Proceedings of the Annual Meeting of the Cognitive Science Society

Human perception often relies on names when forming first impressions of unfamiliar people. As text-to-image (T2I) models are increasingly used in everyday content creation, names may also influence what these models generate under otherwise neutral prompts. We study name-based bias in mainstream T2I models along two dimensions: appearance patterns and socioeconomic status patterns. We curate name sets from China, America, Britain and France, generate over 6,280 images, and use automated VLM-based scoring to assess age appearance and styling, as well as occupation and living-environment attributes in work and home scenes. Across models, changing only the name yields reliable shifts in these attributes. The strength and consistency of the pattern vary by model, with Nano Banana and ChatGPT-5 showing more uniform name-linked defaults. We propose a quantitative framework to identify and compare name-based biases, helping advance fairer AI systems.