Getting machines to make AI art is sort of like trying to present a fake ID to a bartender at 10 p.m. on a Saturday: If it's good enough and the bartender's busy enough, he might let it pass.

Share story

Robbie Barrat was 17 and bored in West Virginia when he started experimenting with artificial intelligence and art.

First, he trained a computer to write original rap songs by feeding it 6,000 Kanye West lyrics. Then, he taught it how to make landscape paintings and nude portraits by feeding it thousands of images scraped off the internet. He uploaded the code to GitHub, the code-sharing platform, so that others could download it and learn from it.

And many did — including the French art collective known as Obvious, whose AI portrait sold for $432,500 at Christie’s last week in an internationally celebrated art auction. That’s nearly 45 times higher than Christie’s original estimated sale price of $10,000, the auction house said in an article on its website.

“I was really expecting people to use (the code) as components for their own project. But I never thought anybody would sell it, just because it’s not high-quality work,” said Barrat, now 19 and working at a Stanford University AI research lab. “It was a project I did in my free time when I was 17.”

Obvious, made up of three 25-year-old students, made the AI portrait using an existing algorithm and Barrat’s code, among other things, as the group acknowledged in an in-depth article by The Verge last week.

Titled “Portrait of Edmond Belamy,” the artwork was the subject of wide attention for its unconventionality in a major auction house such as Christie’s, which lined it up in the same room as work by Warhol and Lichtenstein and heralded it for its novelty. The portrait emits a quality known in aesthetics as the uncanny valley: that eerie feeling you get when you look at something that resembles a human, but isn’t quite there. The man in the portrait has no nose. He has a blob for a mouth and his eyes resemble that of Frosty the Snowman. It appears, as its bottom-right signature indicates, to be the work of an algorithm.

But for all its apparent peculiarity, the relatively small AI art community has described it as utterly conventional, lagging behind advancements by other artists in the AI art world. Mario Klingemann, a Germany-based AI artist whom Obvious has cited as an inspiration, said he was shocked that it sold for hundreds of thousands of dollars. He said he believed for a second that “maybe this is just a practical joke among oligarchs.”

“It’s horrible art from an aesthetic standpoint,” he said. “You have to put some work into it to call it art. It’s something that everybody can do. You can clone this (code) from GitHub, start your computer and start doing it. I don’t know that that’s what art is about. You have to put your own handwriting on it, make your own mark with these tools. That takes some learning and work and finding something different to say.”

Obvious has never denied that their work has relied on the innovations made by others. In fact, the name of the painting itself is a reference to the inventor of the algorithm who has made this entire field of AI art possible, a Google Brain AI research scientist named Ian Goodfellow. “Belamy,” the name of the fictional man in the portrait, is a play on “Bel ami,” which means “good friend” in French.

Obvious could not be reached for comment after the auction, but in an interview with the Verge, Obvious co-founder Hugo Caselles-Dupré confirmed elements were borrowed from Barrat, who said he began corresponding on GitHub with the group in October 2017. “If you’re just talking about the code, then there is not a big percentage that has been modified,” Caselles-Dupré said. “But if you talk about working on the computer, making it work, there is a lot of effort there.”

Using Goodfellow’s algorithm, getting machines to make AI art is sort of like trying to present a fake ID to a bartender at 10 p.m. on a Saturday: If it’s good enough and the bartender’s busy enough, he might let it pass.

It’s the same thing with the computer networks: One computer network basically tries to deceive a second computer that the image it generated can actually pass as real art. It’s like the Turing test, except both participants are machines.

To train the first computer network to produce a portrait, for example, you have to feed it hundreds or thousands of images of portraits so that it learns what a good portrait looks like. Then, the second computer, called the discriminator, plays the referee and decides if it passes the smell test. If it doesn’t, like if a bouncer confiscates a fake ID, the first computer has to “go back to the drawing board” and must try harder at producing convincing art, Klingemann said.

Once it succeeds, the result is something usually bizarre and abstract, and perhaps even imaginative.

“(The computers) learn from the ground up,” Klingemann said. “Initially both parties don’t know what criteria to look for. It’s only over time, in this feedback loop, that they improve their capabilities of spotting the differences. Eventually it becomes so good that even a human cannot distinguish it from fake or real anymore.”

Obvious told Christie’s that it fed the system 15,000 portraits painted between the 14th and 20th century as part of its “training” for the computer.

In a statement following the sale of  “Edmond Belamy,” Obvious tipped their hats to Goodfellow and Barrat.

“We would like to thank the A.I. community, especially to those who have been pioneering the use of this new technology, including Ian Goodfellow, the creator of the GAN algorithm, who inspired the name of the Famille de Belamy series, and artist Robbie Barrat, who has been a great influence for us. It is an exciting moment and our hope is that the spotlight on this sale will bring forward the amazing work that our predecessors and colleagues have been producing.”