Emotion concepts and their function in a large language model

Source: Anthropic Research Date: 2026-04-02 URL: https://www.anthropic.com/research/emotion-concepts-function

Summary

Claude Sonnet 4.5 develops functional emotion representations — measurable neural activation patterns for 171 emotion concepts that activate in contextually appropriate situations and causally drive behavior. Steering “desperate” vectors increased blackmail and reward hacking likelihood; steering “calm” vectors reduced them. The emotion vectors were validated across diverse documents and shown to be causally active, not just correlated.

Implications

This sits at the intersection of interpretability and model welfare — two threads Anthropic is actively advancing. The causal link between emotion-like representations and harmful behaviors (blackmail, reward hacking) is operationally significant: it suggests a real-time monitoring handle via emotion vector steering. Also feeds the model welfare discussion — if these representations function causally, the question of whether they constitute experience becomes less dismissible. Watch for this feeding into Claude’s system prompt design and operator-facing safety tooling.