GPT's Worldview Is Culturally Skewed Toward the West [pdf] (scholar.harvard.edu)

A new study from Harvard researchers shows that large language models (LLMs) — when probed with psychological questionnaires and cognitive tasks — systematically reflect a Western, Educated, Industrialized, Rich, and Democratic (WEIRD) worldview rather than a global human average. Using the seventh wave of the World Values Survey (94,278 respondents across 65 countries) and multiple standard cognitive tests, the authors sampled LLM outputs (1000 responses per question via the OpenAI API) and found the models’ behavioral profile to be a clear outlier versus large-scale cross-cultural data. LLM similarity to human populations declines sharply as those populations become less WEIRD (reported correlation r = -0.70), indicating models align closely with Western patterns of individualism, trust, and moral attitudes and diverge from many non-WEIRD societies. This matters for researchers and practitioners who treat LLMs as “human-like” benchmarks: claims about human-level performance or using models as proxies for human subjects risk being culturally narrow and misleading. The paper highlights technical causes (Internet and English-dominant training corpora, uneven access to data), methodological pitfalls (debiasing and moderation reflecting WEIRD norms), and ethical implications (misrepresenting global perspectives, reinforcing cultural bias). The authors call for cross-cultural evaluation, broader language and data sourcing, and culturally diverse human feedback as mitigation paths for future generative models.