Generative artificial intelligence for sustainable architectural design within urban contexts: evaluating potentials, constraints, and barriers through cost–dissimilarity scenario analysis

Generative Artificial Intelligence (AI) is increasingly used to support early-stage sustainable architectural and urban design through the rapid production of technical representations. However, existing AI-based cost evaluation and Multi- Criteria Decision Analysis (MCDA) approaches typically assess predefined design options rather than systematically linking representational variability to economic outcomes. This paper proposes a tow-phase comparative framework that explicitly operationalize this link. The empirical analysis is based on a structured dataset of paired design alternatives, consisting of baseline solutions and AI-generated variants developed at multiple levels of representational definition. In Phase I, representational dissimilarities are formally identified and classified according to their impact on technical meaning (e.g., geometry, element definition, and implied quantities). In Phase II, these differences are translated into scale-consistent cost estimates, enabling the calculation of cost-dissimilarity relationships across scenarios. This structure allows testing the casual hypothesis that reductions in representational definition systematically increase cost uncertainty. Results show a non-linear increase in cost deviation as representational definition decreases, with a threshold effect at lower levels of detail, where divergences shift from geometric variation to changes in inferred quantities and system composition. The framework thereby delineates an operational range in which AI-generated outputs remain economically interpretable, and a non-optimal range where cost reliability degrades. To address the sustainability dimension, the analysis incorporates cost-based proxies linked to material intensity and construction systems, enabling a first-order assessment of resource implications across alternatives. Overall, the proposed framework differs from existing approaches by explicitly coupling representational analysis with cost translation in reproducible workflow, positioning generative AI (GenAI) as a transparent decision-support tool in costaware and sustainability-oriented design processes.