AI models like DALL-E 2 continue to make art that looks way too European

At the end of September, OpenAI made its DALL-E 2 AI art generator widely available to the public, allowing anyone with a computer to create one of those striking and slightly bizarre images that seem to float more and more on Internet these days. DALL-E 2 is by no means the first AI art generator to be released to the public (competing AI art models Stable Diffusion and Midjourney also launched this year), but it does come with a solid pedigree: its cousin, the text generator model. known as GPT-3 – itself subject to many intrigues and multiple fanciful stories – was also developed by OpenAI.

Last week, Microsoft announced that it would add AI-generated art tools – powered by DALL-E 2 – to its Office software suite, and in June DALL-E 2 was used to design the magazine cover. Cosmopolitan. The more techno-utopian proponents of AI-generated art say it provides a democratization of art for the masses; the cynics among us would say it’s copying human artists and threatening to end their careers. Either way, it seems clear that the art of AI is out there and its potential is only just beginning to be explored.

Naturally, I decided to try it.

As I browsed through examples of DALL-E’s work for inspiration (I had determined that my first attempt had to be a masterpiece), it struck me that the AI-generated art n didn’t have any particular aesthetic other than, perhaps, being a little weird. There were pigs wearing sunglasses and flowery shirts while riding motorcycles, raccoons playing tennis, and Johannes Vermeer A girl with an earringslightly modified in order to replace the titular girl with a sea otter. But as I continued to scroll, I realized that there was a unifying theme underlying each piece: AI art, the more often resembles Western art.

“All AI is just retrospective,” said Amelia Winger-Bearskin, professor of AI and the arts at the University of Florida’s Digital Worlds Institute. “They can only look at the past, then they can make a prediction of the future.”

For an AI model (also known as an algorithm), the past is the data set it was trained on. For an AI art model, this data set is the art. And much of the fine art world is dominated by white Western artists. This leads to AI-generated images that look mostly Western. It’s, frankly, a little disappointing: AI-generated art, in theory, could be an incredibly useful tool for imagining a more equitable view of art that looks very different from what we take for granted. Instead, it is simply to perpetuate the colonial ideas that drive our understanding of art today.

To be clear, models like DALL-E 2 can be asked to generate art in the style of any artist; requesting an image with the “Ukiyo-e” modifier, for example, will create works that mimic Japanese woodblock prints and paintings. But users must include these modifiers; they are rarely, if ever, the default.

DALL-E 2’s interpretation of the “Hokusai painting of Artificial Intelligence” prompt
Neel Dhanesh/Vox; Courtesy of OpenAI

Winger-Bearskin saw firsthand the limits of AI art. When one of her students used images generated by Stable Diffusion to make a video of a nature scene, she realized that the twilight backgrounds produced by the AI ​​model looked suspiciously like the scenes painted by Disney animators in the 1950s and 1960s – who themselves had been inspired by the French Rococo movement. “There are a lot of Disney movies, and what he got back is something we see a lot,” Winger-Bearskin told Recode. “There is so much missing from these datasets. There are millions of night scenes from around the world that we would never see.

AI bias is a notoriously difficult problem. Unchecked, algorithms can perpetuate racist and gender bias, and that bias also extends to AI art: as Sigal Samuel wrote for Future Perfect in April, previous versions of DALL-E spit images of white men when asked to represent lawyers, for example, and depict all flight attendants as women. OpenAI has worked to mitigate these effects, refining its model to try to eliminate stereotypes, though researchers still disagree on how effective these measures are.

But even if they work, the art style problem will persist: if DALL-E manages to portray a world free of racist and sexist stereotypes, it will still do so in the image of the West.

“You can’t refine a model to be less Western if your dataset is mostly Western,” Yilun Du, a PhD student and artificial intelligence researcher at MIT, told Recode. AI models are trained by scraping the internet for images, and Du thinks models created by groups based in the United States or Europe are likely predisposed to Western media. Some models made outside of the US, like ERNIE-ViLG, which was developed by Chinese tech company Baidu, do a better job of generating images that are more culturally relevant to their place of origin, but they come with their own problems ; as MIT Technology Review reported in September, ERNIE-ViLG is better at producing animated art than DALL-E 2 but refuses to do Tiananmen Square footage.

Because the AI ​​is retrospective, it can only make variations of images it has seen before. This is why, according to Du, an AI model is incapable of creating an image of a plate on a fork, even though it should in theory understand every aspect of the request. The model has simply never seen an image of a plate on a fork, so it spits out images of forks on plates instead.

Injecting more non-Western art into an existing dataset wouldn’t be a very helpful solution either, due to the overwhelming prevalence of Western art on the internet. “It’s a bit like giving clean water to a tree that has been fed contaminated water for the past 25 years,” Winger-Bearskin said. “Even though the water is improving now, the fruits of this tree are still contaminated. Running this same model with new training data does not change it significantly. »

Instead, creating a better, more representative AI model would require creating it from scratch — which is what Winger-Bearskin, who is a member of the Seneca-Cayuga Nation of Oklahoma and an artist herself, does. when she uses AI to create art about the climate crisis.

It is a process that takes time. “The hardest part is creating the dataset,” Du said. Training an AI art generator requires millions of images, and Du said it would take months to create a dataset that is equally representative of all the art styles that can be found in the world. world.

If there’s one benefit to the artistic bias inherent in most AI art models, it might be this: like all good art, it exposes something about our society. Many modern art museums, Winger-Bearskin said, give more space to art made by people from underrepresented communities than they did in the past. But this art still only represents a small fraction of what exists in museum archives.

“An artist’s job is to talk about what’s going on in the world, to amplify problems so that we notice them,” said Jean Oh, associate research professor at Carnegie University’s Robotics Institute. Mellon. AI art models are unable to provide their own feedback – everything they produce is at the request of a human – but the art they produce creates a sort of accidental meta-commentary that, according to Oh, is worth noticing. “It gives us a way to observe the world as it is structured, not the perfect world we want it to be.”

That’s not to say Oh thinks more equitable models shouldn’t be created — they’re important in circumstances where representing an idealized world is useful, such as for children’s books or commercial applications, he said. – she told Recode – but rather that the existence of imperfect models should cause us to think more deeply about how we use them. Instead of just trying to eliminate biases as if they didn’t exist, Oh said, we should take the time to identify and quantify them in order to have constructive discussions about their impacts and how to minimize them.

“The main goal is to help human creativity,” said Oh, who researches ways to create more intuitive human-AI interactions. “People want to blame the AI. But the final product is our responsibility.

This story was first published in the Recode newsletter. register here so as not to miss the next one!

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *