AI Models Keep Picking Japan When Asked About Culture
A study of eight language models found that Japan appeared most often in cultural answers that referenced a country outside the prompt's language.
Large language models may not have personal interests, but their answers can still show strong cultural patterns. A study from researchers at the University of the Basque Country and Cardiff University found that Japan appeared unusually often when eight AI models answered open-ended questions about culture.
The researchers tested ChatGPT, Gemini, Claude, Meta Llama, Command-r, Magistral, Qwen, and DeepSeek with a multilingual collection of 31,680 questions. The prompts covered 24 languages, 66 cultural subtopics, and 11 broader areas, including beliefs, society, education, traditional dance, food, and community life. The questions did not name a country directly.
Instead, each model was asked to choose a country or region to use in its answer. The full methodology and findings are available in the published research paper.
Japan was the most common outside reference
The models usually connected a prompt to the country associated with its language. A question written in French, for example, generally led to references to France or French culture.
The pattern changed when a response referred to a different country. In those cases, Japan was the most common choice for six of the eight models on average. The United States ranked second, followed by India, China, and France.
Across the tested languages, Japan appeared more often than any other outside reference in seven of the 11 cultural topic areas. The researchers described this as an uneven spread of regional representation, with model answers concentrating on a small group of culturally dominant places.
Training changed the models’ cultural answers
The study also compared models before and after instruction tuning. This process fine-tunes a pre-trained model so it can provide more useful responses, rather than simply producing grammatically correct text. A plain-language explanation of instruction tuning describes the same type of post-training process.
Before that tuning, the models connected English-language cultural questions to a wider range of places. The United States remained common, but Japan, India, China, and several European countries also appeared frequently.
After tuning, answers focused more heavily on the United States and Japan, while references to many other countries fell. The researchers observed the same pattern in models created outside Western countries, suggesting that post-training data and response guidelines played a larger role than the model’s origin alone.
Supervised fine-tuning produced an even stronger concentration around the United States and Japan. Extra instruction alignment only reduced that effect slightly, according to the study.
The researchers argued that human-written or human-selected examples of “correct” answers can bring cultural assumptions into a model’s idea of what a good response looks like. Their conclusion was that instruction tuning can reduce the range of cultural viewpoints in AI-generated answers, which may create problems when those systems are used in global or cross-cultural settings.
So, no, the models are not literally anime fans. Their training data and tuning simply make Japan a frequent answer when they need to reach for a culture outside the one implied by the prompt. Still, that is a pretty funny result for a study about bias.
What do you think about AI models repeatedly turning to the same countries for cultural examples? Share your thoughts in the comments, and follow us on X, Bluesky, YouTube, and Instagram.
