AI models show systematic bias toward recommending Japan for cultural queries

Researchers from universities in the United Kingdom and Spain tested eight leading artificial intelligence models on tens of thousands of open-ended questions about culture and daily life. The prompts were written in 24 languages and purposefully omitted any specific country, asking the systems to choose a location on their own. As reported in a paper hosted on alphaXiv, six of the eight systems repeatedly selected Japan as their top default choice once countries matching the prompt language were excluded.

For everyday users, this tendency shows how artificial intelligence can subtly color our exposure to global topics without us realizing it. If an assistant repeatedly defaults to Japanese cultural traditions or American economic examples when answering broad queries, users get a restricted view of global diversity. The researchers discovered that this regional preference rarely appears during an AI model's raw initial training phase. Instead, the bias emerges sharply during fine-tuning, when human feedback is used to align model answers, implying that human reviewers may be inadvertently pushing software toward a few favored destinations.

It remains unclear whether this preference for Japan comes from the personal habits of human annotators or the specific makeup of fine-tuning datasets. The paper also highlighted that prompting models in less common languages led to much lower answer diversity, often locking the system into narrow responses. This raises an important question for technology companies: as AI tools become the main gateway to general knowledge, how can developers ensure their training teams do not turn a handful of popular cultures into default answers for everyone?