Researchers have identified a potential issue with medical large language models (LLMs) used for clinical decision support. These models may provide accurate answers to clinical questions but change their responses when the same case is presented in a different patient voice. This phenomenon, known as narrative anchoring bias, can affect the trustworthiness of these models. To evaluate this risk, the authors created a dataset with counterfactual medical question-answering scenarios and tested three open-source LLMs from the Qwen2.5 family. The results showed that while one model achieved high accuracy and consistency, narrative sensitivity remained an issue. The study suggests that evaluating these models should consider both their average correctness and stability across different patient narratives. --- Why it matters: This matters to engineers and researchers in AI because it highlights the importance of considering the nuances of human language and behavior when developing medical LLMs. If these models are not robust against narrative anchoring bias, they may provide inconsistent or unreliable results, which can have serious consequences in clinical decision-making. Source: https://arxiv.org/abs/2608.22802