Tech

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

At the end of September, OpenAI widely distributed its DALL-E 2 AI art generator. available to the public, allowing anyone with a computer to capture one of those startling, slightly whimsical images that seem to be circulating more and more across the Internet these days. DALL-E 2 is by no means the first AI art generator open to the public (competing AI art models stable diffusion as well as middle of the road also launched this year), but it has a solid pedigree: its cousin, the text generation model known as GPT-3, is itself the subject of much intrigue and several useless stories – was also developed by OpenAI.

Microsoft last week announced it will be adding DALL-E 2-based AI-powered art tools to its Office software suite, and in June DALL-E 2 was used to develop cover of Cosmopolitan magazine. The most techno-utopian proponents of AI-generated art say it provides a democratization of art for the masses; the cynics among us will argue that this is copying human artists and threatening to end their careers. In any case, it seems clear that the art of AI is already here and its potential has just begun to be explored.

Naturally, I decided to try.

When I looked at examples of DALL-E work for inspiration (I decided that my first attempt should be a masterpiece), it seemed to me that the art created by AI did not have any special aesthetic, except maybe a little strange. Pigs in sunglasses and floral shirts rode motorcycles, raccoons played tennis, and Johannes Vermeer Girl with a pearl earring, slightly modified to replace the titular sea otter girl. But as I continued to scroll, I realized that at the heart of every piece is one unifying theme: AI art is most often similar to Western art.

“All AI only looks back,” said Amelia Winger-Berskin, professor of artificial intelligence and the arts at the Digital Worlds Institute at the University of Florida. “They can only look into the past, and then they can predict the future.”

For an AI model (also known as an algorithm), the past is the dataset it was trained on. For an AI art model, this dataset is art. And the fine art world is dominated by white Western artists. This results in AI-generated images that look overwhelmingly Western. To be honest, this is a little disappointing: AI-generated art could, in theory, be an incredibly useful tool for creating a fairer vision of art that is very different from what we used to take for granted. Instead, it’s worth simply perpetuating the colonial ideas that define our understanding of art today.

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

DALL-E Interpretation 2 Clues “Hokusai Draws Artificial Intelligence”
Neil Dhanesha/Vox; Provided by OpenAI

Winger-Bearskin saw firsthand the limits of the art of AI. When one of her students used images generated with Stable Diffusion to create a nature video, she realized that the twilight backgrounds produced by the AI ​​model looked strangely like scenes drawn by Disney animators in the 1950s and 60s. years, which in themselves were inspired French Rococo movement. “There are a lot of Disney movies out there and what he brought back was something we see a lot,” Winger-Bearskin told Recode. “So many things are missing from these datasets. There are millions of night scenes from all over the world that we will never see.”

AI bias is a notoriously difficult problem. If left unchecked, algorithms can perpetuate racist and sexist biases, and this bias extends to AI art: As Segal Samuel wrote for Future Perfect in April, previous versions of DALL-E produced images of white men when asked to pose as lawyers. for example, and depict all flight attendants as women. OpenAI was working to mitigate these effects by tweaking their model to try and weed out stereotypes, though researchers still disagree about whether these measures worked.

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

“You can’t fine-tune the model to be less western if your dataset is mostly western” Ilong Du, a graduate student and artificial intelligence researcher at MIT, said Recode. AI models are trained by looking up images on the Internet, and Du believes models built by teams based in the United States or Europe are likely to be biased towards Western media. Some models made outside the United States, such as the ERNIE-ViLG developed by Chinese technology company Baidu, are better at producing images that are more in line with the culture of their place of origin, but they have their own problems; how It is reported by MIT Technology Review. in September, ERNIE-ViLG is better at drawing anime than DALL-E 2, but refuses to make images of Tiananmen Square.

Because the AI ​​is past-oriented, it can only create variations on images it has seen before. That’s why, according to Du, the AI ​​model can’t create an image of a plate on a fork, even though it needs to understand every aspect of the request. The model has simply never seen a picture of a plate on top of a fork, so instead it outputs pictures of forks on top of plates.

Adding more non-Western art to the existing dataset would also not be a very useful solution due to the overwhelming prevalence of Western art on the internet. “It’s like giving clean water to a tree that’s been fed polluted water for the past 25 years,” Winger-Bearskin said. “Even if the water is getting better now, the fruits from this tree are still infected. Running the same model with new training data does not significantly change it.”

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

This is a labor intensive process. “The hardest part is creating the dataset,” Du said. It takes millions of images to train an AI arts generator, and Du said it would take months to create a dataset that is equally representative of all art styles found around the world.

If the artistic bias inherent in most AI art models has a positive side, then perhaps it is this: like all good art, it reveals something about our society. According to Winger-Bearskin, many contemporary art museums have more space for art created by people from underrepresented communities than in the past. But this art is still only a small fraction of what exists in the museum archives.

“The job of an artist is to talk about what is happening in the world, to highlight problems so that we notice them,” he said. Jin Oh, adjunct professor at the Robotics Institute at Carnegie Mellon University. AI art models can’t comment on their own—everything they create is human-made—but the art they create creates a kind of random meta-commentary that O thinks deserves attention. “It gives us the opportunity to observe the world the way it works, and not the ideal world we want it to be.”

This is not to say that O believes that fairer models should not be created—they are important for circumstances where depicting an idealized world is useful, such as for children’s books or commercial applications, she told Recode—but rather, the existence of imperfect models should push us to deeper reflections on how we use them. Instead of just trying to eliminate biases as if they don’t exist, Oh said, we should take the time to identify and quantify them so that we can constructively discuss their impact and how to minimize it.

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

This story was first published in the Recode newsletter. Sign here so you don’t miss the next one!


Source link

Leave a Reply

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

Back to top button