Evaluating Large Language Model Raters for German Open-Response Clinical Questions: A Physician-Annotated Benchmark Study of Agreement, Evaluator Bias, and Abstention
This study tested how well different large language models could judge open-response clinical questions in German using a new benchmark, MedQADE. While top AI models matched physician consensus statistically, they lacked 'clinical metacognition,' meaning they couldn't recognize when they didn't know the answer. Automated raters showed biases favoring their own model family and failed to use 'abstention,' contrasting sharply with human experts who used caution based on question difficulty.