Quantifying the Affective Gap: A Zero-Shot Evaluation of LLMs on Fine-Grained Emotion Taxonomies

The paper doesn't just show models fail at emotion tasks; it reveals that frontier LLMs are being deployed in mental health and conversational AI without validation on fine-grained affective reasoning, despite this being measurable and testable before release.

This connects directly to the Anthropic safety testing precedent from last week. Anthropic cleared export restrictions by submitting to structured safety evaluation, establishing that rigorous benchmarking can satisfy governance concerns and unlock market access. The emotion taxonomy paper suggests the inverse problem: models are already in production without equivalent rigor. The gap between what gets tested (safety, alignment) and what gets validated (task-specific reasoning on safety-critical domains) is widening. Meanwhile, recent work on evidence-grounded LLM outputs in finance showed that even with structured inputs, hallucination detection remains unreliable, hinting that emotion classification may face similar verification challenges at scale.

If major labs release emotion classification benchmarks as part of standard pre-deployment validation within the next two quarters, that signals the field is closing this gap. If not, and if mental health applications continue shipping without this evaluation, that confirms emotion reasoning remains an unmonitored risk in production systems.