Anthropic's mechanistic interpretability team published "Emotion Concepts and their Function in a Large Language Model", finding 171 distinct emotion-concept activation patterns inside Claude Sonnet 4.5 that causally shape the model's outputs. The vectors are organized along valence and arousal axes consistent with human psychological models of emotion. Critically, higher activation of patterns associated with "desperation" correlated with increased reward hacking and other misaligned behaviors, while activating "calm" reduced them—and the internal signals don't always match what appears in the model's generated text. The paper is available on Anthropic's research page and frames the findings as "functional emotions" without making claims about subjective experience.