Generative AI So White: Racial Biases in AI Imagery Across the United States and China

Abstract

This study investigates racial biases in AI-generated occupational images across models developed in the United States and China. Situated at the intersection of human–AI communication and postcolonial theory, we conceptualize generative AI as an active participant in science communication that shapes visual knowledge and racial representation within a global postcolonial order. Constructing a dataset of 9,600 images generated by four models (GPT-4o, Llama 3, Wanx2.0, and Wenxin 3.5), we examine three levels of racial biases—representational bias, positional bias, and racialized meaning bias—using a mixed-methods approach. Findings show that White individuals are overrepresented, granted spatial dominance, and encoded through aesthetic and symbolic conventions that racialize non-White bodies. We reveal how generative AI reproduces global racial hierarchies under algorithmic neutrality, advancing a cross-national auditing framework and contributing to decolonial science communication by foregrounding a human-centered AI perspective.

Ähnliche Arbeiten