Large language models (LLMs) excel at creative text generation, but they're not immune to biases and hallucinations. Retrieval Augmented Generation (RAG) helps ground LLMs in factual information, but what happens when those facts are disputed, and across languages? Researchers are exploring this complex issue with a new dataset, BORDIRLINES, designed to evaluate how well RAG systems handle cross-lingual information retrieval, particularly in the context of geopolitical disputes. Imagine asking an AI, "Does this territory belong to Country A or Country B?" The answer can vary drastically depending on the language of the query and the sources the AI consults. BORDIRLINES tackles this by gathering information from Wikipedia articles in multiple languages related to territorial disputes. The researchers then investigate how different combinations of these sources influence an LLM's response. Early experiments reveal that existing RAG systems struggle with consistency across languages. For instance, when asked about Crimea, an LLM might answer "Russia" in Russian but "Ukraine" in English or Ukrainian. Adding information from different languages can further complicate the issue. This research has important implications for building AI systems that can provide balanced and unbiased information in a multilingual world. Future research directions include expanding beyond Wikipedia to other sources and developing frameworks to better reconcile conflicting viewpoints. The ultimate goal is to create AI systems that can navigate the complexities of cross-cultural understanding, ensuring that information is presented fairly and accurately, regardless of language.