Mustaches, Unibrows, and Shalwar Khameezes: How I Learned New Stereotypes about Myself through AI
Recently, The Verge posted a story about attempting to create an Asian man with a white woman If you read the original article, you will find that the author struggled to generate the images with image generators attempting to give the woman “Asian features.” Because the majority of these image generators are trained on datasets that predominantly feature white individuals, the AI struggled to accurately represent an Asian man without relying on stereotypes and tropes. In addition, stereotypical images of Asian men are prominent throughout the Internet.
At one point, Meta banned keywords that were related to Asians. Likewise, Google paused Gemini’s ability to generate images of people after concerns about diversity. There are some attempts at fixing this bias through the use of “Diversity Fine-tuning”:
However, this requires that there be a significant amount of photos of the group in question. Since I grew up as a Pakistani in Mississippi, I didn’t have high hopes that there would be a large dataset of photos of people who looked like me. Out of curiosity, I decided to try generating images of myself using various AI image generators. The results were… interesting, to say the least. Despite the fact that African Americans make up almost thirty-eight percent of Mississippians, Mississippi’s cultural image is predominantly white. Usually, any Southern state is conflated with Texas, so you will see cowboy boots and hats even though no one I know wore these. I assumed that adding Mississippi to Pakistani would create results that relied on these tropes.
I decided to use a simple prompt of “Pakistani boy from Mississippi” from nine different AI image generators. While some were able to create pictures of Pakistani boys—and one even put him in a polo shirt—the majority relied on stereotypes. Most of the pictures placed the boys in Shalwar Kameez, something I have never seen a Pakistani child wearing in Mississippi outside of cultural events. What was even more strange were stereotypes that I did not even know existed. These were largely oriented around facial hair. The boys depicted often have unibrows that are well beyond what most children have, and others have full mustaches. I guess it goes without saying that the latter is not something I have ever known to exist and the idea of “hairy” Pakistani children was not a particularly prominent trope that I heard growing up. Nonetheless, it is interesting to see how the AI has internalized these stereotypes and applied them to the generated images.
Another interesting observation was the lack of diversity in the generated images. Despite specifying “Pakistani boy from Mississippi,” the AI seemed to struggle with generating images that accurately represented the cultural and regional context. The images lacked any visual cues that would indicate the boy’s connection to Mississippi, such as clothing or surroundings that are typical of the American South. Instead, the AI relied heavily on stereotypical markers of Pakistani identity, such as traditional clothing and exaggerated facial hair features.
Below, you can observe the results that I got for the various image generators: