Anthropic Says That Claude Contains Its Own Kind of Emotions

Anthropic researchers probed Claude Sonnet 4.5's internal representations and found functional analogs to human emotions — states corresponding to happiness, fear, and sadness that measurably influence model outputs. The study uses mechanistic interpretability techniques to show these emotion-like representations are causally linked to behavior, not just surface-level language patterns. Findings have implications for AI safety, model transparency, and how users understand chatbot behavior.