The author evaluates multilingual LLMs' factual accuracy using a novel pipeline and highlights geographical biases in fact generation.
