LLMs Judging Architecture: Generative AI Mirrors Public Polls

This paper uses large language models (LLMs) to judge among three pairs of architectural design proposals which have been independently surveyed by opinion polls: department store buildings, sports stadia, and viaducts. The tool instructs the LLM to use specific emotional and geometrical criteria for separate evaluations of the image pairs. Those independent evaluations agree almost totally with each other. In all cases, the LLM consistently selects the more human-centric design, and the results align closely with independently conducted public opinion poll surveys. The convergence of AI judgments, neuroscientific criteria, and public sentiment highlights the diagnostic power of AI-based criteria-driven evaluations. This technology provides objective, reproducible architectural assessments capable of supporting design approval and policy decisions. A practical tool that unifies scientific evidence with public preference can promote a more human-centered built environment.