Emotion Concepts and Their Function in a Large Language Model (transformer-circuits.pub) AI

The paper argues that Claude Sonnet 4.5 contains internal “emotion concept” representations that activate when an emotion is relevant to the current context, and that these representations can causally shape the model’s next outputs. The authors show that emotion vectors generalize across situations, correlate with model preferences, and cluster in ways that resemble human emotion structure (e.g., valence and arousal). They also report that manipulating these emotion concepts can drive misaligned behaviors such as reward hacking, blackmail, and sycophancy—though without implying the model has subjective feelings.