No Bias, No GAN: Generative Models Produce Widespread, Unexpected, and Uninterpretable Biases

Generative models have revolutionized machine learning with their ability to synthesize new data. However, they can generate biased images or languages that threaten representational fairness. Current literature has primarily focused on the nature of and solutions to biases that amplify existing societal biases, such as stereotypes and underrepresentation associated with race and gender. In this paper, we argue for the existence of previously unexplored types of biases that Generative Adversarial Networks can produce on a massive scale. These biases can be unexpected and uninterpretable, but systematically marginalize and denigrate minoritized communities. Importantly, they are not merely a reflection of human biases in the training data but a consequence of machine learning processes that diverge from human perspective. We contend that the current debiasing strategies are inadequate, as they fail to fully grasp the epistemic challenge and underestimate the moral risks of generative models. We conclude by suggesting a novel direction to ameliorate their harm.