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

In late September, OpenAI unveiled its DALL-E 2 AI art generator to the wider public. This allowed anyone with a computer to create one of those striking and slightly bizarre images of him that seems to be floating around the internet more and more these days. DALL-E 2 is by no means the first AI art generator released to the public (competing AI art models Stable Diffusion and Midjourney were also launched this year). Known as GPT-3 (itself the subject of many conspiracy and multiple gimmick stories) was also developed by OpenAI.

Last week, Microsoft announced the addition of DALL-E 2-powered AI-generated art tools to its Office software suite, and in June it used the DALL-E 2 to design the cover of Cosmopolitan magazine. Most techno-utopian proponents of AI-generated art say it offers the democratization of art for the masses. The cynics among us will argue that it is copying human artists and threatening to end their careers. , the possibilities are just beginning to be explored.

Naturally, I decided to give it a try.

Scrolling through examples of DALL-E’s work for inspiration (I decided my first attempt had to be a masterpiece), the AI-generated art probably had a little Strange.A pig in sunglasses and a floral shirt riding a motorcycle, a raccoon playing tennis, by Johannes Vermeer girl with a pearl earring, tweaked it slightly to replace the girl in the title with an otter. AI art often looks like Western art.

Amelia Winger-Bearskin, Professor of AI and Arts at the University of Florida’s Digital World Institute, said: “They can only see the past, they can predict the future.”

For AI models (also called algorithms), the past is the data set they were trained on. For AI art models, that data set is art. And much of the art world is dominated by white Western artists. Frankly, this is a little disappointing. AI-generated art could, in theory, be a very useful tool for imagining a more equitable vision of art that is very different from what we take for granted. . Instead, it merely perpetuates the colonial ideas that drive our understanding of art today.

To clarify, models like the DALL-E 2 can be asked to create art in the style of any artist. For example, commissioning an image using the qualifier “Ukiyo-e” creates a work that mimics Japanese woodblock prints and paintings. However, the user must include these modifiers. They are rarely the default.

Neel Dhanesha/Vox; Courtesy of OpenAI

Winger-Bearskin has seen the limits of AI art. When one of her students made a video of a nature scene using images generated by Stable Diffusion, the twilight background created by the AI ​​model was used in the 1950s and her 60s. I noticed an odd resemblance to scenes drawn by Disney animators. Influenced by the French Rococo movement. “There are a lot of Disney movies, but what he brought back was something we see a lot,” Winger Bearskin told his Recode. “These datasets are missing a lot. There are millions of nightscapes around the world that we will never see.”

AI bias is a very difficult problem. Left unchecked, algorithms can perpetuate racist and sexist biases that extend into AI art. As Sigal Samuel wrote in his April in his Future Perfect, previous versions of DALL-E spat out images of white men when asked to portray a lawyer. For example, portray all flight attendants as female. OpenAI is working to mitigate these effects, fine-tuning its models to eliminate stereotypes, but researchers are still divided on whether those measures worked.

But even if they worked, it wouldn’t solve the problem of artistic style. Even if DALL-E were able to portray a world without stereotypes of racism and sexism, it would be portrayed in a Western image.

MIT PhD student and AI researcher Yilun Du told Recode: AI models are trained by collecting images from the internet. Du believes that models created by groups based in the United States or Europe are more susceptible to Western media. Some models made outside the US, like his ERNIE-ViLG, developed by Chinese tech firm Baidu, are better at producing images that are more culturally relevant to their country of origin, but their own I have a problem with As MIT Technology Review reported in his September, ERNIE-ViLG is better at creating anime art than his DALL-E 2, but refuses to create images of Tiananmen Square.

The AI ​​is backward-looking, so it can only create variations of previously seen images. According to Du, this is why the AI ​​model needs to understand each aspect of the request, but cannot create an image of a plate on a fork. The model has never seen an image of a plate on a fork, so it spits out an image of a fork on a plate instead.

Injecting non-Western art into existing datasets is also a less profitable solution. This is because Western art is overwhelmingly popular on the Internet. “It’s like giving clean water to a tree that has been given polluted water for the last 25 years,” says Winger his bearskin. “Even though the water quality has improved, the berries are still contaminated. Running the same model with new training data doesn’t change much.”

Instead, to create better, more representative AI models, you have to start from scratch. This is what Winger-Bearskin, a member of his nation Seneca Cayuga of Oklahoma and an artist himself, does when he uses AI to create art. climate crisis.

It’s a time consuming process. “The hardest part is creating the dataset,” he says. Millions of images are required to train an AI art generator. Du says it will take months to create a data set that is equally representative of all the art styles found around the world.

If there’s an advantage to the artistic bias inherent in most AI art models, it’s probably this. Like all great art, it reveals something about our society. According to Winger Bearskin, many contemporary museums are giving more space than before to art made by people from underrepresented communities. However, this art is only a fraction of what still exists in the museum’s archives.

“An artist’s job is to talk about what’s going on in the world, magnify the problem, and make us aware of it,” said Jean Oh, an associate professor at the Robotics Institute at Carnegie Mellon University. An art model cannot provide its own commentary: everything AI-generated is at the dictates of humans, but AI-generated art creates a kind of accidental meta-commentary that draws attention to itself. Oh thinks it deserves: “It gives us a way to see the world in a structured way, not the perfect world we want.”

That’s not to say that Oh believes we shouldn’t create fairer models — they are important in situations where portraying an idealized world is useful, such as for children’s books or commercial applications. Yes, she told Recode — rather, the existence of imperfect models should prompt more thought about how to use them. Rather than simply trying to eliminate biases as if they don’t exist, we should take the time to identify them, quantify them, and have constructive discussions about their impact and how to minimize them. said Mr Oh.

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

This story was first published in the Recode newsletter. SIGN UP HERE Don’t miss the next one!