Hi Computer, Draw Me Another Building

Is a drawing still a drawing when it is literally drawn by a computer? As we enter into the milieu of signalization and data, John May has drawn the conclusion that: “no more drawings, only images.” However, beyond this narrow definition of drawing as orthographic gestures, could we not encompass its digital turn within a larger lineage of crafting lines and figures, albeit with its own specific set of peculiarities. Artificial Intelligence (AI) has recently found its place among the multitude of computational techniques we wield within our discipline, one that serves as an explicit manifestation of such a transformation. Thus, instead of merely fantasizing about AI’s capacity for efficient spatial solutions, we must explore its unique potentials and aesthetic properties as a medium in its own right.

While there is a diverse array of AI models available, the ability for Generative Adversarial Network (GAN) to generate ‘fake’ images is the most disruptive way AI finds its way into our discipline. With roots in creating realistic deep-fake faces of people who do not exist, one can appropriate the algorithm to generate plans that do not exist as well (Figure 1-3). Underlying this algorithmic technique is a feedback loop between two neural networks: the generator and the discriminator. The discriminator, after being trained using a dataset of Scottish tower poché plans for instance, will judge the random noise produced by the generator comparatively to determine its authenticity. As the generator learns to improve its algorithm to better draw novel poché plans, it will begin to fool the discriminator into believing that this generated noise is the ‘real.’ Even though this process of learning and mimicry appears similar to the traditional model of precedent studies, the notions of form, space, or cultural histories within the original plans are irrelevant to these algorithms. Rather, all images, both ground-truth and GAN-generated, are inherently pixels of abstract numeric variables for the machine to consume, evaluate, and create. While this set of pixels may appear convincing to our eyes at first, the AI plans unravel themselves as uncanny imitations of buildings upon closer examination. Through approximating and hybridizing, the act of faking becomes designing, drawing imprecise floor plans with blurring edges and gestural figures. Rather than reproducing buildings and plans as we understand, the computer will draw you another building with its own distinctive algorithmic hands.

Figure 1: Another Building, Bar Poché

Figure 2: Another Building, Rectangle Poché

Figure 3: Another Building, Wobbly Poché