Study Reveals US-Centric Bias in Multilingual AI Language Models

A research paper published on arXiv (ID: 2604.19292v1) introduces LocQA, a novel test set designed to measure biases in multilingual large language models (LLMs). The dataset contains 2,156 locale-ambiguous questions across 12 languages, covering facts about laws, dates, and measurements that vary by region. These questions intentionally omit explicit locale indicators, forcing models to rely on implicit priors. Researchers evaluated 32 different LLMs using this methodology. The study identified two distinct structural biases within these AI systems. Inter-lingual analysis revealed a pronounced global bias toward answers relevant to the United States, even when queries were posed in languages other than English. This demonstrates how knowledge and cultural norms can propagate across linguistic boundaries within AI systems, despite advancements in multilingual fluency. The work aims to quantify both inter-lingual and intra-lingual model biases through their performance on ambiguous geographical questions.