How LLMs Feel Emotions Differently: AI Models Treat Feelings as Discrete Categories, Unlike Humans

Large language models handle emotions in fundamentally different ways than humans do, according to research from arXiv examining how these systems internally conceptualize feelings. Rather than embracing the messy, nuanced emotional spectrum that characterizes human experience, LLMs organize emotions into clean, separate mental categories.

The Research Approach#

To understand how LLMs and humans differ in emotion processing, researchers created a benchmark dataset spanning 15 emotion categories grounded in cognitive appraisal theory—a psychological framework that breaks down emotional situations along interpretable evaluative dimensions. They then asked both LLMs and humans to respond to identical emotional scenarios, extracting what they call "behavioral representations" of each emotion concept based on the subjects' external responses.

Key Finding: Categorical vs. Fluid#

The analysis revealed a striking pattern: LLMs represent emotion concepts in a far more categorical, homogeneous, and deterministic fashion than humans. Where human emotion is inherently subjective, variable, and often contradictory—a single person might feel multiple emotions simultaneously or express the same emotion differently depending on context—LLMs treat each emotion as a distinct, internally consistent unit. LLMs place different emotions further apart in conceptual space and show much less internal diversity when representing a single emotional state.

Robustness and Training#

Interestingly, this categorical structure holds up across different variations in how the models are prompted, including changing task framing and adding different demographic personas. The researchers also traced when this behavior emerged during model training: the discretized representation of emotions appears midway through training and persists regardless of subsequent post-training strategies like alignment tuning.

Why It Matters#

This finding has implications for applications where LLMs interact with users emotionally—customer support, mental health chatbots, or safety-critical systems. If an AI model lacks the inherent subjectivity and fluidity of human emotion processing, it may struggle to genuinely understand or appropriately respond to the complex emotional realities users bring to conversations.