Ask a chatbot to choose any place on Earth and describe how families there welcome the new year, and there is a fair chance it will choose Japan. That curious tendency has attracted attention following a study of cultural preferences in large language models (LLMs). An October 21 post on X went further, claiming researchers had proved that every major AI model was secretly obsessed with Japan — and had finally discovered why.
The study behind that claim is real, and its findings raise important questions about how AI systems represent the world. But the viral interpretation mixed genuine research with unsupported explanations, including claims about corporate safety training, model compression, and the influence of anime communities.
The research tells a more nuanced story. Japan does emerge as a surprisingly frequent cultural reference in many models, but the preference is neither universal nor equally strong across languages. More importantly, the researchers found evidence that the tendency becomes substantially more pronounced during a particular stage of AI training. What they could not establish was precisely why.
What the researchers actually tested
The study, Why Are All LLMs Obsessed With Japanese Culture? On the Hidden Cultural and Regional Biases of LLMs, was first posted to arXiv in April 2026 and subsequently published in October in Findings of the Association for Computational Linguistics: EMNLP 2026. Its authors are Joseba Fernandez de Landa of the HiTZ Center at the University of the Basque Country, together with Carla Perez-Almendros and Jose Camacho-Collados of Cardiff University.
Rather than testing how accurately AI systems understand particular cultures, the researchers wanted to discover which countries the models spontaneously choose when responding to questions without a specified geographic setting.
The team developed a dataset of 1,320 open-ended questions covering 11 broad cultural domains and 66 subtopics, including beliefs, family life, education, the arts, food, health, politics, history, and the economy. The questions were generated with assistance from GPT-5.1, then manually reviewed to remove repetition and unintended geographic references.
Each question deliberately left the location unspecified. Instead of asking what values shape family life in Brazil, for example, a question might ask what values shape family life in a particular region, leaving the model to select the place itself.
The questions were translated into 23 additional languages, creating 31,680 prompts for each model. The languages ranged from English, Chinese, Hindi, Spanish, and Arabic to less extensively represented languages such as Swahili, Hausa, Amharic, Sundanese, Assamese, Galician, and Basque.
Eight AI models were evaluated: GPT-4o mini, Gemini 2.5 Flash, Claude 3.5 Haiku, Llama 4 Maverick, Command R, Magistral Small, DeepSeek V3.2, and Qwen3-Next. The researchers accessed them through OpenRouter using specified generation settings.
A separate AI system then examined the responses to identify the countries or regions mentioned. This automated classification method achieved approximately 98 percent accuracy on a validation set of 264 items. Although that is a strong result, it also means the findings depend partly on an automated evaluator that may occasionally miss geographic or linguistic distinctions.
The resulting dataset allowed researchers to examine something that conventional cultural-knowledge tests often overlook: when an AI system is free to choose a cultural example, which part of the world does it bring into the conversation?
Most models favored countries associated with the question’s language
The study’s largest finding is also the one most easily overlooked in discussions about Japan. Before looking abroad, AI models frequently chose countries where the language of the question is officially spoken.
Depending on the model, between 43 and 78 percent of responses referenced a country where the input language was official. A question written in Greek, for example, was relatively likely to produce an answer involving Greece. This pattern suggests that language itself provides a powerful geographic cue, even when the question contains no explicit location.
Japan’s prominence became apparent when researchers excluded those language-associated references and examined which other countries the models selected.
Here the results were striking. Japan was the most frequently referenced country outside the input language’s official-language regions for five of the eight models, and tied with the United States in a sixth. Gemini 2.5 Flash referenced Japan and the United States exactly 1,493 times each. Claude, Llama, Magistral, Qwen, and DeepSeek all favored Japan, while GPT-4o mini and Command R favored the United States.
Some differences were substantial. Llama 4 Maverick referenced Japan 2,701 times compared with 1,074 references to the United States. Qwen3-Next mentioned Japan 1,861 times compared with 483 references to the United States.
The preference also varied by subject. Across the 11 cultural domains, Japan was particularly prominent in areas such as beliefs, education, the arts, food, and health. The United States led in geography and the economy, while Japan fell to third place in politics and fourth in history.
These findings reveal a concentration of cultural references around a small number of countries, especially Japan and the United States. India and China also appeared frequently, but generally received considerably fewer references when language-associated countries were excluded.
Japan’s prominence is notable given that it accounts for only a small share of the world’s population. However, population alone cannot establish how frequently a country should appear in these answers. The models were selecting illustrative cultural examples, not drawing countries from a population-weighted sample. The stronger evidence of uneven representation lies in the repeated preference for a narrow group of countries and how that preference changes across languages and training stages.
The language of the question changes the cultural perspective
The viral post suggested that AI models default to Japanese culture even when prompted in languages with little cultural connection to Japan. The actual findings are more complicated.
The language used to ask a question substantially affects both the geographic focus and variety of the response.
