Comparing Occupational Gender Bias in AI-Generated Anime-style and Realistic Illustrations

Ayano Ito, Shoma Mizuno, Keiichi Namikoshi, Yuko Sakurai, Satoshi Oyama

Track: Short paper

Keywords: AI-generated anime-style illustrations, AI-generated realistic illustrations, Gender bias

Abstract: The rapid advancement of generative AI has enabled the creation of both realistic and anime-style illustrations across diverse domains. However, such systems risk reproducing and amplifying cultural and social stereotypes embedded in their training data. This study presents a comparative analysis of occupational gender bias in AI-generated anime-style and realistic illustrations. Using Stable Diffusion 3.5, Illustrious, ChatGPT-5, and Gemini 2.5 Pro, we generated images for ten occupations, with each prompt depicting ten individuals. Our analysis examined gender distributions and expressive features such as crossed arms. Results show that anime-style illustrations amplify gender imbalances and cultural stereotypes more distinctly than realistic images. For instance, in the case of nurses, female representations are particularly emphasized. Across occupations, anime-style generation is also strongly shaped by conventions characteristic of Japanese anime. These findings underscore the critical role of stylistic domains in the reproduction of cultural conventions in generative AI and highlight the importance of evaluating bias across different visual styles.

Submission Number: 19