Have you ever wondered what's going on inside the "brain" of a multilingual AI? New research reveals surprising insights into how these models process and generate different languages. It turns out they don't simply switch between languages like a human interpreter. Instead, they rely on a hidden "internal" language, often related to the languages they were trained on. For example, a model trained primarily on English might still "think" in English even when generating text in Japanese, using English as a bridge. Researchers investigated three types of Japanese language models: one trained mostly on English (Llama 2), one initially trained on English then further trained on Japanese (Swallow), and one trained equally on English and Japanese (LLM-jp). They found that while the English-centric model consistently used English internally, the other two models showed more complex behavior. These models employed a mix of English and Japanese internally, with the "closest" language to the target output being more dominant in the model's internal processing. So, when generating Chinese, the model used more Japanese internally due to shared characters. But when generating French, English took precedence due to closer linguistic roots. This discovery has significant implications for understanding how these models work and how to improve their multilingual capabilities. For instance, if a model relies too heavily on its dominant training language, it may struggle to grasp subtle nuances or cultural contexts in other languages. This also explains why efforts to “edit” AI models’ knowledge should focus on the model’s original training language. This research opens exciting avenues for building even more sophisticated and truly multilingual AI models that can seamlessly navigate the complexities of human language.