Images on the internet are even more sexist than texts

The photos that appear on a Google Images search for professions like cardiovascular surgeon, mathematician, developer, soccer player, curator and investment models are of men. But if you instead type in model, hairdresser, nurse, interior designer or art teacher, you’ll get back pictures of women. A study published on Wednesday by the journal Nature analyzed 349,500 photos, 100 for each of the 3,495 most common social categories, observed a consistent gender bias on Google and other platforms like Wikipedia and IMDB. Until now, the most relevant studies analyzing these online biases had focused mainly on text.

Researchers at the University of California Berkeley believe that the finding is important, given the unstoppable rise of visual culture. The amount of time we spend reading is shrinking and the amount of entertainment we consume on “hyper-visual platforms based predominantly on the exchange of images” is growing, the study states. The bias in online images is greater than that seen in surveys and in actual U.S. census data. Although, according to Douglas Guilbeault, a professor at UC Berkeley and one of the study’s co-authors, “one subtlety is that social groups and generations may differ in the type of visual content they produce and consume, and an important area for future research would be to explore how this affects their experience.”

Researchers have also found that bias in online images is more common than in text, and that its effects are more psychologically potent. In the paper, they compared images with text results from Google News. Images show a greater bias than text, and their impact is more lasting: people who saw an image with gender bias retained its influence for longer than those who read a sexist text. “What strikes me most about the work is that a simple shift from textual to visual information can have so many implications for the spread of gender bias, threatening decades of progress,” says Guilbeault.

It is well worth mentioning that the language used for the study is English, a tongue in which the word for most jobs does not vary by gender, unlike in languages such as Spanish, in which the majority of nouns are gendered. The photos, therefore, provide more information than their text description might, like when a male soccer player or female hairdresser appears. But researchers have found effects that go beyond the words themselves: “Several psychological investigations suggest that images may be a particularly potent medium for conveying gender bias. Research on the image superiority effect shows that images are often more indelible and emotionally more impactful than text,” the article states.

How does such bias come to be?

Photos in internet searches reproduce millions of choices by users, who create and upload these images to their accounts and pages: “Gender bias appears to be partially driven both by the content that internet users choose to display on their blogs, and by audiences’ preferences about which news to consume or which images to purchase,” states the study.

But how have the mechanisms of the internet themselves played a role in fostering this online image bias? “It may be that biases are intensified as a result of the network itself,” says Professor Bas Hofstra of the Netherlands’s Radboud University, who analyzed the study prior to its publication. “Perhaps user populations differ, perhaps men are the ones who have used or use the internet the most and, as such, it becomes more gendered,” he says.

The article also speculates that prejudices within the hiring practices of media companies, whose photos are overrepresented on the internet, could be another cause. “The human preference for familiar, prototypical representations of social categories is likely to play a role in perpetuating these biases,” state the study.

One consequence of this dynamic is to limit women’s access to vocations in which they have not yet gained a foothold due to preexisting social dynamics. But there could be more to it, according to Hofstra: “The way people talk about certain social categories, for example: men are talked about more in relation to sports because you see more male athletes online, which could diminish women’s aspirations to enter the sport.”

An obvious fear when it comes to the future that only receives brief mention in the article is that the images on the current internet are one of the primary building blocks of generative artificial intelligence (AI). “Images created with AI can make the internet a more sexist place if they are based on previously generated online images,” says Hofstra. “Our work suggests that gender bias in AI could be due in part to the fact that it is trained with public images from platforms like Google and Wikipedia, which are filled with gender bias,” states the study.

It is difficult to think of concrete solutions for this problem. Google could be giving us biased results because searches and the information that already exists on the internet are biased. “The answers will have to come from technology, academia and civil society,” says Guilbeault. “Our investigation’s goal is to start a critical conversation about the implications of the shift to visual images when it comes to the propagation of gender bias. The methods and tools that we have created are an important step towards the transparency of information regarding the changing panorama of online prejudice, which is a necessary step towards finding solutions.”

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition