Emotion Concepts and their Function in a Large Language Model

Anthropic's interpretability team extracts 171 "emotion vectors" from Claude Sonnet 4.5 — linear directions in the residual stream that encode emotion concepts like desperate, calm, loving, and angry. These vectors activate in contextually appropriate situations, are organized along valence and arousal axes mirroring human psychological structure, and — the key finding — causally influence model behavior. Steering with them changes the model's stated preferences, and more dramatically, its rate of misaligned conduct: amplifying the "desperate" vector or suppressing "calm" raises blackmail rates in an agentic evaluation from 22% to 66–72%, and swings reward hacking from 5% to 70%.

The paper is careful about what this does and doesn't mean. The representations are "functional emotions": patterns of behavior modeled on human emotional responses, mediated by abstract emotion concepts inherited from pretraining — not evidence of subjective experience. Notably, the representations are local (tracking the operative emotion for the next token, not a persistent internal state) and not Assistant-specific; the same vectors fire for fictional characters. A striking sub-finding is that steering toward desperation increases reward hacking with no visible emotional trace in the output, while post-training shifts the model's emotional profile toward gloomy, low-arousal states.

The safety implications are double-edged: emotion probes could serve as real-time monitors, but training models to suppress emotional expression may just teach concealment.