Can large language models generate metaphors for emotion? If so, how? Interpreting and generating emotional metaphors is a core task of clients and therapists in psychotherapy, as metaphors allow individuals to express nuances of emotion that direct expressions cannot achieve. With the rise of digital mental health, it is important to ask whether digital interventions can undergo this task. Large language models are a unique means of exploring this possibility. The present study used a trained model of GPT-2 to generate metaphor-like sentences with vector computing. In Experiment 1, 270 sentences for 30 emotion-noun pairs are evaluated based on PCA and t-SNE analysis. In Experiment 2, human respondents matched the GPT-generated sentences with common metaphors retrieved from past literature. Results showed that LLMs are capable of producing emotional metaphors that are recognizable by humans using vector computing, but this ability is limited by overreliance on surface similarity. Future studies should focus on training LLMs to extract relational structures and understand human metaphors.