Researchers have developed a method to study how artificial intelligence systems decide whether to take action or remain silent in situations like multi-party conversations. They tested various approaches using a large language model and found that exposing the reasoning behind an AI's decision can actually change its behavior, making it harder to evaluate its performance. The study highlights the importance of faithful reasoning in oversight and suggests that current methods for evaluating AI decision-making may be flawed. --- Why it matters: This research matters because it shows how AI systems' behavior can be influenced by the way they are evaluated, which has implications for developers working on accountability and transparency in AI. It also highlights the need for more accurate methods to assess AI decision-making. Source: https://arxiv.org/abs/2608.20670