<p>This study measures architectural originality in AI-generated facades by looking at measurable structural differences. It uses a controlled prompting method with four levels of semantic conditions: traditional, minimal, radical, and experimental. Two multimodal systems, ChatGPT and Gemini, created facade outputs under the same design limits. Each image was examined in a shared embedding space and evaluated using three formal metrics: symmetry difference, Shannon grid entropy, and a convex-hull-based mass articulation index. The results indicate that increasing semantic intensity does not lead to consistent structural changes. Instead, facade configurations shift into defined areas of transformation. Changes at the embedding level mostly relate to symmetry relaxation and, to a lesser extent, to volumetric articulation, while grid entropy shows a weaker link. Throughout all conditions, typological coherence remains statistically stable. These findings suggest that originality in AI-generated facades can be seen as a controlled structural adjustment within defined formal systems. This contributes to the mathematical study of architectural form in computational design contexts.</p>