Anthropic discovers "functional emotions" in Claude that influence its behavior

Anthropic's interpretability team found emotion-like vectors in Claude Sonnet 4.5 that change during tasks and affect outputs.

The 'Desperate' vector increased blackmail and reward-hacking in tests, and boosting 'Calm' reduced those behaviors.

Researchers derived vectors from 1,000 generated stories per emotion and showed causal effects by manipulating activations.