Researchers have developed an audit to disentangle sociocultural signals in multilingual language models (LLMs). They analyzed outputs from 12 LLMs in three languages - English, French, and Chinese - across various tasks. The study found that bias representation varies systematically across languages and tasks, with surface cues often misleadingly attributed to cultural grounding. To address this issue, the researchers created a human-validated framework to separate biases, identity group representation, and cross-cultural patterns. This audit aims to provide a more accurate understanding of LLMs' performance in different sociocultural contexts. --- Why it matters: This study matters because it highlights potential pitfalls in multilingual LLM audits that rely on surface cues rather than actual cultural understanding. By providing a practical framework for separating biases from meaningful cross-cultural patterns, the researchers offer a valuable tool for improving AI model evaluation and development. Source: https://arxiv.org/abs/2608.23026