Answers are grounded in this paper and cite their sources.
AI systems may agree with one another about the quality of buildings and streets, even when their judgments differ from people’s. We tested this possibility using 120 paired building-frontage and streetscape images selected from 39,047 street-view panoramas across six Chinese urban districts. The images were rated by 100 architecture or urban-planning students, 100 students from other disciplines, and five vision–language models (VLMs). The two student groups strongly agreed on how they ranked the images (Spearman correlations of 0.914–0.946 for building frontages and 0.923–0.943 for streetscapes), although architecture/planning students consistently gave lower scores. By contrast, the AI models agreed relatively well with one another but showed little correspondence with either student group in their rankings. We also examined whether visible elements such as vegetation, roads, sidewalks, and sky could explain the ratings. These simple image features predicted AI streetscape scores more successfully than human scores, consistent with the possibility that their evaluations reflect different visual cues, although the study could not establish how either reached its judgments. AI may support preliminary screening of large image collections, but human evaluation remains essential for architectural and urban-design decisions.
No takes yet. Share an insight, caveat, or question.
Mou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: