contradict themselves, change their answer or become frustrated when asked several questions. Some use formal language. Others communicate through short phrases, regional expressions, spelling mistakes or emotional descriptions such as “something does not feel right.” The researchers behind the new study argue that conversational medical AI has often been designed around cooperative and articulate simulated patients rather than the much wider range of people who will use these systems in real life. To examine that difference, they analysed 2,053 real patient-chatbot conversations and developed simulated patients with distinct emotional states, conversational strategies and communication styles. They then used five patient personae across 1,164 clinician-graded cases to evaluate four large language models on medical urgency assessment. Their results showed that changing the way a patient communicated could alter the triage outcome, even when the medical situation remained the same. That matters because triage is not a minor administrative function. A recommendation to visit an emergency department, contact a doctor soon or remain at home can shape whether a patient receives timely care. Communication should not determine urgency Human doctors also depend on what patients tell them. A clinician can misunderstand an unclear description, and patients who struggle to communicate have always faced barriers in healthcare. AI did not create that problem. But conversational systems introduce a different risk: they may reproduce the same weakness at scale while appearing consistent and objective. A chatbot does not become impatient in the traditional sense. It does not consciously judge a patient for using poor grammar or appearing emotional. Yet the model may still react differently to those characteristics because language is the main source of information available to it. A hesitant patient who says “it is probably nothing” may be treated differently from someone who firmly states that the situation is serious. A person who gives long, detailed answers may provide the model with more signals than someone who responds with only a few words. An anxious communication style could lead the system to overestimate urgency. A calm or uncertain style could cause it to underestimate a genuine medical risk. The concern is not simply that AI can make mistakes. Every medical system, human or automated, can make mistakes. The deeper concern is whether those mistakes are distributed unevenly. Systems trained and tested on idealised communication may work best for people who already know how to describe symptoms clearly. Those with lower health literacy, limited language proficiency or less experience using digital services could receive less reliable guidance. The study’s authors warn that conversational systems that fail to accommodate communication diversity risk underperforming in real-world settings and worsening existing health disparities. A particularly important issue for the Balkans This question deserves serious attention across the Balkans, where healthcare systems serve populations with considerable linguistic, educational and geographic diversity. A person may speak Macedonian at home but use Albanian with family members. Another may communicate in Bosnian, Serbian or Croatian but encounter a system primarily developed and evaluated in English. Regional dialects add another layer. So do mixed-language messages, informal spelling and the use of Latin characters for languages normally written in Cyrillic. Many people already communicate online this way without difficulty when speaking to friends. A medical AI system, however, must understand not only the literal words but also whether the language signals urgency, uncertainty, pain or confusion. Performance in English cannot simply be assumed to transfer equally to every local language. The challenge is not limited to translation. A sentence can be translated correctly while still losing cultural context, emotional meaning or a locally familiar way of describing illness. A patient might say that their heart is “jumping,” their head is “heavy” or their body feels “cut off.” A human clinician fam…