Racial bias in AI-generated images

Abstract

Abstract AI-generated images consistently favor White people compared to people of color. This paper examined the image-to-image generation accuracy (i.e., the original race and gender of a person’s image were replicated in the new AI-generated image) of a Chinese AI-powered image generator. We examined the image-to-image generation models transforming the racial and gender categories of the original photos of White, Black and East Asian people ( N = 1260) in three different racial contexts: a single person, two people of the same race, and two people of different races. The findings indicated that White people were more accurately depicted in AI-generated images than people of color in all three racial contexts. Black people, particularly females, were depicted with the lowest AI-generated racial accuracy in the image of a single person, but with higher accuracy in the image of two people of different races. The pattern of Asian people, particularly males, was the inverse: the app had higher AI-generated racial accuracy for Asians in the single-person image but lower accuracy for Asians in the two-people-of-different-races image. In all cases of incorrect racial generating, the AI-powered image generator depicted most people of color as White. This study provides us with insight into racial and gender bias in image generation and the potential representational harms of an AI-powered beauty app developed in China. More broadly, these technological biases reflect a form of postcolonial globalization that impacts image-processing systems in non-White settings, including social values of white supremacy and norms of white beauty.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 28 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.