Can some languages hold the key to AI truthfulness across *all* languages? Large Language Models (LLMs) are notorious for generating false statements—sometimes confidently making stuff up. This “hallucination” problem poses a major challenge, especially when dealing with multiple languages. New research explores the fascinating possibility of cross-lingual truthfulness transfer—essentially teaching an AI to be more truthful in many languages by focusing on a select few. The researchers built a new benchmark, MTruthfulQA, to test LLM truthfulness across nine diverse languages, revealing a significant truth gap between languages, especially for those less similar to English. To address this, they introduced an innovative approach, *Fact-aware Multilingual Selective Synergy* (FaMSS). Instead of overwhelming the model with data from every language, they strategically selected an “optimal subset” of languages based on a clever analysis of linguistic bias. They found that certain core languages (like German, Chinese, and Arabic) can act as linguistic bridges, carrying truthfulness across the entire model. Using targeted translation instruction tuning, they fine-tuned LLMs using this core subset of languages with parallel factual description data. The results are striking. FaMSS significantly boosted truthfulness across all nine languages in the MTruthfulQA benchmark, including those not directly involved in the fine-tuning process. Moreover, the performance on a multilingual common knowledge benchmark (Cross-MMLU) also improved substantially. This research is an exciting step towards more truthful, multilingual AI. However, the study highlights that multilingual alignment isn’t uniform. Some languages ‘interfere’ with each other, while others act as synergistic partners. The insights from FaMSS point to a more strategic approach to multilingual training, suggesting that more isn’t always better—especially when it comes to teaching AI the truth.