Most of the art created by means of AI uses what is referred to as a ‘neural network’, which is a computer system modeled on the human brain and nervous system, meant to mimic certain human patterns of behavioral and cognitive abilities. In 2015, Google created and released a pattern-finding software called “DeepDream”, which prompted artists and techies alike to experiment with and train the network to create new imagery from multiple existing images. This program employs a convolutional neural network (CNN) that finds and enhances patterns from images through algorithmic pareidolia, or a controlled and calculated tendency to interpret vague stimuli as familiar things to the viewer, such as seeing shapes in clouds.
In addition to neural networks such as “DeepDream”, those using AI to create art are specifically using models called “GANs” or “generative adversarial networks”, which, according to researcher Ian Goodfellow’s 2014 article, represent the next step in the evolution of neural networks. GANs are class of machine learning systems; they are trainable pairs of neural networks that can generate imagery on their own after being given a followable framework and an end goal. Training a GAN can be time-consuming and takes a certain knack for understanding the greater system. Educating a GAN also means feeding it a certain grouping of images and information, and then asking it to perform a task based on that imagery and information. For instance, a GAN could be trained to recognize landscapes via many pre-existing images of landscapes, and then programmed to produce a new, original landscape of its own. Most current GANs are limited in their capabilities and can only perform one cognitive task at a time. In order for a GAN to perform a different task, the entire system must be erased and retrained from scratch.
“Art”ificial or Truly Intelligent?
The field of art and AI is an active one, and many artists are working in tandem with computers and programs such as GANs to produce results that are both visually and intellectually stimulating and interesting. The (human) artist must masterfully choose the program they wish to use, decide which imagery or subject matter should be in focus, and then train the program to do as they intend. In doing so, much of the trial and error becomes part of the process, and thereby part of the end result. The artist’s intentions do not always come out as they anticipated, but that, within itself, is much of the appeal to this type of creative process.
In October 2018, history was made when an original work of art—a portrait—created using AI was sold for over $400,000 at auction in New York, wildly outperforming its original high estimate of $10,000. The work was created by an artist collective made up of three French men who go by the name “Obvious”, and who used their own GAN’s data set of portraits to create the aforementioned work. Other noteworthy artists currently working in this medium include Mario Klingemann, Anna Ridler, Robbie Barrat, and Sofia Crespo.
Two camps have seemingly emerged from the discussions and debates regarding the legitimacy of this movement: those who believe that the computer and its programs are their own creative entity separate from the human, and those that believe the programs and the computer are tools that require guidance and manipulation from people to create truly creative and original artworks. Most people working in the field of machine learning feel as though the tool is the medium, and that the end result is a collaboration between machine and artist. As the field is still very much in its infancy, it will be extremely interesting to watch it unfold in the coming years.