Study reveals LLMs exhibit a disproportionate bias toward Japan in cultural responses

LLMs Display Unexpected Cultural Preferences in New Research

A white paper published in April 2025 has exposed a curious phenomenon: frontier AI models display what researchers call a "disproportionate prominence of Japan" when responding to cultural questions1

1

. The study, conducted by researchers at the University of the Basque Country and Cardiff University, tested eight major language models—including ChatGPT, Gemini, Claude, DeepSeek, Meta Llama, Command-r, Magistral, and Qwen—to assess the hidden cultural and regional biases of LLMs1

1

2

2

.

The researchers constructed a multilingual set of 31,680 culturally grounded questions spanning 24 languages and 66 cultural subtopics grouped into 11 higher-level domains1

1

. These multilingual prompts asked about belief systems, traditional dances, daily meals, and the role of neighbors—all without explicitly referencing specific countries. After receiving the initial prompt, the models were instructed to select a country or region for their examples.

AI Obsession with Japanese Culture Dominates Exogenous Responses

While LLMs predictably referenced countries associated with the prompt language—French prompts yielded French cultural references—the pattern shifted dramatically when models provided exogenous responses. When answering in a language but referencing a different country's culture, six out of eight evaluated models preferred Japan1

1

. Across all models and all 24 languages, LLMs overwhelmingly referenced Japan in seven out of 11 cultural prompt topics when excluding own-country mentions1

1

.

The United States came in second, followed by India, China, and France—but none approached Japan's dominance1

1

. "This pattern suggests an uneven regional representation in frontier model outputs, with a strong concentration on a small set of dominant regions," the researchers noted1

1

. The fascination mirrors current human cultural trends, as people increasingly want to travel to Japan, consume its media, and engage with its fashion2

2

.

Instruction Tuning Amplifies LLM Cultural Bias

The study went beyond identifying AI bias to pinpoint when these cultural and regional biases in LLMs emerge during training. Researchers compared English responses from Meta Llama, Qwen, Gemma, and Mistral models before and after instruction tuning—the fine-tuning process that optimizes a pre-trained LLM to provide useful responses beyond grammatical correctness1

1

.

Before instruction tuning, base models displayed broader cultural associations, with "substantial references" to Japan, India, China, and several European countries alongside the United States1

1

. After instruction tuning, however, cultural representation in AI narrowed significantly. "Instruction tuning sharply increases alignment with the United States and Japan while reducing references to most other countries," the researchers found1

1

. This convergence occurred even in models developed outside Western contexts, indicating that post-training induces a homogenization of cultural perspectives rather than merely reflecting model origin.

The effect proved particularly pronounced in models undergoing supervised fine-tuning (SFT), where models train on examples of correct responses using human-generated or human-curated datasets. SFT "sharply increases concentration on a small number of dominant regions (most notably the United States and Japan)"—an effect only marginally mitigated by additional instruction alignment1

1

. These findings suggest that curating or authoring "correct" responses injects cultural bias into the model's response criteria, raising questions about whose cultural perspectives shape AI training data and what this means for global users seeking diverse cultural knowledge from these systems.🟡,