Skewed toward uniformity? Investigating algorithmic stereotyping and amplification in AI-generated images

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

As text-to-image generative artificial intelligence (AI) increasingly shapes visual communication, concerns are growing about its role in reproducing social stereotypes. While recent algorithm auditing studies have advanced our understanding of stereotypical AI-generated imagery, they have largely focused on a narrow set of stereotypes – primarily gender and race – overlooking more nuanced forms. Addressing this gap, our study adopts a case-sensitive approach that conceptualizes stereotypes as multi-layered phenomena composed of multiple, context-specific indicators. Applying the Draw-a-Scientist Test framework, we analyzed stereotypical representations of scientists in images generated by DALL·E 3 and examined potential amplification by benchmarking against Google Images. Specifically, we compared 16 stereotype indicators across six stereotype subjects. Using Latent Class Analysis, we identified six classes reflecting distinct patterns of stereotypical depiction. Our findings show that DALL·E not only produces highly stereotypical portrayals of scientists but also amplifies these patterns compared to Google Images. Both image sources lack adequate visual references for disciplines beyond the traditional laboratory context – a pattern we interpret as epistemic narrowing, whereby diverse forms of knowledge production are reduced to a uniform, laboratory-based representation of science. We conclude by discussing potential drivers of epistemic narrowing, the broader implications of algorithmic visibility, and mitigation strategies.