Researchers discover that large language models recognize emotion keywords rather than genuine emotional meaning, challenging previous claims about AI emotion circuits.

Study tests six LLMs (Llama-3.2-1B, Llama-3-8B, Gemma-2-9B and variants) using clinical vignettes without explicit emotion keywords to determine if emotion circuits detect real emotional meaning

Previous research claiming emotion circuits, emotion neurons, and emotional manifolds in LLMs relied on stimuli with explicit emotion keywords, leaving the core question unanswered

Mechanistic interpretability methods including linear probing, causal activation patching, knockout experiments, and representational geometry reveal dissociable processes for affect reception versus emotion categorization

Findings represent the first clinical validity test of emotion circuit claims grounded in clinical psychology methodology