National elections through the lens of Google: Large-scale audit of gender bias in image search results during the 2023 parliamentary elections in Switzerland

As voters routinely rely on search engines like Google to learn about political candidates, the output of information searches–and in particular, visual cues therein–constitute central information sources for voters’ opinion formation process. While a growing body of evidence has documented gender and partisan biases in visual political communication (e.g. Van der Pas & Aaldering, 2020; Rohrbach et al., 2023) together with gender bias in search engine performance (e.g. Otterbacher et al., 2017; Pradel, 2021; Ulloa et al., 2022), there is currently little empirical evidence of the role of AI-driven search engines in real-world election campaigns. Similarly limited is research on the impact of search engines for elections in the Swiss context despite its distinctiveness due to Switzerland being a direct democracy. To address these two gaps, we conduct a large-scale algorithmic audit of Google image searches to investigate how 5,909 candidates running for the Swiss parliamentary elections in 2023 are visually represented. We focus on Google as the latter constitutes a quasi monopolist on the Swiss search market; it focuses on image search due to it being particularly relevant for visual aspects of gender bias and susceptible to malperformance/manipulation. By conducting the algorithm audit of image search results for all candidates participating in the elections, we aim to investigate how their representation by search engine algorithms varies depending on the gender, the party with which the candidate is affiliated, and the type of region which the candidate aims to represent (i.e. urban vs rural cantons). Specifically, we deploy 120 virtual agents to scrape search engine output in two waves of data collection (4 and 1 week before the election) and use computer vision techniques (based on Amazon Rekognition) to annotate over 3.5 million collected candidate images with information regarding emotional displays, objects present on the images, and face-body ratios (i.e., face-ism). Preliminary results show a dominance of positive emotions, with smiling and happy candidates. Crucially, gender (and party) differences emerge that reinforce gender stereotypes but clash with leadership expectations: Women are shown as more positive and communal, while men are more likely to show negative emotions, such as anger. The findings present large-scale and real-world evidence of subtle biases that reinforce women’s double bind in politics. References: Pradel, F. (2021). Biased representation of politicians in Google and Wikipedia search? The joint effect of party identity, gender identity and elections. Political Communication, 38(4), 447-478. Otterbacher, J., Bates, J., & Clough, P. (2017, May). Competent men and warm women: Gender stereotypes and backlash in image search results. In Proceedi Rohrbach, T., Aaldering, L., & Van der Pas, D. J. (2023). Gender differences and similarities in news media effects on political candidate evaluations: a meta-analysis. Journal of Communication, 73(2), 101-112. Van der Pas, D. J., & Aaldering, L. (2020). Gender differences in political media coverage: A meta-analysis. Journal of Communication, 70(1), 114-143. Ulloa, R., Richter, A. C., Makhortykh, M., Urman, A., & Kacperski, C. S. (2022). Representativeness and face-ism: Gender bias in image search. New Media & Society (Online First).