In English, only about 31 percent of references were to countries where English is an official language. The answers also covered a comparatively broad range of countries. In Assamese, by contrast, approximately 83 percent of references involved countries associated with the language, and responses showed substantially less geographic variety.
The researchers found a strong statistical relationship between the amount of material available in a language within Common Crawl — a large collection of publicly accessible web content — and the diversity of countries mentioned in AI responses.
Languages with extensive online representation, such as English, French, Russian, and Spanish, generally produced a broader range of cultural references. Languages with less available digital content tended to generate more geographically restricted answers or more instances in which the model failed to provide a usable response.
The researchers used Common Crawl coverage as an indicator of language-resource availability, not as a complete record of what the models had actually been trained on. The correlation therefore suggests an important relationship without proving that differences in web content alone caused the observed behavior.
Japan appeared in responses written in Swahili, Hausa, Amharic, and Basque, but the United States appeared more frequently in each of those languages. English produced the largest overall number of references to Japan.
The results therefore contradict the idea that Japanese culture dominates AI responses equally across languages. Instead, they suggest that an AI system’s apparent cultural perspective depends partly on the language in which it is asked to express that perspective.
When does the preference emerge?
Perhaps the study’s most important contribution concerns the stage of AI training at which these geographic preferences become more pronounced.
Modern chatbots undergo several stages of development. During pre-training, a model processes enormous amounts of text and other training material, learning statistical relationships that allow it to predict and generate language. The resulting base model possesses substantial linguistic and factual knowledge but is not necessarily optimized to behave like a conversational assistant.
The next stage commonly involves supervised fine-tuning, or SFT. During this process, the model learns from examples of instructions paired with desired responses. It becomes better at following requests, answering questions, and presenting information in a helpful format.
Additional post-training methods may then refine the model’s behavior using human or AI-generated preferences, reinforcement learning, and other alignment techniques.
To investigate when geographic preferences emerge, the researchers compared base models with instruction-tuned versions from several open model families, including Llama 3.1, Qwen 2.5, Gemma 2, and Mistral. They also examined OLMo models developed by the Allen Institute for AI, whose publicly available intermediate checkpoints made it possible to distinguish changes occurring during supervised fine-tuning from those associated with subsequent training stages.
The base models generally produced a broader distribution of geographic references. The United States remained prominent, but Japan, India, China, and several European countries also appeared regularly.
After instruction tuning, however, references became more heavily concentrated on Japan and the United States.
The OLMo checkpoints helped identify when this change occurred. The strongest shift appeared during supervised fine-tuning rather than the later alignment stage.
For example, in the smaller OLMo 2 model, a measure of how evenly references were distributed across countries — known as normalized entropy — fell from 0.84 in the base model to 0.63 after supervised fine-tuning. Subsequent alignment changed the measure only slightly, to 0.64.
Interestingly, the fine-tuned models sometimes mentioned more distinct countries overall while concentrating a larger proportion of responses on a handful of favored locations. Their geographic vocabulary had not necessarily become smaller; the distribution of their choices had become less balanced.
This distinction matters. The experiments suggest that supervised fine-tuning substantially amplifies preferences already present in some form, rather than establishing that pre-training contributes nothing to them.
The researchers also cautioned that their stage-by-stage comparisons were conducted in English only. The findings therefore cannot establish that the same training-stage effects occur identically across all languages.
Why Japan? The researchers have not established a cause
Identifying when a preference becomes stronger is not the same as discovering why it develops.
The researchers investigated whether the frequency with which countries appeared in post-training datasets could explain the countries favored in model outputs. Their analysis, presented in the paper’s appendix, revealed a more complicated relationship.
For example, France appeared frequently in the OLMo 2 fine-tuning and preference-training datasets, sometimes more often than the United States. Yet France remained relatively uncommon in the model’s generated cultural references, while Japan appeared far more frequently despite having a smaller presence in those datasets.
The contrast was even more pronounced in some OLMo 3 experiments.
China received 52,509 mentions in one supervised fine-tuning dataset, compared with 20,883 mentions of Japan. Yet the final OLMo 3 instruction-tuned model referenced Japan 478 times and China only 19 times in the corresponding English-language evaluation.
These results show that simply counting how often a country appears in training examples does not adequately explain which countries the model later favors. The context in which a country appears, the types of tasks used during fine-tuning, and the relationships learned during training may all matter, but the study did not isolate a definitive mechanism.
The authors concluded that the underlying question of why these preferences develop remains unresolved.
There are plausible possibilities. Japanese culture is extensively discussed in English-language media and has many widely recognizable traditions, from tea ceremonies and cherry blossom viewing to seasonal festivals and distinctive artistic forms. Such topics could make Japan a convenient example in instructional material designed to teach AI systems how to respond to cultural questions.
But that remains a hypothesis, not an established explanation. The study does not demonstrate that these traditions, the popularity of Japanese entertainment, or any particular collection of training examples caused the preference.
