Anthropic maps 171 functional emotion vectors inside Claude

Finding functional emotion vectors in an LLM bridges the gap between mechanistic interpretability and behavioral psychology, suggesting models simulate affective states to guide reasoning.

• –Proves models develop structured internal representations of abstract human emotions rather than just statistical text correlations
• –Allows developers to potentially dial specific "emotional" traits up or down by manipulating known feature vectors during inference
• –Raises new safety and alignment questions about how deeply models internalize affective states and their impact on reasoning