Fair Representation in AI

AI's Regimes of Representation: A Community-Centered Study of Text-to-Image Models in South Asia

The case study evaluates how text-to-image models depict South Asian cultures. It argues that these models often fail to represent cultural concepts with specificity, default to broad visual tropes, and require community-centered evaluation in addition to standardized benchmarks.

Specificity means representing cultural elements in context rather than collapsing them into broad or stereotyped categories. This matters because cultural representation can preserve nuance, but it can also become complicated. Culture is not fixed, uniform, or immune from criticism. Any evaluation of representation must therefore ask whose culture is being represented, who gets to define it, and which internal differences might be overlooked.

Power relations are central. Concepts such as Orientalism, developed by Edward Said, help explain how Western institutions have historically represented parts of Asia, the Middle East, and North Africa as exotic, static, or inferior. When AI systems reproduce similar visual patterns, they risk carrying older representational habits into new technical systems.

I would evaluate model outputs by asking whether they preserve nuance, avoid rigid boundaries between "self" and "other," and represent concepts accurately across regions, languages, and communities. Benchmarks can show aggregate progress, but they may miss the lived significance of errors such as script erasure, generic clothing, or visual stereotypes.

Developers can improve inclusivity by working with communities, expanding datasets across subregions, including local engineers and researchers, and building qualitative harm checks into evaluation. Datasets also need regular updates because culture is dynamic.

My discussion question is: who should be responsible for correcting cultural misrepresentation?

There is no simple answer. Direct community engagement is important, but access is shaped by language, geography, class, and institutional power. Diasporic or elite voices may be easier to reach, but they may not represent everyone. This tension makes community-centered evaluation necessary and difficult.

I liked this exercise because it sits directly at the intersection of computation and social science. It showed how technical evaluation can benefit from interpretive and historical analysis.