Online images amplify gender bias

The world is shifting from a text-centric to an image-centric information landscape. People are spending less time reading and more time viewing images, which are proliferating online at an unprecedented rate. This shift is driven by the popularity of image-based social media platforms (Instagram, TikTok, Snapchat) and the increasing use of images by news agencies and digital advertisers to capture attention. The ease and low cost of image circulation online, facilitated by the internet, potentially magnifies the impact of images in reinforcing social biases, a concern raised even in the era of early photography by Frederick Douglass. This study directly addresses this concern, focusing on the impact of online images on the large-scale spread of gender bias. While existing research primarily focuses on textual gender bias, this study uniquely compares the prevalence and psychological impact of gender bias across both images and text, considering the 'picture superiority effect' – the tendency for images to be more memorable and emotionally evocative than text – and the increased salience of demographic information in images compared to text. The authors hypothesize that online images amplify gender bias, both statistically and psychologically.

Most quantitative research on online gender bias focuses on textual data. A few recent studies examine gender bias in small samples of Google Images, but these lack a comparison of bias prevalence and psychological impact across image and text data. This study builds upon prior work highlighting the 'picture superiority effect' demonstrating images' superior memorability and emotional impact compared to text. The authors also consider research showing images' implicit influence on text comprehension and the increased salience of demographic information presented visually. This existing literature supports the study's hypothesis that online images amplify gender bias more effectively than text.

The study employed both computational and experimental methods to compare gender bias across massive online corpora of images and texts. The main analysis compared data from Google, Wikipedia, and IMDb. **Computational Analysis:** * **Image Data:** The top 100 Google Image results for 3,495 social categories from WordNet were analyzed. This involved over 349,500 images, significantly larger than previous studies. Data collection used multiple Google accounts and servers to mitigate algorithmic bias. The gender of faces in images was determined by a team of 6,392 human coders on Amazon Mechanical Turk (MTurk). Intercoder reliability was assessed (Gwet’s AC = 0.48). Results were robust to various controls (coder demographics, image number, search location, etc.). A replication using a celebrity image dataset from IMDb and Wikipedia (72,214 celebrities, 511,946 images) corroborated the findings. * **Text Data:** Gender bias in text was measured using word embedding models (Word2Vec, trained on the 2013 Google News corpus), which capture the co-occurrence of words. The method positioned each category along a -1 (female) to 1 (male) axis. Robustness checks involved using multiple word embedding models (GloVe, BERT, FastText, ConceptNet, GPT-3) and different corpora. * **Bias Measurement:** Gender bias was quantified across three dimensions: (1) gender association with social categories, (2) representation of women compared to men, and (3) comparison with public opinion and US census data. **Experimental Analysis:** * **Participant Pool:** A nationally representative sample of 450 US participants from Prolific were recruited. * **Experimental Design:** Participants were randomized into three conditions: (1) Text (using Google News for textual descriptions), (2) Image (using Google Images for images), and (3) Control (using Google for unrelated categories). Participants searched for and uploaded descriptions of 22 randomly selected occupations from a set of 54, then rated the gender association. An Implicit Association Test (IAT) measuring implicit bias towards associating women with liberal arts and men with science was administered immediately after and 3 days later. * **Data Analysis:** The gender balance of uploaded descriptions (coded by annotators) and participants' explicit gender ratings were compared across conditions. The study also analyzed the correlation between uploaded descriptions and explicit ratings, as well as the IAT results. Robustness checks included accounting for the number of online sources, time spent on the task, and participant gender.

The study's key findings consistently demonstrate that gender bias is amplified in online images compared to text: 1. **Stronger Gender Associations in Images:** Gender associations for social categories were significantly more extreme (both more female and more male) in Google Images compared to Google News, holding even when controlling for various factors (linguistic features, search frequency, etc.). 2. **Greater Underrepresentation of Women in Images:** Women were significantly underrepresented in images compared to text, public opinion, and US census data on occupations. This underrepresentation was markedly more pronounced in images than in text. 3. **Experimental Amplification of Bias:** The experiment revealed that exposure to images of occupations significantly increased participants' reported gender associations compared to exposure to textual descriptions. Participants in the Image condition showed significantly stronger explicit gender bias than those in the Text and Control conditions. The correlation between the gender associations in uploaded descriptions and participants' explicit ratings was strong (r = 0.79, P < 0.0001). Even when controlling for bias prevalence, images primed stronger gender bias in self-reported beliefs than text. 4. **Amplification of Implicit Bias:** While suggestive, the results indicate that the Image condition may have amplified participants' implicit gender bias, as measured by the IAT. Participants in the Image condition exhibited stronger implicit bias compared to the Control condition, with a potentially enduring effect 3 days after the experiment. A correlation existed between the strength of explicit and implicit bias across conditions.

The findings directly address the research question of whether online images amplify gender bias, providing strong evidence in the affirmative. The consistent observation across computational and experimental analyses shows that gender bias is not only more prevalent but also more potent psychologically in images compared to text. This has significant implications, considering the increasing dominance of visual content online. The experimental results highlight the causal link between exposure to gendered images and the reinforcement of gender stereotypes in beliefs. The amplification of both explicit and implicit bias underscores the pervasive and insidious nature of this problem. The significant difference in bias between image and text conditions highlights the unique influence of visual media. These findings contribute significantly to the understanding of how online platforms perpetuate and amplify gender inequalities, going beyond previous studies focused solely on text. The study’s rigorous methodology and large-scale dataset strengthens the conclusions, making the findings highly relevant to the fields of social psychology, media studies, and computational social science.

This research demonstrates that the prevalent use of images online significantly exacerbates gender bias, both statistically and psychologically. The increased salience of gender in images and their potent psychological impact contribute to this effect. The increasing reliance on visual content in online platforms warrants immediate attention to mitigate the amplification of gender bias. Future research should investigate the social and algorithmic processes that contribute to bias in online images, extending the analysis to other demographic dimensions and content modalities (audio, video). Developing a multimodal approach to computational social science is crucial for addressing these challenges and creating a more fair and inclusive digital environment.

While the study employed a large dataset and a rigorous methodology, some limitations exist. The IAT results, while suggestive, should be interpreted cautiously given ongoing debates about its reliability. The study focused primarily on gender bias; future research could explore other biases (racial, ethnic, etc.). The reliance on Google Image search may reflect biases inherent in the algorithm or its indexing of the web. The study didn't directly address the causal mechanisms underlying the amplification of bias; future research should explore this in more detail.