Do Androids Dream of Curtain Walls? How Image Generating AIs Represent Contemporary Architecture

American author Philip K. Dick introduced the world to the concept of androids that could be trained on the memories of real people, in a similar fashion the development of Ais is based on the use of human-produced data to learn in a statistical way how to describe or even draw something. Said training can be executed using dataset with different degrees of specificity or curation in a specter going from the punctual curation of individual data up to the employment of vast, un-curated datasets. The latter case is representative of how modern open-source models like stable diffusion are trained to be able to generate images either from a textual or a graphical input. Using said models the authors run a series of tests, to assess how models that didn’t undergo a heavily specific training tend to represent contemporary architecture, what are the limitations in using generic pre-trained models, and what is their potential in directing the output toward specific topological features. The paper concludes summarizing the features that the model identified with contemporary architecture in terms of morphology, stylistic characteristics, and strategical aspects, while defining the limits emerged from testing.