Purdah and Patriarchy: Evaluating and Mitigating South Asian Biases in Open-Ended Multilingual LLM Generations

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

This work addresses overlooked intersectional and culturally specific biases in Large Language Model (LLM) outputs, particularly within underrepresented multilingual regions like South Asia. Researchers conducted a multilingual and intersectional analysis across 10 Indo-Aryan and Dravidian languages, identifying how cultural stigmas influenced by purdah and patriarchy are reinforced in generative tasks. A novel, culturally grounded bias lexicon was constructed to capture previously unexplored intersectional dimensions, including gender, religion, marital status, and number of children. This lexicon was used to quantify intersectional bias and assess the effectiveness of self-debiasing in open-ended generations such as storytelling and to-do lists, where bias often manifests subtly. The study also evaluated two self-debiasing strategies, simple and complex prompts, for their efficacy in reducing culturally specific bias in these languages.

Key takeaway

For NLP Engineers developing multilingual LLMs for South Asian markets, you must move beyond Eurocentric bias evaluations. Your models likely reinforce cultural stigmas related to purdah and patriarchy in generative tasks. Implement culturally grounded bias lexicons and evaluation frameworks to identify these subtle biases. Experiment with both simple and complex self-debiasing prompts to mitigate culturally specific biases effectively, ensuring more equitable and appropriate model outputs for diverse user groups.

Key insights

LLMs reinforce South Asian cultural biases related to purdah and patriarchy, requiring culturally-grounded evaluation and mitigation.

Principles

• Intersectional and culturally specific biases are often overlooked.
• Bias manifests subtly in open-ended generations.
• Eurocentric bias evaluations are insufficient for diverse contexts.

Method

Constructed a culturally grounded bias lexicon for 10 Indo-Aryan and Dravidian languages, then quantified intersectional bias in open-ended LLM generations, and evaluated simple and complex self-debiasing prompts.

In practice

• Develop culturally specific bias lexicons.
• Evaluate LLM bias in open-ended generation tasks.
• Test simple and complex self-debiasing prompts.

Topics

• Large Language Models
• Multilingual NLP
• Algorithmic Bias
• Cultural Bias
• South Asia
• Bias Mitigation
• Generative AI

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• Cross-Lingual Consensus: Aligning Multilingual Cultural Knowledge via Multilingual Self-Consistency — LLMs' latent cultural knowledge in local languages can be surfaced and transferred to improve English query performance and reduce Western bias.
• Gender bias in machine translation: an evaluation in the English-Portuguese language pair — Machine translation systems exhibit gender bias, favoring masculine and stereotyped occupational translations from English to Portuguese.
• Cognitive Effects and Biases in Large Language Models — This tutorial clarifies cognitive effects and biases in LLMs by bridging psychology and NLP.
• Evaluating Pluralism in LLMs through Latent Perspectives — LLMs struggle to represent diverse perspectives, especially rarer ones, diverging from human text distributions.
• Machine ‘culture’ beyond language — Language reliably shifts an LLM's "psychological profile," influencing social orientation and cognitive style.
• Isolating LLM Lexical Bias: A Curation-Free Triangulated Metric for Preference-Stage Learning — A new metric quantifies LLM lexical bias from preference learning without manual curation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.