What the viral explanation got wrong
The October 21 X post offered a much more specific explanation than the research supported.
It claimed that AI models acquire their preference for Japan during the safety-training phase, when developers teach them to be helpful, harmless, and polite. According to the post, models begin mapping corporate safety guidelines onto Japanese cultural concepts, producing a disproportionate preference for Japanese norms and values.
The researchers did not establish such a connection.
Their experiments showed that the most substantial change occurred during supervised fine-tuning. The subsequent alignment stages did not produce a comparable increase in geographic concentration. Moreover, the research did not identify Japanese cultural values as an influence on corporate safety guidelines or demonstrate that politeness-related training was responsible.
The post also claimed that quantization — the process of reducing the numerical precision of a model to lower its computational and memory requirements — amplifies the preference for Japan.
This claim is absent from the study. The researchers did not test quantization or compare compressed models with their full-precision counterparts.
Related research has examined whether model compression affects other forms of bias. A 2026 study by Plawan Kumar Rath and Rahul Maliakkal found that aggressive quantization could cause previously unbiased responses to display stereotypical behavior in a demographic-bias benchmark. Separate research led by Federico Marcuzzi found that quantization could slightly increase stereotypes and unfairness in certain text-generation tasks, particularly under heavy compression.
These studies provide legitimate reasons to investigate the relationship between compression and AI behavior. Neither, however, established that quantization increases references to Japan or amplifies the geographic preferences identified in the cultural-bias study.
Another unsupported assertion was that the preference affects how models generate stories, present facts, and write computer code. The researchers evaluated responses to open-ended cultural questions, not performance across those downstream applications.
The observed preferences might influence other tasks, but the study provides no direct evidence demonstrating that they do.
Finally, the viral post attributed the models’ cultural preferences to the interests of active Reddit users and contributors to anime-related websites. That explanation was not established by the research either.
Online material undoubtedly contributes to many AI training datasets, but the researchers’ central observation was that geographic preferences became more concentrated during post-training. Their findings do not identify anime communities, Reddit users, or any specific online subculture as the cause.
What the findings mean for AI users
The study reveals a measurable pattern in the cultural examples chosen by the tested AI systems, but several limitations are important when interpreting it.
First, the researchers examined specific model versions available during their experiments. The results should not automatically be extended to newer systems or every AI model currently in use. Training procedures, datasets, and model behavior can change substantially between versions.
Second, the study measured geographic references, not the quality of cultural understanding. A model that frequently mentions Japan is not necessarily knowledgeable about Japanese society, just as a model that seldom mentions another country does not necessarily lack information about it.
The frequency of a reference tells researchers something about a model’s response tendencies. It does not establish whether the model views a country favorably, understands its traditions accurately, or can explain its culture with appropriate depth.
Third, the results concern questions that deliberately left the location open. Real-world users often provide geographic context, ask about specific cultures, or request comparisons among multiple societies. Those instructions can change which examples a model selects.
For ordinary users, the practical lesson is to be specific about the cultural perspectives they want. Someone interested in family traditions across Asia, for example, may receive a more useful answer by explicitly requesting examples from Japan, Korea, China, India, and Southeast Asia than by asking the model to choose representative traditions on its own.
The findings also raise broader questions for AI developers. Training methods designed to make models more useful and responsive may unintentionally reinforce particular cultural reference patterns. Understanding those effects could help developers improve the diversity and relevance of responses without sacrificing the benefits of instruction tuning.
For speakers of languages with less extensive digital representation, the issue is especially significant. The researchers found that these languages often produced narrower geographic coverage and more unsuccessful responses, potentially limiting the range of cultural perspectives available through AI systems.
What we teach AI shapes what it chooses to show us
The most important finding is not that chatbots have developed an affection for Japan. The study offers no evidence that AI systems possess such preferences in a human sense. Rather, it demonstrates that the processes used to teach models how to respond can substantially influence which cultural examples they select.
The distinction between what a model knows and what it habitually mentions is important. A system may contain information about a wide range of countries yet repeatedly draw on only a few when asked to provide examples. Those recurring choices can shape the geographic perspective presented to users, even without an explicit instruction to favor any particular culture.
An old Chinese saying observes that those who stay near vermilion turn red, while those who stay near ink turn black. The comparison is a metaphor rather than an explanation of machine learning, but it offers a useful reminder that the influences shaping behavior deserve examination.
In AI development, those influences include not only the enormous datasets used during pre-training but also the examples, objectives, and feedback applied afterward. The study indicates that these later stages can have a substantial effect on the cultural references models produce, even though the precise reasons remain uncertain.
The research leaves an important question open: Why does instruction tuning make Japan such a frequent choice when the training data’s raw country frequencies do not adequately explain the result?
Answering that question will require further investigation. In the meantime, the findings give developers a reason to examine the cultural assumptions reflected in their training methods — and give users a reason to look beyond the first example an AI assistant happens to offer.
Follow us on X, Facebook, or Pinterest
