Examining typological configurational compression in AI-generated housing through space syntax

Purpose This study investigates whether housing plans generated by generative artificial intelligence (AI) reproduce the spatial logics embedded in established residential typologies and examines the implications of AI-driven housing design for spatial diversity and, more broadly, for discussions of cultural sustainability within the built environment (SDG 11). While prior research has focused largely on visual realism and functional plausibility, the configurational integrity of AI-generated housing remains underexplored. The study evaluates typological sensitivity and proposes a quantitative framework for assessing potential configurational compression in spatial organisation. Design/methodology/approach Residential plans representing three typologies – courtyard, row house and apartment – were generated using DALL·E 3 under controlled conditions and analysed using space syntax metrics (VGA, Integration, Choice and Connectivity). Principal component analysis (PCA) projected the plans into a configurational morphospace. Centroid separation, silhouette analysis and an exploratory configurational compression index (CCI) were employed to evaluate positional differentiation and cluster cohesion. Findings Configurational variation was largely organised along a single centralisation gradient (PC1 = 89.22%). Although centroid distances indicated measurable positional differentiation, the silhouette coefficient (0.272) and the exploratory CCI (0.335) suggested limited cluster cohesion and partial overlap among the analysed typologies. The findings suggest that AI-generated housing designs preserve recognisable visual typological characteristics while tending towards simplified relational spatial organisation within the analysed dataset. Originality/value The study introduces the concept of typological configurational compression and proposes an exploratory analytical framework integrating morphospace projection, cluster validation and compression indexing to evaluate configurational convergence in AI-assisted housing design.