研究称主流 AI 回答常偏向日本和美国

Somehow, every AI answer leads back to Japan

Published: 8 May 2026

Last updated: 8 May 2026

• Chris Stokel-Walker Contributor

Costumed dancers perform a scene from the Kagura play 'Momijigari'. Artur Widak/NurPhoto/Getty.

Ask an AI questions about pretty much anything, and they’ll pivot to talking about how great Japan and its culture are. What’s going on?

Ask an AI chatbot a question rooted in anything to do with culture – what foods are eaten daily, what dances are traditional, what film industries exist – and the model will anchor its answer in one of two countries: the United States or Japan.

That's the conclusion of a new study from researchers at the University of the Basque Country and Cardiff University, who tested eight of the world's leading large language models across 24 languages and uncovered theAI’s obsession with Japan.

The researchers built a dataset of 31,680 open-ended cultural questions across 11 broad topics and 66 subtopics, then prompted models to answer while forcing them to pick a specific country or region.

The results were surprisingly skewed: five of the eight models favoured Japan above all other non-native countries, while only two leaned more strongly toward the United States.

Young dancer performs a coordinated routine. Artur Widak/NurPhoto/Getty.

Across GPT-4o-mini, Gemini 2.5 Flash, Claude 3.5 Haiku, Llama 4 Maverick, Command-R, Magistral, DeepSeek and Qwen, the models kept spewing out answers referencing Japan, the US, India, China and France at the expense of almost everywhere else.

GPT-4o-mini alone referenced Japan 944 times in its outputs, second only to references to countries where the prompt language was official.

Big in Japan

The AI kept bringing up Japan and Japanese culture even when the prompt language had no obvious cultural tie to Tokyo. Ask a question of the AI in Bengali, Catalan, Galician, Hausa or Swahili, and Japan still ends up cropping up, suggesting the bias isn't being introduced by users, but baked deep into the models themselves.

To find out where the obsession is coming from, the researchers compared base models that have only been pre-trained on raw internet text with their instruction-tuned counterparts – the polished iterations released to consumers.

Base models distributed cultural references relatively evenly across countries. After supervised fine-tuning, the stage where models are taught to follow instructions using curated example data, the distribution collapsed sharply onto Japan and the US.

Pedestrians in the Ameyoko market in Tokyo. Shawn Goldberg/SOPA Images/LightRocket/Getty.

"The most substantial shift in country reference distributions occurs during supervised fine-tuning," the researchers write, adding that subsequent instruction alignment "only marginally mitigates these effects and does not recover the more balanced distributions observed in base models."

Data in, data out

The conclusion is that the obsession is being trained in deliberately by whoever is curating the fine-tuning datasets that shape modern AI assistants.

While you might think that the internet’s obsession with Japan could be a cause, it turns out it’s not. The skew arrives when humans start telling models what good answers look like, rather than pre-training on the open web.

Beyond an obsession with all things Japan, the findings also found that lower-resource languages produced more self-referential outputs. Models prompted in Amharic, Sundanese, Assamese or Basque overwhelmingly stuck to making references to their home countries, partly because there's less training data to draw on, and partly because they refuse to answer more often.