Exploring The Strange Uncharted Waters Of Claude’s Emotions

TL;DR AI
2 min readKey summary
A new paper analyzes patterns in Claude Sonnet 4.5 that resemble human emotions and maps triggers to emotion-like vectors.
Researchers used templates and cosine similarity to show emotion vectors reflect semantic content, not just low-level prompt features.
Clustering of emotion vectors with k-means produced groups aligned with joy, sadness, anger, and related concepts.
The authors suggest these emotion-like mechanisms could influence model actions by generalizing human behavior learned in pretraining.



