In 1968, Harold Cohen moved 5,000 miles across the globe. Although the British painter had been one of the most influential artists in the London art scene, he was beginning to question what constitutes art and what makes someone an artist. So he went to California.
He would never move again. During his residency as an art professor at UC San Diego, a colleague introduced him to the college’s punch-card computer system. Cohen was hooked. He quickly learned primitive coding languages, and using the resources of the college’s not-ten-years-old CSE program and nearby Stanford, began making a machine of his own, not to do calculations or analyze data, but to create art.
To do this, the “drawing machine” relied on thousands of rules set by Cohen, rules that a person would know naturally, but that an AI must be taught: a person has two legs, a tree has one trunk, an apple is round but not perfect. Over the course of decades, Cohen refined his rules, turning the machine’s creations from basic, monochrome squiggles into complex and colorful abstractions.
“I am constantly being asked whether I am a scientist or an artist, as if there’s a law that says I can’t be both,” Cohen said in 1995.
60 years later, Cohen’s art is no longer a strange experiment, but the foundational text of a new technological movement. Although they may be much more complex than in Cohen’s day, AI-generated images are now more accessible, popular, and inescapable than ever. Whether this is the future or end of art- or something else entirely- depends on who you ask.
However, the generators of today are quite different from Cohen’s machine, largely due to three key innovations. The first is the creation of vast compilations of photos, artworks, and illustrations- one modern generator, Stable Diffusion, has a database of over 2.3 billion images.
Both of the other two involve the generation itself. Generative adversarial networks, or GANs, were theorized in 2014 by a team led by Ian Goodfellow. Prior to GANs, AI generators would simply attempt to mimic real photos, often creating blurry and nonsensical images. GANs added another component: the discriminator. Similar to an email checking for spam, discriminators measure the generated images against real samples. If the data doesn’t pass, then both parts improve- the generator learns how to fool its opponent, and the discriminator learns what to watch for.
Although GANs produced remarkably stylistic images, the high electrical costs and generation time limited general appeal. Diffusion technology, which saw major use with the launch of DALL-E in 2021 (Wall-E and Salvador Dali, according to the creators), took that barrier down. Although diffusion is incredibly complex, for most people the results were easy to understand; AI art was cheaper, easier and faster than ever before.
Since then, AI-generated art has exploded in general popularity. DALL-E 2, which recently went into open beta after being in closed testing since April, reports having 1.5 million users create over 2 million images each day, and generation tool Midjourney, which emphasizes ease-of-use, has over 2.7 million users on its discord server. AI Art has also gained some credence within the traditional art world; In 2018, the generated art Portrait of Edmond de Belamy was put up for sale at the venerated Christie’s auction house in New York City. The initial price estimate was for $10,000. It sold for $432,000.
“As soon as I got onto Midjourney, I was like a kid in the candy store,” Jason Allen said. “There’s something really special about watching your imagination appear on the screen.”
Last month, Allen became the focus of an online firestorm after his AI-generated image, Théâtre D’opéra Spatial, took first place in the digital artwork category of the Colorado State Fair. Allen, who runs a board game company out of Colorado Springs, had gotten into AI art earlier this year, spending dozens of hours perfecting the prompt and variation for each work. He says he decided to enter the Fair both to show off his work as well as “make a statement about Midjourney.”
While he expected debate, Allen says that he was surprised at the level of backlash. After outlets like the New York Times and Washington Post covered his story, online users leapt to both defend him as an innovator and deride him as “actively anti-artist.” Allen says he started receiving hate mail and even some death threats.
“No one has the right to tell another person how to be an artist, or how to express themselves creatively,” Allen said. “If you’re creating your work from the soul, something like this shouldn’t affect you. You choose to represent or to illustrate your work in the medium that’s meaningful to you, that you want to express yourself. I chose Midjourney to do that.”
Many artists, however, are concerned that AI isn’t creating art, but stealing it. Due to the fact that generation tools use millions of other past and present artist’s works, many artists have grown angry that their work is being cataloged and used without their consent. One fantasy artist, Greg Rutowski, found that his name was used as a prompt in Stable Diffusion over 93,000 times, and recently, the website Have I Been Trained? allows artists to see whether or not their work has been used by an AI for its generation.
Not everything that generators are collecting is just art, though. As the tech has become more widespread, so have fears of it being used for misinformation and non-consensual depictions of celebrities and public figures. While some sites, such as DALL-E 2, attempt to prevent the creation of violent or misleading content, others have few restrictions as to what a user can create.
“Once you start filtering something, where do you stop?” Emad Mostaque, founder and CEO of the company behind Stable Diffusion, told The Verge.
This legal and ethical complexity is why sites such as Getty Images and Shutterstock have banned the sale of AI-generated images from their platforms, as well as why the U.S. copyright office has largely rejected any copyright attempts, claiming that the “human authorship” element was lacking. However, last month, the Office did grant the first known US registration for AI art on Zarya of the Dawn, a graphic novel featuring images created by Midjourney.
A large part of the online backlash has also centered on the fear that generation tools will remove work for smaller artists, especially for jobs that require numerous samples, such as concept artists. Magazines such as The Atlantic, Cosmopolitan and The Economist have already used AI art for some covers and illustrations.
“Artists kind of feel cheated because they spend so much time and effort putting intention and meaning into the best way to represent whatever subject that they want,” junior Eva Chong said. “But AI art just kind of comes along and does it in a second.”
Allen says that the ethical question of AI-generated art isn’t shouldn’t be about the existence of the technology, but on whether companies are going to use it to “save a buck.” Some artists have also stated that the use of AI could help their creative process, while not fully supplanting the human element.
“Plenty of old-fashioned artforms are still alive and well,” professional artist and Trinity graduate Julia Clift said. “People crave human expressions about the human condition. It nourishes us and makes us feel less alone in the world. I can’t imagine that ever going away.”
Much of the debate around AI-generated images has come down to philosophical arguments on what makes up art. Soul? Intention? Effort? Whatever the answer may be, the companies and users furthering this new technology aren’t waiting.
“We are in a new art movement, we are the new avant garde,” Allen said. “If this is an artwork, and I’m not an artist, then art is dead. And this is better than art.”
Through over five decades, even as AI generators became mind-numbingly complex, Harold Cohen continued tuning his machine, rule by rule. The machine’s paintings were put on display at the San Francisco MOMA, exhibited on PBS and Good Morning America, and displayed in global traveling exhibits. Cohen’s works became more intricate, often 3D and towering over the viewer. He often spoke of it in friendly terms, like a “painter’s assistant,” and when asked the meaning behind the machine’s title of ‘Aaron,’ the artist said that it wasn’t an acronym, but a name. When Cohen died in 2016, Aaron did too.
“To be realistic, I rather suspect that AARON will end when I end.” Cohen said in 2011. “Why would anybody want to take up my other half?”
