Data Diversity and Algorithmic Bias: Do Racially-Imbalanced Training Data Produce Human-Like Bias in Face Classification?

Can a lack of racial diversity in training data produce human-like forms of AI face classification bias? We examined the causal effect of racially-imbalanced training data on face classification model disparities and whether these disparities resemble three well-documented human biases—the cross-race effect, hypodescent, and the gendered-race effect. We further examined whether balanced training data mitigate these disparities. In three experiments, we trained deep Vision Transformer models for face classification on 16,000 face images, experimentally manipulating racial composition while holding other dataset features and training procedures constant. Study 1 showed that models trained on low-diversity (majority-White) datasets classified White faces more accurately than Black, East Asian, and Indian faces, resembling the human cross-race effect. In Study 2, using parametrically-morphed Black-White biracial face images to estimate race-classification thresholds, lower-diversity models required more Black than White content to reach this threshold, yet represented biracial faces with more Black content. Lastly, Study 3 revealed a gendered-race effect: regardless of training set imbalance, models classified Black men more accurately than Black women, but East Asian women more accurately than East Asian men. The use of racially-balanced training data nearly eliminated disparities in monoracial classification, reduced bias in biracial classification, but it did not alter the gendered-race effect. These findings show that training data composition causally shapes algorithmic performance in ways that both mimic and diverge from human biases. We discuss implications for AI-based discrimination and both the promise and limitations of data balancing as a strategy to mitigate algorithmic bias.