How text-to-image models encode caste, and why audits that check one label at a time miss it

A text-to-image model turns a written prompt into a picture. Ask it for “an Indian person”, and whose face and home appear? Researchers have asked that question about caste since 2023. Their studies find Dalits shown through poverty, and a 2026 paper argues that most audits look for caste the wrong way: one label at a time, when caste is a ranking between groups.

What have the audits found?

In a 2024 paper presented at the AAAI/ACM Conference on AI, Ethics, and Society, run by two computing societies, Sourojit Ghosh examined how Stable Diffusion, an image generator, depicts people of different castes, generating 100 images per prompt and comparing them with the model’s default depictions of an “Indian person”. The outputs, the paper reports, “perpetuate systems of ‘castelessness’”, equating “Indianness with high-castes” and depicting “caste-oppressed identities with markers of poverty”. It also reports that Dalits are “prominently depicted as living in rural areas and always at protests”.

A 2025 paper by Ghosh, Sanjana Gautam, Pranav Venkit and Avijit Ghosh, Documenting Patterns of Exoticism of Marginalized Populations within Text-to-Image Generators, looked at everyday activities such as going to work across 13 countries it labels “Global South”, India among them, and compared them with the USA, the UK and Australia. It found that “individuals from ‘Global South’ countries are depicted in similar attire irrespective of the performed activity”, and that these effects fall “doubly upon marginalized populations within ‘Global South’ countries”.

A 2023 FAccT study by Rida Qadri, Renee Shelby, Cynthia Bennett and Remi Denton, AI’s Regimes of Representation, took a different route: it treated South Asian communities as experts on what the models get wrong. Its abstract concludes that generative AI “can reproduce an outsiders [sic] gaze for viewing South Asian cultures”.

Why does it matter how audits count caste?

Audits such as Ghosh’s 2024 study check one label at a time: do pictures of Dalits skew toward poverty or degraded settings? A paper at FAccT’26, the ACM Conference on Fairness, Accountability, and Transparency, by Divyanshu Kumar Singh and five co-authors, argues that this misses what makes caste bias caste bias. Caste is a ranked relationship between groups, so a model can render each group without an obvious stereotype and still rank them against each other. The authors call the unmarked order behind this “Brahmanical Normativity”, after the Brahmins, the dominant priestly caste, and ask:

“why does the model enforces [sic] and justifies ‘dignity’ for lower-caste, but upper-caste have a default right to dignity?”

— Singh, Das, Subramanian, Saha, Voida and Semaan

i

Source: FAccT’26 paper

Their proposals include evaluating caste through broader markers such as food, culture and architecture rather than caste names, and questioning how models imagine dominant castes, not only oppressed ones. We set out the paper’s argument in August.

What are the limits?

The 2026 paper does not report consulting Dalit communities directly: the authors compiled names from online caste-surname information, which they complemented with, in their words, “our own personal knowledge and histories of growing up in India”. They name a limit of their own method: names are “a form of category”, so the audit still leans on caste labels. The 2026 audit generated its images with Gemini, a text-to-image model, and Ghosh’s 2024 paper tested one model, Stable Diffusion; the authors list testing more models as future work.

The 2023 study did ask. A Dalit academic from India, interviewed for it and identified only as a participant, reviewed what the models made of “daily life of a Dalit person” and found dirt, hardship and poverty, and no Dalit celebrations or dance. The images, she said, missed the “dynamic aspect of Dalit identity”. Representations of Dalit life should “also [be] about their songs, about their cultures, about how they make a difference through their everyday acts”, not “abject poverty… as the only marker of Dalit life”, the study reports.

Key terms

• Text-to-image model: a system that turns a written prompt into a picture.
• Audit: a study that prompts a model many times and compares who and what appears.
• Brahmanical Normativity: the 2026 paper’s name for an unmarked order governing who a model renders with dignity or authority.
• FAccT: the ACM Conference on Fairness, Accountability, and Transparency, where the 2023 and 2026 papers were presented.

What to watch

None of these has a date yet.

• Two groups in one frame. Whether audits test scenes that place castes together.
• Dominant castes. Whether audits examine how models render them as closely as they examine Dalits.
• More models. The 2026 authors, who generated their images with Gemini, list testing more models as future work.
• Who audits. Whether Dalit researchers and communities lead audits. Qadri and colleagues, in a study of South Asian communities, treated communities as experts.