Research Projects

Rationale: India, a culturally diverse country with distinct regional traditions, experiences systematic dominance of majoritarian narratives over minority voices, a phenomenon we term cultural hegemony. Large Language Models (LLMs) inherit this hegemony by learning from datasets that reflect dominant cultural perspectives while marginalizing authentic regional practices. This results in LLMs imposing majoritarian norms as universal truths across India's cultural landscape. Preliminary investigations reveal how cultural hegemony manifests in LLMs, such as depicting Kerala's matrilineal Nayyar traditions through patriarchal lens, mislabeling Bodo Bathouism as Hinduism, and reducing tribal communities to labor stereotypes. This cultural hegemony, as we have preliminarily identified, operates across six dimensions: gender, social, colorism, economic, religious and linguistic, leading to unprecedented digital erasure of India’s cultural diversity and highlighting the need for systematic investigation of hegemonic structures in LLMs. Scientific Objectives: This research aims to map cultural hegemony in LLMs within Indian contexts through four specific objectives: (i) developing a cultural hegemony taxonomy by reviewing sociology case studies to identify how dominant narratives suppress regional authenticity (ii) creating state-specific evaluation benchmarks with culturally representative datasets to detect patterns of cultural hegemony (iii) designing mitigation techniques using Sparse Autoencoder (SAE) frameworks that dismantle dominant bias circuits while preserving cultural variations and (iv) establishing quantitative metrics, including Cultural Authenticity Retention Score (CARS) and Cultural Distinctiveness Preservation (CDP), to measure Indian cultural preservation post-mitigation alongside cultural hegemony reduction. Hypothesis/Model: We hypothesize that cultural hegemony in LLMs arises from the dominance of majoritarian cultural perspectives across Indian states. Targeted interventions can help identify and dismantle these biased narratives while restoring authentic representations of regional cultures. Next, we will utilize the Cultural Bias Intelligence Quotient (C-BiQ) framework to detect cultural hegemony in LLM outputs and implement an SAE-based framework to target and mitigate hegemonic structures. Finally, we will validate our approach using metrics to assess the effectiveness of our interventions in dismantling dominant narratives and restoring authentic representations. Main Experiments: Our methodology includes several key components to investigate cultural hegemony in LLMs within the Indian context. We will start with expert-driven dataset curation to document hegemonic patterns across Indian states, ensuring our datasets reflect diverse regional cultures. Next, we will utilize the C-BiQ framework to detect cultural hegemony in LLM outputs. Additionally, we will implement an SAE-based framework to target and mitigate those hegemonic structures. Finally, we will validate our approach using metrics to assess the effectiveness of our interventions in dismantling dominant narratives and restoring authentic representations. Significance: This work establishes the first systematic framework for identifying cultural hegemony across Indian states, allowing for a nuanced understanding of how dominant narratives affect regional cultural representations. By mapping and analyzing these hegemonic patterns, we aim to provide targeted solutions for countering cultural bias in LLMs. This research would position India as a global leader in developing culturally aware LLMs, offering practical tools for dismantling digital cultural dominance while ensuring the preservation of authentic Indian minority knowledge systems. Ultimately, our work will contribute to a more inclusive and equitable representation of India’s diverse cultural landscape in AI applications.