Whether, Not Which: Mechanistic Interpretability Reveals Dissociable Affect Reception and Emotion Categorization in LLMs

A groundbreaking study by researcher Michael Keeman introduces a clinical validity test for emotion processing in large language models, challenging previous assumptions about how AI understands human emotion. Using mechanistic interpretability methods—including linear probing, causal activation patching, knockout experiments, and representational geometry—the research tested six models (Llama-3.2-1B, Llama-3-8B, Gemma-2-9B and their instruction-tuned variants) with clinical vignettes that evoke emotions through situational cues alone, completely removing explicit emotion keywords like 'devastated' or 'joyful.'

The study discovered two dissociable emotion processing mechanisms within LLMs. The first, 'affect reception,' detects emotionally significant content with near-perfect accuracy (AUROC 1.000) and operates consistently across all tested models, suggesting this capability emerges early in model training. The second mechanism, 'emotion categorization,' which maps detected affect to specific emotion labels, showed partial dependence on keywords, with performance dropping 1-7% when keywords were removed, though it improved with model scale.

Causal activation patching experiments confirmed that keyword-rich and keyword-free stimuli share representational space, transferring affective salience rather than specific emotion-category identity. This finding falsifies the 'keyword-spotting hypothesis' that previously questioned whether LLMs genuinely understand emotion or merely recognize emotion-related vocabulary. The research establishes clinical stimulus methodology as a new rigorous standard for testing emotion processing claims in AI systems, with direct implications for AI safety evaluation and alignment practices. All stimuli, code, and data have been released for replication.