Researchers have developed a mechanism called runtime action interference (RAI) to control AI behavior in real-time. RAI regulates the pace and content of actions taken by an AI agent, preventing it from engaging in toxic behaviors like harassment. The team implemented RAI in a replication of AlphaStar's actor.py code and conducted a human participant study using StarCraft II. They found that when users were aware of the AI's capability, they perceived lower fairness and higher toxicity compared to when the capability was withheld. This suggests that humans' perceptions of an opponent controlled by RAI can vary significantly depending on the information presented about its capabilities. --- Why it matters: This study matters because it highlights the importance of separating control mechanisms from capability disclosure in human-AI interactions. Understanding how users perceive AI behavior is crucial for developing more transparent and fair AI systems, particularly in applications like gaming or autonomous vehicles where humans interact with AI agents. Source: https://arxiv.org/abs/2608.21398