A.I.-Generated Art Has Crossed the Uncanny Valley

Last month, a piece of art called Théâtre D’opéra Spatial (that’s French for “Space Opera Theater”) was entered into the Colorado State Fair’s fine art competition by a man named Jason Allen. The piece is a gorgeous “painting” that depicts a giant baroque hall containing three women in flowing red and white robes. The image won first place in the digitally manipulated photography category, and the artist judges at the state fair said the work was the best of the best.

At the fair, Allen said the piece was made with Midjourney, an artificial intelligence tool that can create art, but no one really understood what that meant. Once they realized an A.I.-generated piece of art had beaten human artist-created images, a debate opened up. When A.I. makes beautiful art, is it somehow a lie, or is it the paintbrush of the future? On Friday’s episode of What Next: TBD, I spoke with Drew Harwell, a tech reporter covering A.I. for the Washington Post, about how the A.I. art boom is reshaping the artistic economy and maybe even the meaning of art. Our conversation has been edited and condensed for clarity.

Lizzie O’Leary: If you look around online, A.I. art suddenly feels like it’s everywhere, from weird memes to altered versions of the Mona Lisa. Multiple A.I. tools can make these images—DALL-E, Stable Diffusion, and Midjourney, to name a few—but they all use a similar process. How does that process work?

Drew Harwell: So they take a bunch of photographs and pieces of art and everything we see on the web, they put it into this big A.I. model and say, “OK, let’s make something new.” They take what they’ve seen before, learn the patterns, learn the styles and the colors, learn how they’re attached to the text given, and they create something. The end result is effectively just a little program that anybody can use.

I’m struck by how quickly these art A.I.s are advancing. DALL-E was released in January of last year and there were some pretty basic images. And then, a year later, DALL-E 2 is using complex, faster methods. Midjourney, the one Jason Allen used, has a feature that allows you to upscale and downscale images. Where is this sudden supply and demand for A.I. art coming from?

You could look back to five years ago when they had these text-to-image generators and the output would be really crude. You could sort of see what the A.I. was trying to get at, but we’ve only really been able to cross that photorealistic uncanny valley in the last year or so. And I think the things that have contributed to that are, one, better data. You’re seeing people invest a lot of money and brainpower and resources into adding more stuff into bigger data sets. We have whole groups that are taking every image they can get on the internet. Billions, billions of images from Pinterest and Amazon and Facebook. You have bigger data sets, so the A.I. is learning more. You also have better computing power, and those are the two ingredients to any good piece of A.I. So now you have A.I. that is not only trained to understand the world a little bit better, but it can now really quickly spit out a very finely detailed generated image.

Is there any way to know, when you look at a piece of A.I. art, what images it referenced to create what it’s doing? Or is it just so vast that you can’t kind of unspool it backward?

When you’re doing an image that’s totally generated out of nowhere, it’s taking bits of information from billions of images. It’s creating it in a much more sophisticated way so that it’s really hard to unspool.

Art generated by A.I. isn’t just a gee-whiz phenomenon, something that wins prizes, or even a fascinating subject for debate—it has valuable commercial uses, too. Some that are a little frightening if you’re, say, a graphic designer.

You’re already starting to see some of these images illustrating news articles, being used as logos for companies, being used in the form of stock art for small businesses and websites. Anything where somebody would’ve gone and paid an illustrator or graphic designer or artist to make something, they can now go to this A.I. and create something in a few seconds that is maybe not perfect, maybe would be beaten by a human in a head-to-head, but is good enough. From a commercial perspective, that’s scary, because we have an industry of people whose whole job is to create images, now running up against A.I.

And the A.I., again, in the last five years, the A.I. has gotten better and better. It’s still not perfect. I don’t think it’ll ever be perfect, whatever that looks like. It processes information in a different, maybe more literal, way than a human. I think human artists will still sort of have the upper hand in being able to imagine things a little more outside of the box. And yet, if you’re just looking for three people in a classroom or a pretty simple logo, you’re going to go to A.I. and you’re going to take potentially a job away from a freelancer whom you would’ve given it to 10 years ago.

I can see a use case here in marketing, in advertising. The A.I. doesn’t need health insurance, it doesn’t need paid vacation days, and I really do wonder about this idea that the A.I. could replace the jobs of visual artists. Do you think that is a legitimate fear, or is that overwrought at this moment?

I think it is a legitimate fear. When something can mirror your skill set, not 100 percent of the way, but enough of the way that it could replace you, that’s an issue. Do these A.I. creators have any kind of moral responsibility to not create it because it could put people out of jobs? I think that’s a debate, but I don’t think they see it that way. They see it like they’re just creating the new generation of digital camera, the new generation of Photoshop. But I think it is worth worrying about because even compared with cameras and Photoshop, the A.I. is a little bit more of the full package and it is so accessible and so hard to match in terms. It’s really going to be up to human artists to find some way to differentiate themselves from the A.I.

