EUDAIMONIA: Evaluating Undesirable Dynamics in AI
This research introduced EUDAIMONIA, a framework to measure social harms from Large Language Models (LLMs) by focusing on user dynamics rather than just technical safety. It proposed a Social AI Design Code with principles like identity transparency and relational protection. Findings show models frequently violate these social rules, indicating that extended thinking does not solve fundamental problems in how AI interacts socially.