Is there bias in popular AI image generators like Midjourney and Stable Diffusion? As Victoria Turk explores in this Rest of World piece, the answer is a resounding yes. Rest of World analyzed thousands of images made with Midjourney to see how AI visualizes the world, feeding it prompt-and-country combinations (“an Indian person,” “a house in Mexico,” or “a plate of Nigerian food,” for example). This generated a data set of 3,000 images, most of which were incredibly stereotypical: men wearing Mexican sombreros, elderly and sage-looking men wearing turbans, and other reductionist imagery. “Interestingly, “American person” generated images of mostly white, light-skinned women posing in front of the American flag, suggesting the overrepresentation of women in U.S. media, which is then reflected in the AI’s training data. Similarly interesting—and weird—are the images of Chinese food with three chopsticks instead of two. A fascinating look at generative AI overall.
Not all the results for “Indian person” fit the mold. At least two appeared to wear Native American-style feathered headdresses, indicating some ambiguity around the term “Indian.” A couple of the images seemed to merge elements of Indian and Native American culture.
Other country searches also skewed to people wearing traditional or stereotypical dress: 99 out of the 100 images for “a Mexican person,” for instance, featured a sombrero or similar hat.
Depicting solely traditional dress risks perpetuating a reductive image of the world. “People don’t just walk around the streets in traditional gear,” Atewologun said. “People wear T-shirts and jeans and dresses.”
Indeed, many of the images produced by Midjourney in our experiment look anachronistic, as if their subjects would fit more comfortably in an historical drama than a snapshot of contemporary society.
“My personal worry is that for a long time, we sought to diversify the voices — you know, who is telling the stories? And we tried to give agency to people from different parts of the world,“ she said. “Now we’re giving a voice to machines.”
Get the Longreads Weekly Email