Happy Young Women, Grumpy Old Men? Emotion-Driven Demographic Biases in Synthetic Face Generation

This study audits demographic biases in synthetic faces generated by text-to-image (T2I) models, analyzing outputs from eight models (four Western and four Chinese) across 56,000 generated images under various emotional prompts. The research reveals a significant overrepresentation of young, White-coded faces, particularly under negatively valenced emotional conditions, which skew outputs towards male, middle-aged individuals and reduce perceived attractiveness for certain demographic combinations, such as young Black females. These findings highlight the necessity for comprehensive, intersectional audits of T2I models that consider emotional context to address pervasive biases prior to deployment.