Regional Bias in Large Language Models
M P V S Gopinadh 1 Kappara Lakshmi Sindhu 1 Soma Sekhar Pandu Ranga Raju P 1 Yesaswini Swarna 1 1 Vishnu Institute of Technology, Bhimavaram, Andhra Pradesh, India · 2026
Read Regional Bias in Large Language Models by M P V S Gopinadh 1 Kappara Lakshmi Sindhu 1 Soma Sekhar Pandu Ranga Raju P 1 Yesaswini Swarna 1 1 Vishnu Institute of Technology, Bhimavaram, Andhra Pradesh, India (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-16349.txt.
Introduction
Gopinadh and coauthors test for place bias in large language models. Their 2026 preprint deals with fair acts of such models. Place bias means a model treats folk or facts by place in a slanted way. Slants can stem from the text hoards the model learned from. The team probes model replies for place tied slants.
Probes span prompts that vary place cues in set ways. Slants found in replies mark where guards are needed. Fair tools must serve users of all places with same care. The study gives a check list for place bias in new models. It comes from a team based in Andhra Pradesh, India.
Catalog record
| CREATOR | M P V S Gopinadh 1 Kappara Lakshmi Sindhu 1 Soma Sekhar Pandu Ranga Raju P 1 Yesaswini Swarna 1 1 Vishnu Institute of Technology, Bhimavaram, Andhra Pradesh, India |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-16349
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Regional Bias in Large Language Models M P V S Gopinadh 1 Kappara Lakshmi Sindhu 1 Soma Sekhar Pandu Ranga Raju P 1 Yesaswini Swarna 1 1 Vishnu Institute of Technology, Bhimavaram, Andhra Pradesh, India Abstract This study investigates regional bias in large language models (LLMs), an emerging concern in AI fairness and global representation. We evaluate ten prominent LLMs: GPT-3.5, GPT-4o, Gemini 1.5 Flash, Gemini 1.0 Pro, Claude 3 Opus, Claude 3.5 Sonnet, Llama 3, Gemma 7B, Mistral 7B, and Vicuna-13B using a dataset of 100 carefully designed prompts that probe forced-choice decisions between regions under contextually neutral scenarios. We introduce FAZE, a prompt-based evaluation framework that measures regional bias on a 10-point scale, where higher scores indicate a stronger tendency to favor specific regions. Experimental results reveal substantial variation in bias levels across models, with GPT-3.5 exhibiting the highest bias score (9.5) and Claude 3.5 Sonnet scoring the lowest (2.5). These findings indicate that regional bias can meaningfully undermine the reliability, fairness, and inclusivity of LLM outputs in real-world, cross-cultural applications. This work contributes to AI fairness research by highlighting the importance of inclusive evaluation frameworks and systematic approaches for identifying and mitigating geographic biases in language models. † † footnotetext: This work was presented at AMRIT-2024. This version includes refined articulation and improved presentation. Keywords Large Language Models \cdot Regional Bias \cdot AI Fairness \cdot Bias Evaluation 1 Introduction Large language models (LLMs) have emerged as transformative tools in artificial intelligence, fundamentally changing how we interact with technology and process information [ 8 ] . These models have demonstrated remarkable capabilities across diverse applications, from content generation to complex reasoning tasks. However, despite their impressive performance, LLMs struggle with a persistent and critical challenge of biased responses that reflect and sometimes amplify existing societal prejudices [ 9 ] . While recent advances in LLM development have promised rapid progress in AI capabilities, they have simultaneously raised important concerns about fairness and equitable representation across different demographic groups [ 12 ] . Bias in LLMs is observed across multiple dimensions, including gender, race, religion, and geographic origin [ 3 ] . These biases can lead to harmful judgments and decisions in real-world applications, potentially promoting discrimination and marginalization of certain social and cultural groups. Among various forms of bias, regional (geographical) bias represents a particularly concerning yet understudied phenomenon. Geographic bias occurs when LLMs exhibit systematic preferences or make errors in their understanding and predictions based on a user’s location or the geographic context of a query [ 2 ] . This form of bias can result in misunderstandings, unfair treatment, and systematic exclusion of specific regions or communities from equitable AI services [ 14 ] . The consequences of geographic bias are especially severe in multilingual and culturally diverse regions, where inaccurate or culturally insensitive model responses can directly affect access to critical information and services. As LLMs become increasingly integrated into essential domains such as education, healthcare, commerce, and public services, their regional preferences create significant disparities in user experiences. These disparities undermine the fundamental