Tasting Emotions with GPT-4o: Comparing AI and Human Norms in Metaphoric Associations

Abstract Large language models (LLMs) have emerged as efficient tools for generating psycholinguistic norm data. However, their capacity to capture embodied cognition, particularly metaphorical associations grounded in sensorimotor experience, remains insufficiently understood. The present study provides the first systematic investigation of taste-emotion metaphorical mappings in GPT-4o by comparing model-generated responses with established human norms across four experiments. In Experiments 1 and 2, we used free-association tasks, prompting GPT-4o via the OpenAI Application Programming Interface (API) to generate the taste word most strongly associated with emotion-related stimuli (Experiment 1) and the emotion word most strongly associated with taste stimuli (Experiment 2). In Experiments 3 and 4, we used Likert-scale rating tasks to quantify the associative strength of taste-emotion pairs (Experiment 3) and concept-taste pairs (Experiment 4). We systematically varied temperature settings (T = 0, 0.7, 1.0) and iteration conditions to assess response stability. Across tasks, GPT-4o showed weak-to-moderate correlations with human norms, indicating meaningful but incomplete alignment in metaphorical mappings. The model aligned more closely with human responses for basic emotion words (e.g., love) than for emotion-laden concepts (e.g., wedding), and performed better in structured rating tasks than in open-ended free-association tasks. GPT-4o also exhibited high internal consistency across temperature settings and iterations. These findings suggest that current LLMs can recover a substantial portion of conventional taste-emotion mappings from language alone, but remain limited in approximating fully embodied metaphorical knowledge. The results have important implications for the use of LLMs in psycholinguistic research, especially in domains where sensorimotor grounding is theoretically central.