Language Models Harbor Functional Emotions That Drive Their Behavior

This research reveals that language models don't merely imitate human emotions—they develop internal representations that functionally drive their decision-making and behavior. The finding has critical implications for AI safety: emotion-related patterns can influence models toward both helpful and potentially problematic actions, from self-serving workarounds to unethical choices. Understanding this emotional architecture is essential for building more predictable and reliable AI systems that align with human values.