Title: Transcending Caste Categories: An Analysis of Caste Bias and Morality in Text-to-Image AI Models

Abstract:

Text-to-Image (T2I) models have demonstrated significant potential across numerous fields. However, these systems are simultaneously exacerbating harmful societal prejudices in the content they generate. Recent studies within the South Asian context have highlighted how Generative AI (GenAI) systems reinforce caste-based stereotypes and biases. Although this existing research provides crucial insights into the obscured narratives of caste discrimination facilitated by GenAI, it predominantly conceptualizes caste as a static identity category.

In contrast, this study adopts a different ontological perspective, prioritizing the relational dimensions of caste. This shift allows for a more sophisticated comprehension of the mechanisms through which T2I models enact and perpetuate caste discrimination. By integrating algorithmic auditing with critical discourse analysis, we employ a theoretical framework that challenges Brahminical Normativity. This approach reveals how caste biases are sustained in ways that extend beyond simplistic dichotomies of upper versus lower-caste classifications.

Our research offers two primary contributions. First, we move beyond the conventional categorical definition of caste. Second, we advocate for an anti-caste methodology to address concerns regarding fairness and bias within AI systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC