The AI Brain in the Cultural Archive

At this point we know much more about human creativity and how it works than we do about “AI creativity.” In philosophy, psychology, cognitive science, and other fields, many alternative theories and types of creativity have been proposed. With time, we are likely to see the same for “AI creativity.” But we are not at this point yet. After neural networks are trained on trillions of text pages or billions of images harvested from the Web, they can compose new texts and images. However, the abilities of the creative nets are not defined by traditional algorithms; instead, they are distributed among trillions of connections between billions of artificial neurons. In other words, we managed to create a technology that is in some ways similar to a human brain—and that is so complex we can’t understand how it works.

The current generation of creative AI systems have been trained on very large and heterogeneous collections. However, it’s equally interesting to limit the training data set to a particular area in the larger space of human cultural history. Unsupervised by Refik Anadol, currently on view at MoMA, is such an experiment. Its dataset is approximately 140,000 artworks in MoMA’s collection. This collection, in my eyes, represents one of the most creative and experimental periods in human visual history. It captures the feverish and relentless experiments of artists, designers, architects, and photographers to generate new visual and communication languages and—most importantly for these artists—to “make it new.”

At first glance, this logic of modernism is the opposite of the process behind creative AI. In many different times and very different places, modern artists wanted to break with traditional art and its key qualities such as visual symmetry or narrative content. Their creativity was based on a fundamental rejection of everything that came before—at least in theory, and as expressed by their manifestos. Deep neural networks, however, are trained in the opposite way: by learning from the historical past and all art created up until now. In a way, a neural network is like a “meta” conservative artist, studying in the virtual “museum without walls” that contains all human historical art.

Contemporary artworks that use neural networks trained on massive cultural databases, such as Anadol’s Unsupervised, stand in a long tradition of making new art from accumulations of images and other media. In this way, these artworks continue expanding possibilities of art and its methods, and especially “database art.”¹ This expansion includes new methods for “reading” cultural databases. The double brain of Unsupervised’s creators (i.e., a human brain and an “artificial brain” of the neural net) does not directly quote from historical art or create media collages like earlier artists. Instead, the neural network extracts complex patterns and relations from hundreds of thousands of artworks and objects collected by MoMA. It then uses everything it learned to generate an endlessly changing visual field that interpolates between the features of these artworks and the relationships between them in unpredictable ways. The surface is constantly changing, oscillating between 2D and 3D, between pure abstraction and seemingly recognizable echoes of particular artworks, between gentle movements and dramatic visual explosions. These are not mechanical imitations or simulations of what already was created. In my opinion, they are genuinely new cultural artifacts with contents and styles that we have never seen before.

Looking at the work, I am reminded of the alien planet in Andrei Tarkovsky’s film Solaris (1972). Solaris synthesizes human-like “guests” from the memories of human inhabitants of the space station orbiting the planet. The surface of the planet changes throughout the film, sometimes showing concrete images, then starting to form islands. Unsupervised also generates something that is both very familiar and also alien, detached from all the thoughts and impulses of the artists who made works in MoMA’s collection. Perhaps it is an early glimpse of the new forms of consciousness that will emerge between the humans and computer systems they engineered—or a vision of a future language of culture, which we are just now learning how to read.