Beauty is in the eye of the beholder. But what happens when the beholder can’t see herself reflected in the giant mirror that is the internet?
This question has been at the forefront of Pinterest’s Inclusive Product team for years. Last month, the social media company launched a new search that uses AI (artificial intelligence) and machine learning to show Pinners more relevant results when looking for hair inspiration. Now, you can narrow down your hair search—think “summer hairstyle” or “braids with curls”—by six different hair patterns: protective, coily, curly, wavy, straight, and shaved/bald.
The new technology for personalized hairstyles builds on Pinterest’s efforts to establish a more inclusive product experience. For a site that’s rooted in a constant search for beauty ideas, it’s a major step in the company’s mission to increase representation for everyone online.
“As a visual platform and a place that we think of as a positive corner on the internet, it’s important for us to think about how we can inspire people no matter who they are,” says Annie Ta, who spearheads the Inclusive Product team at Pinterest. According to internal data from Pinterest, there were more than 120 million searches for “hair” and more than 5 billion hair-related Pins created this past July alone. There’s no data to show what percentage of Pinners identify as BIPOC (Black, Indigenous, and people of color), but Ta says this tool puts underrepresented groups at the center of the company’s product-making decisions. “How might we make experiences more representative online?” she asks.
The new search is powered by an underlying computer vision model that’s trained on a diverse set of hair-related images. Over time, the model learns to associate certain pixels with certain types of hair patterns. The overarching goal was to ensure AI fairness from the very first step in development.
[Image: Pinterest]While AI and algorithms are used in almost all of our technology, they can be rife with biases. In 2015, Google Photos labeled the faces of an African American couple as “gorillas.” And in May of this year, people discovered that Twitter’s image-cropping AI favors white faces over Black faces.
“We’ve learned how important it is to start with a diverse data set,” says Nadia Fawaz, an applied research scientist and the tech lead for Inclusive AI at Pinterest. Fawaz says Pinterest collected anywhere from tens of thousands to hundreds of thousands of images portraying different hair patterns, skin tones, genders, and styles.
“This is an iterative process. You look at how it performs, what errors it makes, and you feed those errors back in the model so it learns from them,” she says. For example, the algorithm had to be taught that not every spherical surface is a bald or shaved head, or that not every braided hair falls under the protective style category.
For a site like Pinterest, where beauty ideals abound, mitigating such biases is particularly important. “Beauty is subjective, and when you factor in an AI or algorithm to determine what pops up in someone’s search, you want to make sure you try to accommodate as many variables as possible,” says Dennis Schultz, executive director of the Blacks in Technology Foundation, whose mission is to bridge the gap between Black workers and the tech industry.
For Schultz, a limited or skewed definition of beauty could alter the way people see themselves without even realizing it. He says Pinterest’s endeavor is noble and innovative, as long as the quest for inclusivity extends behind the scenes. “What brought this about? How was the decision made? Were there diverse voices in the room?” he asks.
Ta says Pinterest doesn’t “keep track of everyone’s race and ethnicity on the team,” but she notes that the product came from “a lot of different backgrounds” and a wide variety of disciplines, including hairstylists like Naeemah LaFond, the global artistic director of hair-care brand Amika, who advised on the overall user experience and language used.
For now, the search tool is available in the U.S., U.K., Ireland, Canada, Australia, and New Zealand. The Pinterest team is working on rolling it out to other markets, where cultural nuances may dictate the need for more hair patterns and different definitions of beauty. Ta says Pinterest gets 80 million beauty-related searches monthly. “Beauty is so much about the physical representation of people,” she says. “We want everyone to be seen.”