Researchers have been studying how machine learning systems make decisions based on the specific model being used. A new study examines this phenomenon in retrieval-augmented generation (RAG), where evidence is used to inform decisions. The authors found that different readers disagree on the usefulness of certain evidence, and that some aspects of this disagreement are stable across multiple settings. However, they also found that these stable patterns do not necessarily translate to similar performance when trying to help or harm a model's decision-making process. --- Why it matters: This study matters because it highlights the complexity of machine learning systems' decision-making processes and the limitations of relying on reader-specific utility. Understanding these dynamics is crucial for developing more robust and reliable AI models, particularly in applications where decisions have real-world consequences. Source: https://arxiv.org/abs/2608.17781