Speaking in Feelings: Facilitating Human Emotion Communication through Analogy by Large Language Models

Measuring emotion remains a persistent challenge due to poor alignment between theoretical constructs and available measurement tools, as well as issues of underspecificity and low reliability. Dimensional approaches—especially those incorporating social-affective dimensions—are being explored for their potential to enhance the quality of emotional data. Recent advances in large language models (LLMs) offer a novel opportunity to improve both emotion measurement and communication, leveraging their exceptional pattern recognition capabilities and capacity for multi-modal generation. In this study, we investigate for the first time whether LLMs can successfully mediate human emotional communication through analogical representations—expressing feelings not by naming them, but by evoking them. We utilized multi-dimensional affective ratings from 547 participants across eighteen emotion labels obtained in a prior study. Based on this human emotion representation and custom-designed prompts, an LLM generated both textual and high-resolution (1024 × 1024) image outputs for each label. Two independent raters selected ten texts and one image per label, which were then presented to a new sample of 138 participants via an online psychology experiment platform. Participants rated these multi-modal stimuli using the same multi-dimensional affective scale. Angular and Euclidean distance metrics were used to evaluate the similarity between participants’ affective responses and the original ratings. Emotional experiences for fifteen out of eighteen emotion labels were successfully reconstructed through exposure to LLM-generated stimuli. However, three emotions—disgust, satisfaction, and gratitude—were not effectively conveyed through either modality. The results also highlighted contextual modulation of emotion expression, influenced by media format and individual mood pathology. Across all categories, image-based stimuli evoked stronger affective responses than text. Moreover, participants with elevated depressive symptoms exhibited heightened emotional reactivity to shared stimuli—a potential affective signature of depression, rather than a general emotional amplification. This study introduces a novel pipeline for eliciting and transmitting human emotional experiences through language models using analogy as a bridge—a method that is both replicable and empirically validated. Still, establishing true equivalence of emotional experiences requires further theoretical and mathematical development. Future directions include incorporating physiological markers, comparing across LLM architectures, and refining generalizability for broader human-centered applications in affective computing.