Emergence of Hierarchical Emotion Organization in Large Language Models

As large language models (LLMs) increasingly power conversational agents, understanding how they model users' emotional states is critical for ethical deployment. Inspired by emotion wheels, i.e., a psychological framework that argues emotions organize hierarchically, we analyze probabilistic dependencies between emotional states in model outputs. We find that LLMs naturally form hierarchical emotion trees that align with human psychological models, and larger models develop more complex hierarchies. We also uncover systematic biases in emotion recognition across socioeconomic personas, with compounding misclassifications for intersectional, underrepresented groups. Human studies reveal striking parallels, suggesting that LLMs internalize aspects of social perception. Beyond highlighting emergent emotional reasoning in LLMs, our results hint at the potential of using cognitively-grounded theories for developing better model evaluations.

As AI systems become everyday conversation partners, they increasingly need to respond to people's emotions. But we still know little about how they represent those emotions internally. We studied whether large language models organize emotions in a structured way, rather than simply labeling text as happy, sad, or angry. Inspired by psychological ``emotion wheels,'' we developed a method that looks at the probabilities a model assigns to many emotion words across thousands of situations and uses them to reconstruct an emotion hierarchy. We found that larger models form richer trees of emotions, and these trees increasingly resemble emotion organizations proposed in human psychology. However, this apparent emotional understanding is not equally reliable for everyone. When models were asked to interpret scenarios from different demographic perspectives, they made systematic errors: for example, some underrepresented personas were more likely to have emotions misread as anger, fear, shame, guilt, or frustration. A human study showed similar patterns in some cases, suggesting that models may absorb social biases present in human behavior and data. Our results provide a new way to evaluate emotional reasoning in AI, while highlighting the need for safeguards before emotion-aware systems are used in sensitive settings such as companionship, counseling, education, or customer support.