Beyond Categories of Caste: Examining Caste Bias and Morality in Text-to-Image AI Models

Researchers examined how Text-to-Image AI models perpetuate caste biases in South Asian contexts, shifting analysis from treating caste as a static identity category to understanding it as a relational system. Using algorithmic audits and critical discourse analysis, they propose an anti-caste framework to address fairness issues in generative AI systems beyond simple upper/lower-caste binaries.

This research addresses a critical blind spot in AI fairness discourse: the treatment of caste discrimination as a categorical problem rather than a relational one. While previous studies documented caste stereotypes in generative AI outputs, this work reframes the investigation to examine how T2I models perpetuate hierarchical social structures inherent to caste systems. By combining algorithmic audits with critical discourse analysis, researchers demonstrate that bias operates through complex relational dynamics that simple demographic categorization cannot capture.

The significance lies in how South Asian social hierarchies have been largely invisible in Western-dominated AI ethics discussions. Most fairness research focuses on race, gender, or ethnicity within specific geographic contexts, leaving caste—a system affecting billions in South Asia—underexamined in generative AI systems. This gap allows harmful stereotypes to propagate unchecked through increasingly mainstream tools like DALL-E, Midjourney, and Stable Diffusion.

The anti-caste approach proposed here has implications for broader AI governance. Rather than treating protected categories as checkboxes for compliance, this framework suggests developers must understand the structural, historical, and relational dimensions of discrimination. For AI companies operating globally, this research underscores that fairness requires culturally-specific, context-aware analysis rather than one-size-fits-all solutions.

Moving forward, the AI industry should expect increased scrutiny on how generative models handle non-Western social hierarchies and identity systems. This work sets a precedent for South Asian researchers and marginalized communities to demand representation in AI ethics discussions, potentially reshaping how fairness audits are conducted and what demographics are included in bias testing frameworks.

• →Text-to-Image models perpetuate caste biases through relational hierarchies rather than simple categorical stereotyping.
• →Existing AI fairness research overlooks caste discrimination despite its impact on billions in South Asia.
• →An anti-caste framework is needed to address structural discrimination in generative AI systems.
• →Western-dominated AI ethics discourse lacks culturally-specific analysis of non-Western social hierarchies.
• →Global AI companies must implement context-aware fairness audits beyond demographic categories.