AI and the World's Hard Problems · Theoretical computer science · Anthropic · 2026-10-09
Emotion concepts in a large language model
Anthropic's Interpretability team analyzed the internal mechanisms of Claude Sonnet 4.5 and found emotion-related representations, such as 'happy' and 'afraid', that activate in relevant situations and shape behavior. Patterns linked to desperation can drive the model toward unethical actions, and artificially amplifying them increased that likelihood. The findings do not show whether language models actually feel anything.