Executive Summary

This study investigates the language dependencies of Large Language Models’ (LLMs) values, ethics, and beliefs across 20 languages, revealing critical implications for AI safety and deployment. Through systematic analysis of responses from four leading LLMs (GPT-4, GPT-3.5, Mistral-large, Claude-opus) to controversial statements, scientific claims, and global issues, we uncovered significant variations in ethical reasoning and safety mechanism effectiveness across different language families.