LLMs Are Biased Toward Japanese Culture - Here's Study Recap

Frontier AI models have a cultural type, and according to new research, that type is Japan. When you ask a large language model an open-ended question about culture without naming a country, it tends to reach for Japanese examples more than any other, a pull that shows up across eight major models and 24 languages.

The finding comes from a white paper published in April by researchers at the University of the Basque Country and Cardiff University. They built a multilingual test set of 31,680 open-ended but culturally grounded questions, spanning 66 cultural subtopics across 11 broader domains. The prompts asked things like which legends explain the land, what role neighbors play, which school subjects carry the most weight, and what people eat at everyday meals. None of them named a country. After each question, the model had to pick a region and answer with specific examples.

The models did not just tilt toward Japan by a hair. The paper describes a "disproportionate prominence of Japan" in responses, consistent enough to hold across the whole 24-language spread rather than clustering in one region's outputs.

The design was meant to expose what the models believe by default.

"This ensures that any regional or cultural assumptions arise from the model's internal priors rather than prompt design."

— The researchers

Language also changed how narrow the models got. The paper reports that in English and other high-resource languages, the models gave more varied answers and leaned less on the country tied to the input language. Lower-resource languages left them with thinner ground to stand on, and they fell back harder on defaults.

The first pattern is the obvious one. Ask in French, get France; ask in Japanese, get Japan. Models lean on whatever country matches the language of the question. The interesting part is what happens when a model looks past that home country. When it reached for an outside reference, it reached for Japan. Six of the eight models preferred Japan in these exogenous answers. The United States came second, then India, China, and France. Across all 24 languages, Japan won seven of the 11 cultural topics once each language's own country was set aside.

The models tested were ChatGPT, Gemini, Claude, Meta Llama, Command-r, Magistral, Qwen, and DeepSeek, so this is not a quirk of one lab. The researchers describe it as a concentration problem, with outputs clustering on a small set of dominant regions. The gap is what makes the result more than a footnote: even models from Chinese and European labs, answering in their own languages, kept drifting back to the same short list.

The pull toward Japan is not hard to explain. Anime, manga, game franchises, and food have been some of the most exported and most discussed pieces of culture online for years, and the models learned from that same well. The internet's long habit of romanticizing Japanese media did not vanish when the training data got scraped; it got encoded. What the data adds is a number: this is not a vibe, it is a measurable tilt that survives translation into two dozen languages.

I run an account about AI, and "the models are weebs" is the funniest research summary I've read all year. The joke writes itself: the internet spent decades fixated on anime and Japanese food, the machines learned from that internet, and now they have a favorite country too. The finding underneath the punchline is quieter and more useful.

The study did not stop at naming the bias. It asked when the bias appears. The team compared model outputs before and after instruction tuning, the training stage that teaches a raw model to answer usefully instead of just producing grammatical text. For that comparison, they used the English answers of Meta Llama, Qwen, Gemma, and Mistral, checking each one before and after the tuning step. Before tuning, the base models had broader taste. The United States still showed up often, but so did Japan, India, China, and several European countries, spread across a wider map.

After tuning, the map shrank.

"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

The effect was strongest in models that went through supervised fine-tuning, where a model learns from curated examples of correct answers, often written or picked by people. That step, the researchers found, pushes outputs toward a handful of dominant regions, and later alignment only nudges it back a little. The homogenization is not baked into the raw internet data; it gets added afterward, when humans decide what a good answer looks like.

The Japan bias is a symptom, and the part I keep coming back to is quieter: a global tool is flattening the world's cultures into two or three defaults. The paper notes this holds even for models built outside the West, so it is not simply American companies exporting American taste. Something about the tuning process itself rewards the safe, dominant answer.

The study frames this as a practical warning rather than a curiosity. For tools deployed across many cultures, the researchers argue, preserving a range of cultural viewpoints "may be critical." That matters most where these systems now sit by default: search, tutoring, translation, and the everyday answers people take at face value.

I don't think it's a scandal, and it doesn't make the models useless. It is a reminder that a model's "neutral" default is a choice someone made, and the choice tends toward the familiar.

Read also, YouTuber sammyuri built a working version of ChatGPT inside Minecraft using only vanilla redstone, a 439-million-block machine running a tiny 5-million-parameter model with a 64-token memory; it can take hours to produce a single reply and often returns nonsense, but as a proof of concept it is striking.

And Cyberpunk 2077's modding scene keeps stretching the game, with one new mod adding a brutal battle axe that leans on the built-in gore and dismemberment systems, and another wiring OpenAI's API into Night City so modders can generate NPC dialogue, side quests, and adaptive enemy behavior on the fly.