As artificial intelligence increasingly shapes architectural and planning practices, understanding how these systems interpret and represent urban environments becomes crucial for critical practice. This research proposes a novel methodology to study the latent urban imaginaries embedded within contemporary AI models, specifically examining how Stable Diffusion and CLIP encode and reproduce urban knowledge. By generating a controlled collection of public space images across 241 global capitals—varying only in location names in otherwise identical prompts—we create a standardised dataset for analysing AI's embedded urban understanding. Our analysis unfolds in two complementary directions: a visual assessment that identifies proto-images through similarity clustering, revealing the geographical extent of the model's urban knowledge, and a semantic evaluation using CLIP to analyse the presence of established urban theory concepts across the generated spaces. Both analyses are mapped geographically, producing cartographies that visualise how AI systems distribute urban knowledge and perpetuate biases across global contexts. As these technologies become increasingly integrated into design tools and workflows, understanding their embedded spatial assumptions and biases becomes essential for informed architectural and planning practice. By highlighting dormant themes implicit in machine learning models, our study offers a critical platform for debating the impact of generative AI on urban planning and speculation, ultimately challenging the narratives and futures it constructs.