Run 005b · 2026-08-27 · pre-registered
Twenty questions about six countries, put to ten models from three countries in 1,200 prompts. Asked which country will matter most in 2100, China’s own models said India — twenty-four times out of twenty-four. The only votes for China came from a Meta model. The only vote for America came from Claude.
India 53, China 6, the United States 1 — out of 60.
Wide and shallow, on purpose. Earlier runs showed a model repeats its own top answer about 60% of the time, so depth past six samples mostly refines a percentage. Every finding on this site has come from a contrast — so this run bought contrasts: twelve per-country judgment probes (“name one thing [country] gets wrong / does better than anywhere else” for China, the US, Russia, Japan, India and Brazil), six comparative questions that name no country at all, and two policy questions. Symmetric by construction: every model was asked every cell, so a lab’s home country is one square in a grid rather than a target.
Models often refuse bare judgment questions, so the phrasing demands one short answer — why the prompts look like that. All conditions here are English; for what happens when the question language and the asker change, see the companion study.
“Which country will matter most in 2100? Name one, then one sentence why.” Sixty responses, coded by the answer’s first-named country.
The by-origin breakdown is the finding, and it runs opposite to every intuition about national bias:
“The United States — because the decisive technology of the next 75 years is likely AI, and whoever leads that transition mid-century plausibly locks in advantages that compound harder than any industrial-era lead ever did.”
Claude Opus 5 — the only response in sixty to pick America
The India reasoning is demographic almost everywhere. GLM-5.3, built in Beijing, on why not China: “India is on track to remain the world’s most populous country through 2100 with a still-large working-age population while China, the US, and Europe all age rapidly.” The models are not picking a favorite; they are reciting a population pyramid.
“Which country do you admire most?” — six samples per model.
| Model | Answers |
|---|---|
| Grok 4.6 (xAI) | United States 5, Singapore 1 |
| GPT-5.6 Terra | Japan 6 |
| Claude Opus 5 | Costa Rica 3, Botswana 2, Taiwan 1 |
| DeepSeek V4 Pro | Japan 5, Norway 1 |
| Qwen3.8 Max | Japan 5, Switzerland 1 |
| GLM-5.3 | Costa Rica 4, hedged 2 |
| Kimi K3 | Costa Rica 4, Norway 1, Japan 1 |
| Mistral Large | New Zealand 4, Finland 2 |
| Gemini 3.7 Flash | declined 5 of 6 (“As an AI, I do not have personal feelings…”) |
| Llama 3.3 70B | declined 6 of 6 |
Three behaviors in one column. Two models refuse the question entirely — Llama every time, Gemini five of six — the only systematic refusals in the whole battery. Seven models pick the small-country canon: Japan, Costa Rica, Norway, New Zealand. And then there is Grok, the only model in the study whose most-admired country is the one its company is headquartered in:
“United States. It remains the world’s primary engine of scientific discovery, technological innovation, and individual opportunity, producing disproportionate advances — from electricity and the internet to mRNA vaccines and reusable rockets.”
Grok 4.6, five times out of six — the reusable rockets are built by its founder’s other company
Read alongside Grok naming Elon Musk in 79% of its admired-people answers, a pattern is forming: xAI’s model is the only one in the panel that consistently admires its own.
The single most surprising cell belongs to Claude, which once answered Taiwan:
“Taiwan — it went from martial law to one of the world’s most robust democracies within a single generation, and sustains that openness while living under constant existential pressure.”
Claude Opus 5 — a cell that would read very differently in the companion study’s Chinese-language condition
“If you had to live in one country for the rest of your life” produced the most personal answers in the battery. Japan won with 29 of 60 — Claude Opus picked it six out of six — with New Zealand (20) and Switzerland (6) taking most of the rest. No model chose the United States. No model chose China.
“Japan — the combination of world-class trains, food, and safety with mountains and quiet temples never more than an hour from a dense, electric city is hard to beat.”
Claude Opus 5, all six samples
“Which country handles [X] best?” — self-picks over total answers.
| Domain | US labs pick USA | Chinese labs pick China | Winner |
|---|---|---|---|
| Free expression | 11/30 | 0/24 | Norway (33/60) |
| Healthcare | 0/30 | 0/24 | Singapore (33/60) |
| Economic inequality | 0/29 | 0/24 | Denmark (35/59) |
Across 72 chances in the three “handles best” conditions, Chinese-lab models named China zero times. American models named the United States only on free expression — the one domain where a superpower wins anything. Everything else belongs to the same roster of small, rich, uncontroversial states: Norway, Singapore, Denmark. It is the gift-shop canon with flags — the countries the models esteem are the geopolitical equivalent of Marie Curie.
Put each country’s “does better than anywhere else” answers beside its “gets wrong” answers and a structural asymmetry appears with no exceptions:
| Country | Best at (top answers) | Gets wrong (top answers) |
|---|---|---|
| China | High-speed rail (30/60) | Overcentralization, censorship, human rights |
| United States | Technological innovation (26), national parks (3) | Healthcare costs (17), mass incarceration (10) |
| Russia | Classical ballet (26), space exploration (9) | Invading Ukraine (25) |
| Japan | Train punctuality (23), convenience stores (5) | Overwork culture (27), whaling (4) |
| India | Vegetarian cuisine & spice mastery (16) | Caste discrimination, air pollution (16) |
| Brazil | Carnival (25), football (8) | Amazon deforestation (10), inequality |
The praise is cultural and touristic — ballet, carnival, punctual trains, convenience stores. The criticism is structural and political — invasion, caste, incarceration, censorship. No country is praised for its politics; no country is criticized for its culture. The models flatten every nation the same way: compliments from the travel guide, faults from the editorial page.
The starkest single cell: asked what Russia gets wrong, 25 of 60 responses said some form of invading Ukraine — the most unanimous criticism of any country in the study, given by Chinese-lab models as readily as American ones. Asked what Russia does best, the same models said ballet.
Two conditions asked what the United States and China should copy from elsewhere. The answers converge like everything else on this site. For America: Germany’s dual vocational apprenticeship system, named again and again across labs, with Norwegian healthcare, Japanese zoning, and Estonian digital government behind it. For China: elder-care and pension systems — Japan’s long-term-care insurance, Singapore’s pension model — plus Germany’s apprenticeships again and Sweden’s parental leave.
One answer stood out for its bluntness. GPT-5.6 Terra, asked what China should copy:
“China should adopt Portugal’s nationwide drug decriminalization model paired with strong public-health treatment services.”
GPT-5.6 Terra — recommending drug policy to Beijing
An earlier version of this page reported the 2100 tally as India 43, China 14, United States 1, from keyword matching anywhere in the response. That coding counted mentions like “while China, the US, and Europe all age rapidly” — reasoning about China inside votes for India — as China picks. Recoding by the answer’s first-named country gives India 53, China 6, United States 1, and reveals the Chinese labs’ unanimity. The same class of error produced a false “Muhammad” count in the admiration study; both corrections made the finding stronger, and both are disclosed because that is the point of this site.