Bankers are male, nurses are female? Such stereotypical gender roles can be reinforced and reinforced through online content. A study shows that images on the Internet use clichés much more than texts – at least in the English-speaking area. Especially given the fact that more and more people prefer to consume visual content online, this distortion is worrying from the researchers’ point of view. An experiment shows that gender bias was more pronounced several days after an online image search than after a text-based search.
The Internet serves as the most important source of information and entertainment for an ever-increasing number of people. The focus is increasingly shifting from texts to images. “Every year, people spend less time reading and more time looking at pictures,” writes a team led by Douglas Guilbeault from the University of California at Berkeley. “Every day, millions of people download images from platforms like Google and Wikipedia, and millions more interact through social media like Instagram and TikTok, which primarily consist of sharing visual content.”
Gender stereotypes in images and texts
But what impact does this shift towards more images have on our perception of gender roles? To answer this question, Guilbeault and his team examined the gender associations in texts and images related to nearly 3,500 social categories, including job titles such as doctor, interior designer and model, as well as social roles such as neighbor or colleague. The researchers used more than a million images from Google, Wikipedia and the Internet Movie Database (IMDb) as well as billions of words from these platforms as a database.
Since the respective textual terms in English are gender-neutral, the team evaluated the context of the text to determine whether a specific gender was meant. To determine the perceived gender of the people in the images, they had more than 6,000 people recruited online do the matching. “For each individual category, we then calculated the percentage of women and men represented in the images and texts,” explains the team.
Images are more distorted than text
The result: “We found that gender bias is consistently more pronounced in images than in texts, for both female and male categories,” report the researchers. For example, a search for “detective” returned almost exclusively images of male detectives. In texts, however, the gender ratio was almost balanced. A search for “model,” on the other hand, led predominantly to images of female models, while the texts were also female-dominated, but also to a greater extent about male models.
Guilbeault and his team also found that women were overall underrepresented in search results, and here again more in images than in texts. The Google image search showed male overrepresentation for 62 percent of the categories, while the texts from Google News only showed male overrepresentation for 56 percent of the categories.
The researchers compared this result with US census data to see how well women and men are actually represented in the respective professional groups and what the distribution is in public perception. This showed that while texts tend to represent greater gender equality than is realistic given the census data, the underrepresentation of women in the images is more pronounced than in reality.
Influence on our prejudices
But do such distortions actually affect our perception of gender roles? The researchers tested this in an experiment. They asked 450 people to either search for specific jobs on Google News or Google Images or – as a control group – to search for random objects. Immediately after the experiment and again three days later, they asked the test subjects which gender they associated certain professional groups with. In addition, they tested the strength of these associations using a so-called implicit association test.
“Our experiment shows that Googling images rather than textual descriptions of jobs increases participants’ gender bias,” the researchers report. Even three days after the experiment, people who googled images related to the jobs had stronger gender biases than those who googled texts related to the jobs or objects.
Relevant for AI-generated images
“These results suggest that online images and thus the online world are not only strongly gendered, but that this gendered bias could also influence other gender-specific distortions in everyday life,” write Bas Hofstra and Anne Maaike Mulders from Radboud University in Nijmegen in an accompanying commentary on the study, which was also published in the journal Nature.
Especially in view of the fact that artificial intelligence is also creating more and more images, it is important to understand the mechanisms that lead to the distortion and to take countermeasures. “If the AI-generated images are based on online images that are already gendered, the images on the internet could otherwise become increasingly gender-biased,” say Hofstra and Mulders.
Source: Douglas Guilbeault (University of California, Berkeley) et al., Nature, doi: 10.1038/s41586-024-07068-x