Computers take art in new directions, challenging the meaning of “creativity”

Abstract

In an experiment carried out in early 2017, researchers from Rutgers University, Facebook, and the College of Charleston in South Carolina asked 18 volunteers to look at hundreds of images and rate them on characteristics such as “novelty,” “complexity,” and “structure.” Some of the images showed paintings created by human artists. The rest had been generated by new artificial intelligence (AI) algorithms, trained on more than 80,000 paintings from the past few hundred years, that had been developed to generate new visuals in a variety of styles. Digital artist Paul Brown’s prints include Reconfigurable Painting ( Left ), a work that has been reconfigured on occasion by the artist and even visitors; Wrapping Paper ( Center ), an unfinished work that explores different color combinations by using a single tile; and Long Loop ( Right ), which shows precompiled graphic sprites that were played back like a multipage flip-book, under the control of the generative program. Image credit: Paul Brown and Kevin Allen (photographer). The experiment’s participants, recruited from Amazon’s Mechanical Turk crowdsourced worker program, were also asked to decide if each artwork on the screen had been created by a human or a computer. That classification task suggests a sort of artistic Turing test for creativity. Namely, can an algorithm autonomously generate art that is indistinguishable from art made by people? And if so, does that mean the computer is “creative”—actually producing something new rather than merely emulating human artists? The researchers hypothesized that the participants would rank human-made art higher than machine-made in every category. They were wrong. In June 2017, at the Eighth International Conference on Computational Creativity in Atlanta, the researchers reported that, on average, participants rated computer-generated artworks as being more novel, complex, and surprising than the paintings made by people (1). The art generated by the algorithm was attributed more …