Logo image
Back
Can ChatGPT rate body-related images like humans? Effects of temperature and few-shot prompting on valence and arousal judgments
Journal article   Peer reviewed

Can ChatGPT rate body-related images like humans? Effects of temperature and few-shot prompting on valence and arousal judgments

Yueyang Xiao, Sanle Zhao, Qianyu Zhou, Yuchen Zhan, Jiayi Han, Feng Ji and Jinbo He
Body image, v 57, 102097
01 Jun 2026
DOI: https://doi.org/10.1016/j.bodyim.2026.102097
PMID: 42068738

Abstract

Life Sciences & Biomedicine Psychology, Clinical Psychology, Multidisciplinary Science & Technology Psychiatry Psychology Social Sciences
Body-related images (e.g., idealized thin/muscular bodies and more diverse body types) are used to study appearance-related affect, body dissatisfaction, and social comparison. Preparing such stimuli typically requires human ratings of core emotional dimensions (e.g., valence and arousal), which is labor-intensive and difficult to scale across contexts. Multimodal large language models (LLMs), such as ChatGPT, offer the potential to automate preliminary norming tasks through multimodal training and adaptability, but their validity for such rating remains unknown. Across two studies, we tested whether ChatGPT models could approximate human emotional ratings (valence and arousal) of body-related images, and whether performance depended on temperature settings and few-shot prompting. Using 80 established images of ideal and non-ideal male and female bodies, Study 1 evaluated the temperature effect by testing GPT-4o-mini and GPT-5-chat-latest across four temperature settings. Study 2 examined the few-shot effect using GPT-4o-mini, GPT-5-chat-latest, and GPT-5 with 0-5 example shots. Intraclass correlations, Pearson correlations, and error metrics were computed to examine AI-human alignment. Results revealed that across models and settings, valence ratings aligned more closely with human scores than arousal ratings. Few-shot prompting significantly improved alignment, particularly for GPT-5. However, improvements were inconsistent across both temperatures and few-shot learning. The models showed promising alignment when guided by a few examples but exhibited sex-specific biases. Findings suggest that LLMs can support preliminary emotional norming of body-related images, particularly for valence when guided by examples, but sex-specific asymmetries highlight the need to develop bias-reducing strategies to ensure responsible use in future applications.

Metrics

1 Record Views

Details

Logo image