Generative bias: widespread, unexpected, and uninterpretable biases in generative models and their implications

Abstract

Generative models, with their ability to create new data, have significantly advanced machine learning. However, these models can also produce biased outputs, threatening representational fairness. Current literature mainly addresses biases in synthetic data that amplify social biases, such as racial and gender stereotypes, alongside underrepresentation and misrepresentation. In this paper, we argue for the existence of previously unexplored biases that generative models can produce on a massive scale. Using Generative Adversarial Networks as a case study, we show that these unexpected and hard to interpret biases can systematically marginalize and denigrate minoritized communities. Importantly, these biases are not merely a reflection of human biases in the training data but a consequence of machine learning processes that diverge from the human perspective. We contend that current debiasing strategies are inadequate, as they fail to fully grasp the epistemic challenges and underestimate the moral risks of generative models. We conclude by suggesting a novel direction to ameliorate their harm.