This is making me wonder about the humans underneath the data sets that the A.I. is trained on. The criticism is, of course, that these businesses are making money off thousands of artists’ work without their consent or knowledge and it undermines their work. Some people looked at the Stable Diffusion and they didn’t have access to its whole data set, but they found that Thomas Kinkade, the landscape painter, was the most referenced artist in the data set. Is the A.I. just piggybacking? And if it’s not Thomas Kinkade, if it’s someone who’s alive, are they piggybacking on that person’s work without that person getting paid?

I’m really sympathetic to that argument, actually, because some people could call this plagiarism. This is a data set that literally includes images of artists like the Painter of Light. Thousands of his images are in that data set and when the A.I. wants to make something that it thinks needs to look like that, it’s going to look at those images and infer things from them. And not just people like him, but living artists who see this as a job, their images are feeding in the machine that is taking their jobs.

Jason Allen’s argument, and the A.I. booster’s argument, is: How did you learn how to paint? You looked at a bunch of other people’s paintings, right? You went to art school. You learned the techniques of the masters who came before you. Isn’t this how we all learn? We look at the creative product of other people and we copy over time. It’s flattery. We duplicate the things we’ve seen before. That’s what A.I. is doing. It’s just doing it way better and at a scale of billions of times more than we do it.

There’s no question that A.I. can make compelling and beautiful images. But when you start to think beyond technical skill to the emotional resonance or cultural commentary contained in great art, a computer seems destined to fall far short of a human. I’m thinking of the artist Kara Walker. A lot of her art draws on hyperracialized, hypersexualized images of Black people and Black women. I could very quickly see an A.I. trained on Kara Walker’s work, not having the commentary, not being subtle or interesting or critiquing in that way and just popping out some racist images.

Totally, and that’s where the humans get back the upper hand. We can understand things about cultural context and these broader, mushier sort of social debates that the A.I. is just nowhere close to understanding. The problem with these A.I.s is that they’re trained off the internet and you’ve probably been on the internet, you’ve seen it’s not all sunshine and rainbows. There’s a lot of bad crap on the internet.

This has happened with chatbots, text A.I., where you train it off the seedier parts of Reddit or 4chan, and voila, it starts to become this neo-Nazi weirdo because it’s learning to copy what it’s seen. And so for A.I. art, it’s always going to be a big but finite view of the world. Right now, it’s just doing a really good job of mirroring other images.

Jason Allen seemed to take the stance that A.I.-generated art is the future. How did that go down with people who were judging it and also people who have reacted to this piece of work on the internet?

Some people were very upset about it. They felt there was deception. They felt there was plagiarism. They felt like he had stolen from these human artists who had worked in the “right way.” Then, I talked to the judges who gave him the win, I talked to one of the artists that he beat, who won third place from hand-drawing something on an iPad, and I talked to an oil painter who does these gorgeous paintings and who could potentially be out of a job in the future because of this. I expected them to all hate it, but they surprised me by saying, “Actually, I think it’s pretty cool.” They felt like, “Who are we to say what is art and what is not art?” They all felt like, “Well, maybe we should have a category for A.I. art in the future so it’s not competing with human art.” The beauty of a piece of artwork is what we, as the human viewers, get out of it. Does it matter that it’s by a person’s hand, and if it does, how far do you have to go?

The pro-A.I. argument is that art is always evolving, and so are the tools that humans use to make it—they’re just different depending on the era.

A hundred or 200 years ago, there was a view that even photography was not real art, it was just somebody hitting a button on a camera capturing what they saw in real life, and now we understand that the best photography is about composition and color and style and tone, and all of these human subjective judgements that Jason Allen would say he made in his own piece of art. He said he made a ton of different iterations: He went through and there were a lot of bad creations and he chose the one that he felt was the most evocative.

Not to get too philosophical, but these tools do raise the question of what makes art, art? And who gets to define that anyway? In your story, you quoted one of the judges basically saying she didn’t know that this was A.I.-generated art but she wouldn’t have changed her judgment—that there was a concept and a vision, he brought it to reality, and it’s a beautiful piece. So is art, however we use that word, ultimately in the eye of the beholder?

Art is what we take from it. This judge, who is an artist herself and an art historian, she did not feel deceived. She felt like the emotion she first felt when she saw it was still real. It really caught her eye when she was walking through the state fair, looking at a lot of great pieces of art. She said it told a story and it made her question, “What are these figures in this painting looking at? What is the deeper meaning to any of this?” She didn’t feel like there was some bait and switch. She felt like, “Oh, this is a pretty cool image, and I don’t exactly know how it was made, but it hit something in me.” And that’s really all she was looking for when she was looking at art.