As we descend ever further into an AI-fueled dystopia, one question about our future becomes ever more pressing: When the internet is dominated by machines, who’s going to be around to continue its long-standing tradition of fetishizing Japanese culture and media?
Well, there’s no reason to worry, because it turns out the AIs are weebs, too. According to a white paper published back in April, experiments indicated that—when asked open-ended questions about culture—frontier LLMs like Claude, Gemini, and DeepSeek display a “disproportionate prominence of Japan” in their responses.
The study, conducted by a team of University of the Basque Country and Cardiff University researchers, was designed to assess the cultural and regional biases of LLMs. To do so, they constructed a multilingual set of 31,680 “open-ended yet culturally grounded questions,” including prompts in 24 languages that spanned “66 cultural subtopics grouped into 11 higher-level domains.”
The prompts included questions about belief and society, asking “What legends explain the land?” and “What is the role of neighbors?” They asked which subjects are most valued in school, what types of traditional dances exist, what foods are eaten during daily meals—all without directly referencing specific countries and cultures. After being given the initial prompt, the LLMs were then instructed to explicitly select a country or region to provide examples in their response.
“To reduce prompt variability, all questions follow a standardized template, and no explicit regional cues are provided,” the researchers write. “This ensures that any regional or cultural assumptions arise from the model’s internal priors rather than prompt design.”
After prompting eight models—ChatGPT, Gemini, Claude, Meta Llama, Command-r, Magistral, Qwen, and DeepSeek—the most apparent bias is an unsurprising one: Across all models, LLMs will most often answer cultural prompts with whichever country or region is associated with the language used in the prompt. In other words, when asked cultural questions in French, LLMs will tend to give responses referencing France and French culture.
However, whenever LLMs provided an exogenous reference—that is, answers referencing a different country than the one associated with the language used—they had a clear favorite: AI loves talking about Japan.
On average, six out of the eight evaluated models preferred to reference Japan during exogenous responses. The United States came in second, followed by India, China, and France—but across all models, in all 24 languages, LLMs overwhelmingly preferred referencing Japan in seven out of 11 cultural prompt topics when excluding own-country mentions.
“This pattern suggests an uneven regional representation in frontier model outputs, with a strong concentration on a small set of dominant regions,” the researchers say.
The study wasn’t just intended to assess LLM bias, but also when that bias emerges during LLM training. To do so, they compared the English responses of Meta Llama, Qwen, and Gemma, and Mistral models both before and after instruction tuning—the fine-tuning of a pre-trained LLM intended to optimize its ability to provide responses that are useful to the user and not just grammatically correct.
The researchers’ findings indicate that, before instruction tuning, base models displayed a broader set of cultural associations: “While the United States remains prominent” in base model responses to English-language cultural prompts, “substantial references are also made to Japan, India, China, and several European countries.”
After instruction tuning, however, LLMs show more marked cultural preference.
“Across all examined model families, instruction tuning sharply increases alignment with the United States and Japan while reducing references to most other countries,” the researchers write. “This convergence toward culturally dominant regions occurs even in models developed outside Western contexts, indicating that post-training induces a homogenization of cultural perspectives rather than merely reflecting model origin.”
The disparity is particularly pronounced for model variants that have undergone supervised fine-tuning, a type of LLM optimization in which the model is further trained on examples of correct responses—often using human-generated or human-curated datasets.
The researchers found that “SFT sharply increases concentration on a small number of dominant regions (most notably the United States and Japan)”—an effect that’s “only marginally” mitigated by additional instruction alignment. Those findings suggest that the process of curating or authoring “correct” responses injects a cultural bias into the model’s response criteria.
“Taken together, these results demonstrate that instruction tuning systematically reduces cultural diversity in model outputs, steering responses toward a limited set of culturally dominant perspectives, particularly those associated with the United States and Japan,” the study reads. “This finding has important implications for the deployment of instruction-tuned models in cross-cultural or global applications, where preserving diverse cultural viewpoints may be critical.”
