Robots that can create art!?

By Liz Ribeiro

In October, 2018, the painting entitled “Portrait of Edmond De Belamy” was auctioned at Christie’s in New York for the price of 432 thousand dollars. This event itself did not represent anything extraordinarily new for the art world, since works of art are auctioned at art galleries all over the world on a daily basis. Nevertheless, if one bothers to pay closer attention to the lower right corner of the painting, one will see a relatively unusual signature: an algorithmic formula. The portrait of Edmond De Belamy was the first auctioned painting to be entirely made by Artificial Intelligence (AI). In spite of this achievement, paintings made with the help of AI are not perceived as a great novelty. Since 1983 artist and professor at the University of California, Harold Cohen, has taken part in the AARON program, a robot that has made its works of art autonomously for over 2 decades. Besides that one, there is also a creation by Simon Colton called “The Painting Fool”, which has used newspaper articles, among others means, to generate its paintings since 2013.

Behind the artwork auctioned at Christie’s and several other galleries that employ similar technology is the Collective Obvious, co-founded by Hugo Caselles-Dupré, Pierre Fautrel and Gauthier Vernier. Obvious, based in Paris, aims to broaden the possibilities of artistic creation with the use of artificial intelligence and machine learning. The Project that gave rise to the “Belamy family” started from the discovery of the Generative Adversarial Networks (GAN), which are machine learning algorithms used in the generation of images. The system was “stocked” with 15 thousand classical portraits from the fifteenth to the twentieth centuries. From this database, two algorithms work in a kind of competition: the “generative part” creates the images and the “discriminative part” discards those clearly made by an Al. The final result is an image painted with inkjet on canvas, created by the algorithm that managed to ‘deceive’ the discriminative part. Thereafter, it is framed and signed by the formula that originated it. Would the use of GAN in art production be able to reproduce the “essence” of human creativity and then always produce unique works with this technique?

Creativity is one of the key elements in modern art, being directly linked to the notions of “endowment”, “inspiration” or “enlightenment”. It is clear, however, that over time there have been countless deconstructions, revolutions and re-significations of this concept, as well as the concepts of beauty, artworks and artists themselves. On the other hand, contemporary art is always broadening the ontological and axiological boundaries of artistic production. The use of algorithms in the creation of works of art such as GAN represents exactly this radicalization, especially of what is meant by human creativity.

Through machine learning, GAN’s generative networks are progressively better at generating images that are more similar to the dataset they receive, since their function is to model the distribution of legitimate artworks. Simultaneously, discriminatory networks are more efficient in selecting images more similar to this set by learning the boundary between the class of artificial images and the class of real works.

It is important to realize that (1) the generative network is fed by a random noise, unknown to the network, and from it, a “novelty”, which will turn out to be the image, is generated; and (2) that the training set is determined by humans. This means that even with a random result, the possibilities of the image will always be within this specific set. Thus, creativity would reside in both randomness and its entry into the generative network. The algorithm is not creating a painting from scratch, but rather generating, from a dataset, an image more similar to the ones in it.

In this sense, the artistic production by or with robots makes us reassess the limitations of both humans and machines all the time. It is noticeable that the possibilities of creation by GAN, though extremely wide, are still limited by the dataset the algorithm has. On the other hand, it may be able to generate many more works, which, in turn, may be a lot more distinct from one another than the human physical capacity would allow. We are left to wonder if the concepts of what it is to be human and what it is to be a machine will blur in the near future, given the emergence of robots that can actually create art.

Liz Ribeiro is an independent researcher at Ateliê de Humanidades and master student in anthropology at IFCS-UFRJ.

This article was originally published in the Brazilian newspaper Jornal do Brasil on january 30th, 2019 (republished on the website of the Ateliê de Humanidades). Translated by Marco Aurélio de Carvalho Silva.