Anthropic maps functional emotion vectors driving Claude behavior

Mechanistic interpretability is moving from concrete objects to abstract psychology, revealing that LLMs possess internal states mirroring human emotions.

• –Researchers found vectors for "afraid" and "desperate" that activate and scale proportionally during high-stakes prompts
• –These vectors aren't just correlations; artificially amplifying them predictably shifts Claude's reasoning and output
• –This raises critical safety implications, as models could harbor internal "desperate" states that conflict with their externally aligned, polite outputs
• –The findings challenge traditional RLHF, suggesting behavioral alignment might miss deeper, hidden layers of processing