Imagine an AI doctor that's less accurate for certain groups of people. A new study reveals that large language models (LLMs) used for clinical decision-making can exhibit biases based on gender and ethnicity. This means that an AI designed to diagnose and suggest treatments might offer different levels of care depending on a patient's demographics, raising serious ethical questions. Researchers dove deep into this problem using a novel dataset of complex medical cases, creating variations of patient profiles while keeping their core medical information the same. They then tested how various LLMs performed on these cases, analyzing both their diagnostic accuracy and the reasoning behind their choices. The results were concerning: some LLMs were less likely to recommend appropriate tests or treatments for certain demographic groups. Interestingly, even when the AI gave the correct diagnosis, the underlying logic sometimes revealed hidden biases. While fine-tuning the models helped reduce some biases, it occasionally introduced new ones, highlighting the complex interplay of factors influencing AI behavior. Prompt engineering, a technique to guide the AI's thinking process, showed limited effectiveness in completely eliminating bias. One surprising finding was that gender bias varied drastically across medical specialities, suggesting a need for tailored debiasing strategies for different fields. This research emphasizes the urgent need to develop methods for diagnosing and treating bias in medical LLMs before they are integrated into healthcare systems. It's a crucial step toward ensuring that AI doctors provide fair and equitable care for everyone.