Online images amplify gender bias

The world is increasingly visual. People spend more time viewing images than reading text, a trend accelerated by the proliferation of images online through platforms like Google, Wikipedia, Instagram, and TikTok. News agencies and advertisers also capitalize on the power of images to capture attention, leveraging their memorable and implicitly processed nature. This study explores the consequences of this visual shift, focusing on its impact on the spread of gender bias. Frederick Douglass, in his 1861 lecture, warned of the potential for images to reinforce social biases, a concern only amplified by the internet's ability to disseminate visual content at scale. While much research has focused on gender bias in text, this study uniquely compares the prevalence and psychological impact of this bias across images and text, anticipating that online images will amplify gender bias both statistically and psychologically.

Existing quantitative research on online gender bias predominantly focuses on textual data. A few recent studies have explored gender bias in Google Images but with limited scope, lacking comprehensive comparisons with text and a thorough investigation of the psychological effects. The 'picture superiority effect' suggests images are more memorable and emotionally evocative than text, possibly implicitly influencing text comprehension. Images also directly convey demographic cues, making gender more salient than in text where gender can be easily minimized or omitted. This difference indicates images may be a potent medium for transmitting gender bias.

The study employed both computational and experimental methods. For the computational analysis, the researchers examined gender associations of 3,495 social categories (occupations and social roles) in over one million images from Google, Wikipedia, and IMDb, along with billions of words from those platforms. Gender bias in images was measured by automatically retrieving the top 100 images for each category from Google Images and having a team of human coders classify the gender of faces. Gender bias in text was measured using word embedding models that capture the extent to which each category co-occurs with words associated with men or women. The study compared image-based and text-based gender bias measures, also considering public opinion and US census data on occupations. The experimental component involved a nationally representative sample of US participants (n=450) using Google to search for either textual descriptions (via Google News) or images of occupations. Participants then rated the gender they most associated with each occupation. An implicit association test (IAT) was also administered to measure implicit biases. Statistical analyses compared gender associations across images, text, public opinion, and census data, as well as the effect of image versus text searches on participants' explicit and implicit biases.

The study revealed that gender bias is consistently more prevalent in images than in text, for both female- and male-typed categories. The underrepresentation of women online is significantly worse in images than text, public opinion, and US census data. The observational analyses demonstrated a strong correlation between gender associations in images and text (r=0.5), but image-based associations were consistently more extreme. Specifically, the average gender bias was more than four times stronger in images than text (μ=0.14 vs. μ=0.03). The experiment showed that googling for images of occupations, rather than textual descriptions, significantly amplified gender bias in participants' explicit beliefs. Participants in the image condition reported significantly stronger gender associations than those in the text condition (mean difference=0.06, p<0.001). A correlation (r=0.79) was observed between the gender of uploaded descriptions and participants' explicit gender ratings. Even when controlling for the prevalence of gender bias, images were found to prime gender bias more strongly than text. Finally, suggestive results indicate that images also amplified participants' implicit gender bias, although this effect was less robust than the effect on explicit bias.

The findings strongly support the hypothesis that online images amplify gender bias. The greater prevalence of bias in images compared to text, along with the experimental demonstration of image-based priming, suggests images are a more potent medium for transmitting and reinforcing gender stereotypes. This has significant societal implications, particularly given the increasing dominance of image-based social media and the integration of images into search engine functionality. The use of multiple data sources (Google, Wikipedia, IMDb) and multiple methodologies (observational, experimental) strengthens the study's conclusions. The study's results resonate with ongoing concerns regarding algorithmic bias and the need for a multimodal approach to computational social science.

This study demonstrates that the increasing reliance on visual communication online comes at a cost, amplifying gender bias both statistically and psychologically. Addressing this issue requires a multi-pronged approach focusing on content creation, algorithmic design, and media literacy. Future research could investigate the social and algorithmic factors contributing to bias in online images, extend the multimodal framework to other communication modalities (audio, video), and compare human-generated content with AI-generated content, which also demonstrate biases. Developing a multimodal approach to computational social science is crucial for building a more fair and inclusive internet.

While the study employed rigorous methods, several limitations exist. The IAT results, while suggestive, were less robust than the findings regarding explicit bias. The observational analyses focused primarily on Google Images, although results were replicated using Wikipedia and IMDb data. The study's focus on gender bias leaves open the question of whether similar effects exist for other demographic dimensions. The study's experiment focused on STEM and liberal arts occupations, potentially limiting the generalizability of findings to other occupational categories.