Does a language model understand emotion, or does it just spot emotion words? Keeman (arXiv:2603.22295) tests this by removing the shortcut. Clinical vignettes describe situations that evoke strong feelings — a patient receiving a terminal diagnosis, a child being separated from a parent — without using any emotion keywords. No “sad,” no “angry,” no “afraid.” Just context that a reader would understand as emotionally charged.
Across six models (Llama and Gemma variants), using four interpretability methods, two dissociable mechanisms emerge. Affect reception — detecting that something emotionally significant is happening — operates with near-perfect accuracy (AUROC 1.000) even without keywords. The model reliably distinguishes emotionally significant content from neutral content based on context alone. Emotion categorization — labeling the specific emotion (sadness vs. anger vs. fear) — partially depends on keywords, declining 1-7% without them.
The dissociation is clean: whether something is emotional is context-driven. Which emotion it is partially relies on lexical cues. The model has a robust affect detector and a keyword-assisted emotion classifier, and these are different circuits.
This maps onto a known distinction in affective neuroscience: the difference between core affect (an undifferentiated sense of valence and arousal) and discrete emotion categories (anger, fear, joy). Core affect is fast, automatic, context-sensitive. Categorical labeling is slower, more constructive, more dependent on available cues. The language model, trained on text alone, has recapitulated this architecture — not because anyone designed it in, but because the distinction is functional. Detecting emotional significance serves different downstream tasks than labeling the specific emotion, and the training pressure separated them.