principles of fairness and equity that should guide the development and deployment of AI systems globally. While existing research has extensively examined gender and racial biases in language models [ 16 , 13 ] , geographic bias particularly affecting non-Western and global south regions remains significantly understudied. Recent work by Manvi et al. [ 2 ] has begun to highlight how LLMs demonstrate systematic geographical biases, while Decoupes et al. [ 18 ] emphasize the importance of evaluating geographical distortions as a crucial step toward equitable representation. However, comprehensive frameworks for systematically measuring and quantifying regional bias across multiple state-of-the-art LLMs remain limited. This research addresses this gap by conducting a systematic empirical evaluation of regional bias across ten prominent LLMs: GPT-3.5, GPT-4o, Gemini 1.5 Flash, Gemini 1.0 Pro, Claude 3 Opus, Claude 3.5 Sonnet, Llama 3, Gemma 7B, Mistral 7B, and Vicuna-13B. We introduce FAZE (Framework for Analysing Zonal Evaluation), a prompt-based evaluation framework designed to quantify behavioral regional bias by measuring unwarranted region-specific commitments under contextually neutral conditions. Using a carefully curated dataset of 100 contextually neutral prompts that probe forced-choice decisions between regions, we systematically evaluate how different models handle geographic contexts. Our evaluation generates 1,000 model responses, which we analyze to identify patterns of regional preference and bias. The FAZE framework quantifies bias on a 10-point scale, where higher scores indicate greater tendency to favor specific regions. Our findings reveal substantial variation in bias levels across models, with scores ranging from 9.5 (GPT-3.5) to 2.5 (Claude 3.5 Sonnet). These findings suggest that geographic bias varies across model architectures, training methods, and design choices. This work makes an empirical and evaluative contribution to the study of bias in large language models. FAZE is introduced as a lightweight, behavioral screening framework that captures user-facing regional commitment tendencies under controlled neutral conditions. The framework is intended to provide a reproducible and interpretable basis for comparative analysis across models, without attempting to exhaustively characterize all possible manifestations of geographic bias. As AI systems become increasingly embedded in global applications, ensuring these technologies work fairly across all geographic regions is essential for building trust and effectiveness in cross-cultural contexts. This research represents an important step toward developing language models that understand and respect diverse geographic and cultural perspectives worldwide. Throughout this paper, the term regional bias refers specifically to geographic or location-based preferences exhibited by language models, distinct from other demographic dimensions such as nationality, ethnicity, or language. 2 Literature Review The evaluation of bias in large language models has become an increasingly important research area as these systems are deployed in real-world applications. While significant attention has been devoted to examining gender and racial biases [ 16 , 12 ] , geographic and regional biases have only recently begun to receive systematic investigation. Li et al. [ 1 ] developed the Hierarchical Regional Bias (HERB) evaluation method to quantify bias in pre-trained language models using information from sub-region clusters. These frameworks demonstrate the growing recognition that geographic bias requires systematic measurement and evaluation. Recent work has explicitly examined geographic bias as a distinct phenomenon. Manvi et al. [ 2 ] demonstrated that large language models exhibit systematic geographical biases in their world knowledge and reasoning, with particular disparities affecting non-Western regions. Faisal and Anastasopoulos [ 14 ] showed that language models reflect and potentially amplify existing global power imbalances, stemming from imbalanced training data that over-represents certain regions. Lyu et al. [ 4 ] revealed that even within English-speaking contexts, models exhibit preferences for certain geographic varieties. Decoupes et al. [ 18 ] emphasized that evaluating geographical distortions is crucial for achieving equitable representations across different regions. Prompt-based evaluation methods have emerged as effective approaches for examining bias. Alnegheimish et al. [ 5 ] showed that natural sentence prompts provide ecologically valid ways to examine various forms of bias. Nozza et al. [ 6 ] introduced the HONEST framework to assess harmful patterns in sentence completion tasks, demonstrating the value of structured prompt-based evaluation methods for uncovering implicit biases. These approaches offer advantages over purely statistical methods by revealing how biases manifest in actual model outputs that users would encounter in