In the competition for cultural influence, AI has a default country

Artificial intelligence is quietly reshaping the global contest for cultural influence. Unlike traditional search engines, which present users with competing options, AI chatbots increasingly make the choice for them. And when asked an open-ended question about family life, culture or tradition, they tend to choose one country over and over again: Japan. Confused? You’re not alone.

“What values shape family life in ______? Choose the place yourself. Be brief.” Researchers at the University of the Basque Country and Cardiff University put that question to eight AI models in 24 languages in 31,680 variations on festivals, food and family. Between 43% and 78% of references, depending on the model, stayed at home. Set those aside, and six of the eight models reached for the same country: Japan.

Call that a favorite and you misread it. A favorite is picked against alternatives. Here nobody offered any. The blank carried no preference and the machines filled it with the same place again and again. That is a default, and the power of a default is that it never has to win again. Most people never question it.

Marketers know the old contest. Google returned 10 links, sold the top slots, published enough of its logic to be studied and recalculated with every query. Japan’s tourism board could bid, optimize and measure. Now many people put the question to a chatbot instead, and a chatbot answers rather than lists.

Ask it an open question and it names one place. In English, Japan drew 2,893 mentions to America’s 2,028; even Alibaba’s Qwen named Japan 1,861 times and China 434. Attention that search spread across a page, the model gathers into a few names. And when a model answers from what it has learned rather than from a live search, the preference was set in training, long before the user typed a word, and it surfaces wherever a question leaves the choice open. No advertiser can buy the country a model recalls first from training. There is no keyword tool and no console.

The authors do not say why Japan and not France, which ranks fifth despite an export catalogue of its own. The internet says it knows. Since July, near-identical posts on X have announced that researchers “finally figured out why,” and one version supplied a tidy answer: Japan is the safe choice, a culture no filter objects to. The paper does not test it and training data does not settle it either: the authors found Japan only modestly present in the post-training data of the open models they checked. OLMo-3, an open model from the Allen Institute for AI, saw China mentioned 52,509 times in fine-tuning and named it 26 times when asked 1,320 English questions.

What the paper does show is when. In English-only tests of open models, base models trained on raw text scatter their references. Supervised fine-tuning, where people teach models how to answer, narrows them: OLMo-2-32B named Japan 61 times before that stage and 361 after. Later alignment trimmed that to 221, still more than triple the starting count. That is one open model, not proof about the eight commercial ones. But it is the only lever the paper points to and it sits in a stage Tokyo does not run.

Is the machine a mirror or a maker? Decades of Japanese film, food and design sit in its training text, and the Cool Japan Fund, set up in 2013 after Douglas McGray coined “Gross National Cool” in 2002, was an attempt to cash that appeal in. A mirror would explain why Japan. It would not explain the count rising nearly sixfold between two stages of one model’s training. The machine did not invent Japan’s appeal. Training turned up the volume.

Europe has been here before. After Japan appeared at the 1867 Paris Exposition, prints and ceramics became fashionable and Van Gogh copied Hiroshige. Japonisme was Europe’s picture of Japan, not Japan’s picture of itself. Paris bought the prints; it did not learn the language. The machines show the same selectivity. Japan leads on beliefs, education, arts, food and health; America on geography, politics, history and the economy. The model reaches for Japan to illustrate culture and for America to explain power.

The price of a default is paid by everyone else. None of the study’s 12 most-referenced countries is in Southeast Asia, a region of about 700 million people. Even South Korea, whose pop culture fills playlists worldwide, ranks third only in media and entertainment. Asked in Indonesian, models named Japan 411 times and America 246. Asked in Sundanese, spoken by about 32 million people in West Java, 79% of references stayed close to home, and answer diversity was among the lowest of the 24 languages.

Start by measuring. The dataset is public. Japan’s tourism board and exporters should run it on the models their customers consult; ministries in Jakarta, Hanoi and Seoul can follow. Singapore committed 70 million Singaporean dollars in 2023 to a national program for Southeast Asian language models, SEA-LION among them, and SEA-LION was not among the eight tested. Sakana AI launched Sakana Chat in March for Japanese language and social norms, but a domestic model cannot decide what foreign systems tell the world. The aim is not to game the ranking. It is to learn whether the machine that names you can also describe you.

For a generation, Japan worried about whether the world was listening. The machines are eager to talk about it.

Countries have spent decades competing to be chosen. What happens when the choice is made before the competition begins?

David Kim is a former investment banker turned journalist who writes at the intersection of capital, technology and geopolitics.