The authors argue that generative AI systems not only produce biased outputs but also encode the dominant culture's way of knowing as the default infrastructure of knowledge. This 'epistemic subordination' occurs during training, where a single probabilistic model is created based on the languages, assumptions, and cultural frameworks of the majority. Minority epistemologies are absorbed into this model, but structurally subordinated in the output. The authors claim that existing laws regulating anti-discrimination, cultural and linguistic rights, and democratic viewpoint pluralism fail to address this issue because they focus on downstream decisions rather than the training process itself. --- Why it matters: This matters to researchers in AI because it highlights a fundamental flaw in current generative models: their inability to represent diverse perspectives. Understanding and addressing epistemic subordination is crucial for creating more inclusive and equitable AI systems. Source: https://arxiv.org/abs/2608.18758