Examining Cultural Representation in Generative Image AI Models: A case study of Stable Diffusion

This study investigates AI fairness in cultural representation, focusing specifically on generative Text-to-Image AI Models, an emerging research area. Through a case study on Stable Diffusion, the research examines cultural representation by analyzing a diverse group of 7 cultures. Employing a mixed-methods approach involving quantitative image quality assessment, computer vision detection, surveys, and socio-semiotic analysis, the study uncovers both strengths and limitations in cultural representation. The findings highlight instances of inaccuracies and biases across all cultures, revealing the presence of more negative biases in images representing Oriental cultures in the global South. Additionally, the study identifies gender disparities and cultural biases within and across cultures, shedding light on underlying gender inequalities. By contextualizing these biases within broader societal frameworks, incorporating concepts such as Orientalism, colonialism, and intersectionality, the study explores the causes and far-reaching implications of these biases, particularly in perpetuating stereotypes and inequalities against marginalized cultures. The research also proposes potential mitigation strategies and underscores the need for improvements in AI-generated image assessment methods. This research contributes to the field of AI fairness by uncovering multifaceted cultural biases within generative AI image models and enhancing the shared understanding of the complex socio-techno dynamics of cultural representations in AI. The findings underscore the critical importance of examining cultural biases and striving for accurate and fair cultural representation in AI technologies.