Law, S., Valentine, C., Kahlon, Y., Seresinhe, C., Tang, J., Gath Morad, M., Fujii, H.  In preparation

Here, we test how well AI-generated building facades match three types of human descriptors: geographical (e.g., "London townhouse"), objective (e.g., "angular" vs "curvy"), and affective (e.g., "utopian" vs "dystopian"). Using a Latent Diffusion Model to generate facade images, we evaluated alignment through both AI and human assessments.

We find that AI performs well with geographical prompts, though with notable regional biases. Objective descriptors produce better alignment than affective ones, suggesting that whilst AI can replicate visual patterns and regional styles, it struggles to capture the emotional and experiential qualities that shape how we actually experience architecture. These findings highlight both the potential and current limitations of AI in architectural design, pointing to fundamental gaps in how these models understand the embodied experience of space.