Online Gender Stereotypes are Stronger in Images than Text

Join online: https://berkeley.zoom.us/j/94627196292

Advancements in natural language processing have spurred the proliferation of studies examining gender stereotypes in online texts, including news and social media. Yet, while these studies suggest a reduction of gender bias in recent years, research indicates that progress toward gender equality has slowed or stalled in vital areas of social life, from hiring practices to household management. Textual measures of online stereotypes are at risk of underestimating the gender gap, which may be more salient in online images that visualize the demographics of people. In this talk, I show that online gender stereotypes are more prevalent in images than texts using a novel dataset comprising over one million images from Google, Wikipedia, and IMDb, mapped to over 3,400 distinct social categories, including occupations (e.g., “doctor”) as well as generic social roles (e.g., “friend”) and lifestyles (e.g., “vegan”); stereotypes in these images are then compared to stereotypes measured by word embedding models trained on billions of words from the same online platforms. To characterize the empirical consequences of these findings, I use an online experiment to show that googling for visual rather than textual descriptions of occupations amplifies people’s implicit bias toward associating men with science and women with liberal arts, a stereotype linked to pervasive inequalities in academia and industry. I conclude by showing how text and images can differ in the kinds of stereotypes they encode; for example, I show that gendered ageism, whereby women are pressured to appear younger than men, is particularly pervasive in online images. Implications for algorithmic bias are discussed.