On October 24, 2026, a research paper titled “WikiOriginQA: Automated Construction of a QA Dataset with Knowledge Origins for Analyzing Cultural Bias” was published on the arXiv platform (ID 2610.30677v1), officially introducing WikiOriginQA—a automated question-answering dataset designed for analyzing cultural bias.
This dataset abandons the traditional approach of relying on crowd-sourced annotators familiar with various cultures, instead using the creation dates of different-language Wikipedia editions as an objective proxy for the origins of knowledge, enabling fully automated construction. Experimental results show that current large language models (LLMs) perform better when handling questions rooted in Anglophone cultural contexts, confirming their potential cultural bias; at the same time, they highlight the critical role of knowledge propagation into the Anglosphere.
The study emphasizes that this approach significantly enhances the scalability of cultural bias assessment, enabling systematic evaluations across multilingual settings without relying on extensive human annotation. The paper explicitly states that this work constitutes methodological innovation and empirical analysis, and does not involve training specific models or commercial deployment.