Proceedings of the Annual Meeting of the Cognitive Science Society

Human perception relies on crossmodal correspondences, systematic links between different sensory modalities. As text-to-image (T2I) models increasingly serve as aesthetic infrastructure, they risk codifying cultural biases embedded in sensory language. We analyze the mapping of cross-sensory bias in mainstream T2I models, examining how gustatory, tactile, auditory, and olfactory adjectives map onto visual attributes. Using a modality _ carrier design, we evaluate five bias dimensions in 6,000 images via automated VLM-based assessment: color, demographic, entity, situational, and layout. Our results reveal widespread cross-sensory homogenization, with models projecting abstract sensations onto a narrow set of cultural prototypes. This study presents a framework for quantifying cross-sensory bias in T2I models and offers tools for auditing and mitigating their broader cultural and social impact.