Culturally uneven urban perception in large language models

The paper introduces a measurement framework to assess cultural bias in urban perception by large language models (LLMs), utilizing a globally stratified street-view image dataset. Findings reveal that outputs from three multimodal models exhibit a bias towards European and North American cultural framings, and while prompting can align AI responses with specific regional human perspectives, it fails to capture the diversity of human responses, leading to potential biases in urban evaluations. This highlights the risks of deploying LLMs as neutral tools in urban analysis, underscoring the need for careful consideration of cultural contexts in model application.