# Casteist but Not Racist? Quantifying Disparities in Large Language Model Bias between India and the West

**Authors:** Khyati Khandelwal, Manuel Tonneau, Andrew M. Bean, Hannah Rose Kirk, Scott A. Hale  
**arXiv:** [2309.08573](https://arxiv.org/abs/2309.08573)  
**Date:** 2023-09-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.08573v1/

## Key points

- Study titled "Casteist but Not Racist? Quantifying Disparities in Large Language Model Bias between India and the West"
- Addresses societal biases encoded in Large Language Models (LLMs)
- Developed Indian-BhED dataset to quantify stereotypical bias in LLMs
- Dataset includes examples related to caste and religion contexts
- Most LLMs tested exhibit strong biases towards stereotypes in an Indian context compared to a Western context
- Need for more diverse voices when evaluating LLMs
- Instruction Prompting investigated as intervention to mitigate biases
- Instruction Prompting significantly reduces both stereotypical and anti-stereotypical biases for GPT-3.5 in most cases
- Importance of considering biases in LLMs across different cultural contexts and their potential representational harms
- Sheds light on disparities between India and the West regarding LLM bias levels

## Summary

In the study titled "Casteist but Not Racist? Quantifying Disparities in Large Language Model Bias between India and the West," authors Khyati Khandelwal, Manuel Tonneau, Andrew M. Bean, Hannah Rose Kirk, and Scott A. Hale address the issue of societal biases encoded in Large Language Models (LLMs). To bridge the gap between research on bias levels and potential harms in the Global South versus a Western-centric perspective, they developed a novel dataset called Indian-BhED (Indian Bias Evaluation Dataset) to quantify stereotypical bias in popular LLMs. The dataset includes examples related to caste and religion contexts that are both stereotypical and anti-stereotypical. The findings reveal that most LLMs tested exhibit strong biases towards stereotypes in an Indian context compared to a Western context. This highlights the need for more diverse voices when evaluating LLMs. Additionally, Instruction Prompting was investigated as a simple intervention to mitigate such biases. The results show that Instruction Prompting significantly reduces both stereotypical and anti-stereotypical biases for GPT-3.5 in most cases. Overall, this research emphasizes the importance of considering biases in LLMs across different cultural contexts and their potential representational harms. By focusing on an Indian-centric frame, it sheds light on disparities between India and the West regarding LLM bias levels.