In the 1970s, when researchers asked children to draw a scientist, 99 percent of them drew a man. As this experiment was repeated over 50 years, the number of women drawn increased, and within the past decade, more than half of girls will draw a woman when asked what a scientist looks like.
Today, Google search results tend to agree with these children’s drawings. Type in an occupation, and you’ll be met by a wall of stock photos that tell a story of gender parity across many professions. But despite the apparent gender equality, the results capture other elements of societal expectations of women. A close observer might notice that many of the women depicted seem rather young—and the men tend to appear older.
It’s no secret that women are encouraged—by advertisers, popular media, and well-meaning comments of “you don’t look your age”—to appear youthful. Aging is often depicted as a negative for women, while older men are regarded as wiser and more experienced. Most women have encountered this personally, but data on the phenomenon has been scarce.
To find exactly how widespread this bias is, a team of Berkeley researchers surveyed images and text across some of the most well-trafficked places on the internet, such as ChatGPT, IMDb, Google, and Wikipedia and found that women are regularly depicted as younger than men—and devalued because of it.
This in itself didn’t surprise Solène Delecourt, an assistant professor at UC Berkeley’s Haas School of Business, and a co-author of the study published in Nature on Wednesday, but “the effects we see are much, much broader, and potentially carry effects in the labor market for women at a scale that was maybe more than I even expected,” she said.
The researchers analyzed over half a million images from Google search, in which women consistently appeared younger than men. This serves as a measure of cultural bias because “it’s basically trying to give you content that you’re most likely to click on,” added co-author Douglas Guilbeault, who is now an associate professor at Stanford. “That has a way of being prone to bias, because it ends up just amplifying whatever most people click on.”
Across the internet, women are most commonly shown in their 20s, while men are usually shown in their 40s and 50s. And it’s not just that women in the images look younger. Often, they are younger. On IMDb and Wikipedia, the researchers were able collect information about the actual ages of people in the photos, and this trend persisted.
“If you have biased data going in, you will in all likelihood replicate the bias. And we see this again and again.”
This reflects hiring biases in the entertainment industry, Guilbeault noted, but these are also the most visited pages. For the people looking up Hollywood profiles on IMDb, “the influence and attention is biased towards older men.”
In both online images and text, the researchers found similar, skewed depictions of men and women in thousands of occupations and other categories. But in census data, across most of the fields examined, there were no age differences among men and women. In the few professions where an age gap existed, the women were older, on average, than the men. But the online images presented an inverted picture.
“The pattern we see in the data just does not match reality,” Delecourt said. “The average woman in the US, and actually in the world, has a higher life expectancy. The average woman is older, so what we see in…online images and text and videos is wrong.”
This age-gap myth also affects how people view women in the workplace. As a part of the study, the researchers asked participants to find photos of people working in different professions. When the participants selected photos of women, they assumed that people with that job were generally younger and had less experience.
By influencing people’s perceptions, this pervasive imagery can have real-world consequences in hiring decisions. And the online age gaps were most extreme for higher-status and higher-earning positions, potentially contributing to the gender pay gap, which is more pronounced among people with post-graduate and professional degrees.
Online data is also used to train AI, which perpetuates biases. In the case of age and gender, the researchers found that ChatGPT assumed women were younger and less experienced and rated resumes from older men the best—a concerning fact as more and more companies use AI in hiring, from screening resumes to conducting and recording interviews.
“Computer-driven decisions have this veneer of objectivity,” said Hilke Schellmann, a professor at NYU and author of the book The Algorithm who was not involved in the study. “The problem lies in that we as humans often think the results of models seem objective, thematically correct, but in reality, if you have biased data going in, you will in all likelihood replicate the bias. And we see this again and again.”
The large AI models “require consuming all of the internet’s data,” Guilbeault told me. “When you start dealing with data at that scale of human culture, it’s inevitable that it’s going to just be fraught with biases and stereotypes and mythologies and illusions, and so it’s really problematic.”
And the stereotypes can become more deeply rooted as AI develops. “The model learns from the previous model, and if the bias is baked in, we have a lot of evidence that it may amplify the bias,” Schellmann said. There aren’t guardrails built into the models, and there’s very little oversight, she added.
The researchers studied age and gender because those were two categories they could confidently measure across the internet—but they are far from the only ways online data and AI models are biased. Other research has found that AI image generators often produce racist and sexist stereotypes, which more people will likely encounter as AI images and text pervade search results and the wider online landscape.
“People are increasingly relying on the internet and these algorithms to learn about their social world, to filter information, to give them images and videos and content that then they use to inform their views of who people are and how the world works,” Guilbeault said. “These popular algorithms used by millions and millions of people every day are entrenching these biases.”
This post has been syndicated from Mother Jones, where it was published under this address.