Users prefer generative AI to reflect real over idealized gender proportions

Generative artificial intelligence (GAI) systems must make decisions about how to represent gender in images of social categories. Different standards exist for evaluating whether such representations are appropriate: systems can reflect current statistical distributions (proportional equality), depict equal proportions (uniform equality), or actively promote diversity (substantive equality). While different authors apply different standards, what constitutes an appropriate representation from users' perspective is rarely investigated systematically. Understanding user expectations is crucial. Our study therefore investigates users' evaluations of gender representation in various social categories. In an experiment with 719 US participants, we find that uniform equality and substantive equality play little role in user evaluations. Users prefer a representation that reflects their perception of current reality. Our results show that users prefer statistical alignment.