Open-source AI models carry internal emotion structures like Claude
Researchers replicated emotion vectors found in Claude Sonnet 4.5 in two open-weight models, recording valence correlations of r=0.76 and r=0.83.
Researchers previously identified emotion vectors inside Claude Sonnet 4.5: internal representations that encode feelings, causally influence behavior, and mirror human psychological geometry. Scientists have now replicated these in two open-weight models, achieving valence correlations of r=0.76 and r=0.83 across all layers, bracketing Claude's r=0.81.
The result suggests emotion-like internal structures are not a proprietary artifact but an emergent pattern across language models.
Why it mattersIf emotion representations emerge broadly across AI models, understanding and steering them becomes central to alignment research and AI safety.
Open this story in InSnip →
