The standard audit for bias in a text-to-image model runs on a simple design: choose a demographic label, generate a batch of images against it, and compare the results to a neutral baseline. Applied to caste, this means prompting a model for images of Dalits, the communities once branded “untouchable,” roughly 200 million people who sit at the base of South Asia’s caste hierarchy, and checking whether the outputs skew toward poverty, manual labor, or degraded settings more often than chance would predict. The method is categorical: it treats each caste label as a bucket to be checked for stereotype, one bucket at a time.
A paper presented at the ACM Conference on Fairness, Accountability, and Transparency (FAccT’26) in Montreal, held June 25 to 28, 2026, argues that this categorical method, whatever it catches, misses the structure that makes caste bias caste bias. Caste is not a set of freestanding labels; it is a ranked relationship between groups, historically enforced through endogamy and purity-pollution hierarchies that only mean anything in comparison to what sits above and below. A model can generate individually unstereotyped images of a Dalit person and a Brahmin person and still encode caste bias in how it places the two in relation to each other: who serves whom, who is depicted with authority, whose spaces the other appears in. The paper, by Divyanshu Kumar Singh, Dipto Das, Deepika Rama Subramanian, Koustuv Saha, Stephen Voida, and Bryan Semaan, names the thing this relational structure enforces directly: Brahmanical Normativity, an unmarked moral order that assigns dignity and utility to bodies (who a model renders as worthy of comfort or authority), governs social relations (who is permitted to interact with whom, and how), and shapes material-spatial relations (whose mobility and environment the model treats as constrained by default). The authors put the asymmetry bluntly:
“Why does the model enforce dignity for lower-caste, yet upper-caste have default right to dignity?”
Discrimination in these systems, on their account, is relational, not categorical, and audit method has to change to catch it.
That reframing has a direct consequence for how caste bias in AI gets fixed, and the paper doesn’t stop at the reframe: it proposes combining anti-caste epistemologies with decolonial method, explicitly warning that “Brahmanical forces are now weaponizing decolonial approaches to impose Brahmanical worldviews” if that combination isn’t done carefully. Concretely, the authors argue for testing universal, non-caste-coded markers (food, dress, architecture) to expose default Brahminism rather than only auditing caste labels directly; for building datasets from community narratives and oral histories instead of top-down datafication; for auditing upper-caste representation as rigorously as lower-caste representation, rather than treating only the latter as a problem to fix; and for using local land revenue records as historical counter-evidence against colonial-era caste classifications baked into training data. A categorical fix looks like debiasing each label’s outputs individually: more varied, less stereotyped images of Dalits, of Brahmins, of any caste queried alone. A relational fix tests outputs that put groups in the same frame or the same narrative, the kind of bias a label-by-label audit is structurally unable to see because it never generates the comparison. Fairness researchers working on other axes, gender and race chief among them, have made related arguments before: that debiasing single-category outputs can leave the relationships between categories untouched. Caste, on this account, is a case where the relational structure is not incidental to the bias but definitional to it.
What the paper doesn’t do, by its own account, is consult Dalit or caste-oppressed communities directly — the authors ground their framework in “personal knowledge/histories of growing up in India” rather than named community testimony, a reflexive-standpoint method with limits the paper doesn’t paper over. Relational audits are also harder to design, harder to score, and harder to sell to AI companies that have been slow enough to adopt even the simpler categorical checks. A method that requires generating and coding multi-group scenes, rather than single-label batches, multiplies the audit’s cost well before it multiplies its adoption. Whether FAccT audiences, or the labs whose models get audited, adopt the specific remedies above is not decided by one paper’s presentation.
What the paper does establish, on the evidence available, is that a widely used audit method has a blind spot built into its design, not just its execution. That is a modest claim and a durable one: methodology critiques in AI fairness research tend to outlast the specific systems they were written about, because the next generation of models inherits the same audit tools unless someone has already argued for better ones.
