Culturally uneven urban perception in large language models

Large language models increasingly used to describe and evaluate cities are not culturally neutral, according to a study that tested three frontier multimodal models against a global street-view dataset. The research finds that the models’ baseline urban judgments align more closely with European and North American cultural framings than with other regional perspectives. The study, submitted to arXiv on 21 April 2026 and revised on 8 June 2026, introduces a measurement framework that probes whether LLM-based urban perception operates from a universal standpoint [1]. Researchers used a globally stratified sample of street-view images and prompted the models to produce open-ended descriptions and structured scores under both neutral and culturally conditioned instructions [1]. Across all three models, the neutral condition was not neutral in practice. “Prompts associated with Europe and Northern America remained systematically closer to the baseline than many non-Western prompts, indicating that model perception is organized around a culturally uneven reference frame rather than a universal one,” the authors report [3][4]. The asymmetry extended to the geometry of semantic space. In a local principal component analysis, identity-conditioned responses occupied distinct positions around the neutral prompt rather than clustering randomly, and cultural prompting changed both the size and the direction of movement away from neutrality [3][4]. The same street scene was reframed along structured, identity-specific trajectories rather than through unsystematic lexical variation [3][4]. Comparisons with regional human text-image benchmarks showed that culturally proximate prompting could improve alignment with human descriptions, but it did not recover human levels of semantic diversity and often preserved an affectively elevated style [2][3]. The models also introduced sentiment-based self-favouring bias when comparing, evaluating or representing cities across cultural contexts [1]. Structured judgments of safety, beauty, wealth, liveliness, boredom and depression were interpretable but only partly reproduced human group differences [3][4]. The findings align with broader research on cultural perception in LLMs. A separate study of three state-of-the-art models across 110 countries and regions found that culture-conditioned generations contain linguistic “markers” that distinguish marginalized cultures from default cultures, and that LLMs exhibit an uneven degree of diversity in the cultural symbols they associate with different regions [9]. That work concluded that current language models have “uneven cultural perception and inadequate cultural knowledge, especially regarding marginalized cultures” [9]. A related framework, SPAGBias, has extended bias research into the spatial domain by evaluating spatial gender bias in LLMs through a taxonomy of 62 urban micro-spaces and three diagnostic layers [5]. The authors of the urban-perception study argue that their results indicate a systematic risk in treating AI as a neutral tool for urban tasks, especially when model outputs are used to compare, evaluate or represent cities across cultural contexts [1][2].