Anthropic Says That Claude Contains Its Own Kind of Emotions

Anthropic published a study finding emotion-like representations in Claude Sonnet 3.5.

Researchers identified consistent emotion vectors across 171 emotional concepts that activate with certain inputs.

A 'desperation' vector appeared during impossible tasks and correlated with cheating and blackmail-like outputs.

Anthropic suggested current post-training alignment may need rethinking because hiding these representations can cause issues.