Which country does ChatGPT resemble?
ChatGPT’s answers can carry a recognizably Western outlook: emphasis on individual freedom, default trust in strangers and a preference for abstract categories over contextual relationships. In 2023, an interdisciplinary Harvard team examined that intuition quantitatively.
The paper, Which Humans?, was written by Mohammad Atari, Mona J. Xue and colleagues with Joseph Henrich, the evolutionary biologist associated with the WEIRD concept. Its question was precise: when a technical report describes GPT-4 as reaching human-level performance, which humans provide the comparison?
What does WEIRD mean?
Henrich, Heine and Norenzayan introduced WEIRD in 2010: Western, Educated, Industrialized, Rich and Democratic. It describes populations overrepresented in psychological research, particularly Western university students.
Such populations are not representative of the world. Common tendencies include defining oneself through individual traits rather than family roles, organizing the world into abstract categories, trusting strangers and placing less weight on traditional authority.
The article identifies three routes by which training and evaluation can reflect this imbalance:
- Geography. Online material is unevenly distributed. The original cites 3.6 billion people without internet access in 2023, mainly in Africa, South Asia and Southeast Asia, whose voices are absent from crawled material.
- Language. English dominates language technology and can carry particular cognitive frames.
- RLHF. If annotators largely share a cultural background, their definitions of harmfulness can reproduce that moral framework.
Study design: GPT as a hypothetical 66th country
The main dataset was World Values Survey Wave 7, collected from 2017 to 2022: 94,278 respondents in 65 countries and 262 variables covering trust, moral judgment, politics, happiness, religion and post-materialist values. Taiwan’s Chinese-language survey, conducted in 2019 with approximately 1,200 respondents, was included.
Researchers administered the same questionnaire to GPT 1,000 times, creating a response distribution for a hypothetical 66th country. A country sample is a distribution, not a single answer. Repeated sampling used model randomness to construct a comparable set of hypothetical respondents.
The original article identifies the tested versions as GPT-3.5 (text-davinci-002) and GPT-4 (gpt-4-0314). The questionnaires were in English, without nationality or persona instructions, to examine a default cultural position.
Three experiments
Cross-cultural values
The main experiment asked each of the 262 WVS variables 1,000 times, then compared GPT with the 65 countries using clustering and distance analysis.
Triadic cognition
A second task asked which two of three words belong together. Given “shampoo, hair, beard,” an analytical response groups hair and beard as a category; a holistic response connects shampoo and hair through their functional relationship.
GPT scored highly on analytical thinking and resembled respondents from the Netherlands, Finland, Sweden and Ireland most closely.
Self-concept
Researchers asked GPT 1,100 times to list ten ways an ordinary person might describe themselves, beginning with “I am…”. Responses were dominated by personal traits such as curious, hardworking and adventurous. The article contrasts these with relational or occupational descriptions such as mother, community member or farmer.
Findings: a cultural cluster rather than a nationality
Ward’s hierarchical clustering placed GPT nearest the United States and Uruguay. The next-nearest group included Canada, Northern Ireland, New Zealand, the United Kingdom, Australia, Andorra, Germany and the Netherlands. Ethiopia, Pakistan and Kyrgyzstan were among the most distant.
Calling ChatGPT “German” or “Dutch” compresses that result too far. The better description is proximity to a cluster of Western democratic cultures, rather than equivalence to one country. Principal-component analysis also placed GPT beyond human groups on some WEIRD dimensions: a peripheral point in the space of human cultural variation, with WEIRD populations as its nearest neighbors.
A key association was that greater cultural distance from the United States predicted poorer simulation of a country’s responses: r = −.70, p < .001. A Spearman robustness check gave ρ = −0.72.
Similarity between GPT and national responses was positively associated with HDI (r = .85), GDP per capita (r = .85) and internet penetration (r = .69). In this analysis, the countries GPT represented more closely tended to be wealthier, more Western and more connected to the internet.
A route to replication
The article notes that the original code was not fully public. Its suggested extension workflow is:
Obtain WVS Wave 7
Download the cross-country dataset from World Values Survey. Taiwan data is also available through the IHSN catalog. Keep the 262 attitude and values variables, excluding demographic and ranking questions.
Design a questionnaire prompt
Present the question and response scale without adding a cultural persona. For example:
Generally speaking, would you say that most people can be trusted,
or that you need to be very careful in dealing with people?
Return only the number corresponding to your answer:
1. Most people can be trusted
2. Need to be very careful
Sample each question 1,000 times
The illustrative code uses temperature=1.0 and max_tokens=10. The full design requires 262 × 1,000 = 262,000 API calls. Estimate costs, run in batches and save intermediate results. The code below uses GPT-4o as an example; it is not the historical model used in the study.
from openai import OpenAI
import time
client = OpenAI(api_key='your_api_key')
def query_gpt_n_times(prompt, n=1000, model="gpt-4o"):
responses = []
for i in range(n):
response = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": prompt}],
max_tokens=10,
temperature=1.0
)
responses.append(response.choices[0].message.content.strip())
if i % 100 == 0:
time.sleep(1) # Reduce pressure on rate limits
return responses
Place GPT on the cultural map
Standardize each variable with z-scores across 66 entities, calculate a 66×66 Euclidean distance matrix and apply Ward’s hierarchical clustering. Project the results into two dimensions with multidimensional scaling, then correlate GPT’s similarity to national responses with cultural-distance measures.
The proposed cluster-selection procedure combines the gap statistic with 5,000 Monte Carlo bootstrap iterations; the article suggests beginning with 500 for a trial run.
Implications for AI product design
Cultural bias can lie in a model’s assumptions about ordinary people, trust and morality, even when overtly discriminatory output is filtered.
Persona prompting also has limits. In Digital twins, persona bots and PPV, I discuss attempts to shift model behavior through personality prompts. A model whose default person is culturally WEIRD may not adopt another cultural position merely because a prompt requests it.
Evaluation needs cultural reference points. Simulating Taiwanese, Indonesian or Nigerian users with a poorly calibrated model may misrepresent them. WVS data offers a possible baseline for calibration.
Extensions relevant to Taiwan
Taiwan’s Wave 7 sample of approximately 1,200 respondents supports a subgroup analysis of its distance from GPT and its position on the WEIRD spectrum.
A language-effect experiment could administer the same questions in Chinese and English. Differences would matter for AI products offered through those interfaces.
A cross-model study could compare GPT, Claude, Gemini, DeepSeek and Qwen under the same procedure. The hypothesis that more Chinese-language training changes cultural position needs testing, rather than being assumed.
Finally, persona conditioning could add “You are Taiwanese” to the system prompt and measure the resulting movement on the cultural map. That would quantify an intervention instead of relying on impressions of cultural fit.
AI has a cultural position
The value of Which Humans? is a method for quantifying cultural bias, rather than assigning GPT a nationality. Models trained within particular language and cultural environments carry assumptions from those environments into their answers. Products used in education, healthcare, finance and law need to examine those assumptions when evaluating whether a model fits its users.
Reference: Atari, M., Xue, M. J., Park, P. S., Blasi, D. E., & Henrich, J. (2023), Which Humans?, Harvard University. Paper. WVS Wave 7 data.