Which Humans? LLMs mainly mirror WEIRD minds (Europeans?) (www.hks.harvard.edu)

A new working paper shows that large language models (LLMs) do not reflect a generic “human” mind but instead mirror the psychological profiles of WEIRD populations (Western, Educated, Industrialized, Rich, Democratic). By comparing LLM responses on standard cognitive and psychological measures against large-scale cross-cultural human datasets, the authors find LLM behavior is an outlier relative to global human variation and most closely matches people from WEIRD societies. Model-human similarity falls off sharply as populations diverge from WEIRD norms (correlation r = −.70), highlighting that many claims of “human-level” performance implicitly mean “WEIRD-level” performance. The finding matters for AI/ML research, evaluation, and deployment: training on predominantly WEIRD textual data produces models that generalize unevenly across cultures, raising scientific limits on cross-cultural validity and ethical risks around fairness and misinterpretation. Technically, the paper recommends concrete mitigations—broader, more diverse training corpora; cross-cultural benchmarking and evaluation suites; culturally aware fine-tuning; and interdisciplinary collaboration with social scientists—to reduce cultural bias in future generative models and ensure assessments specify which human populations they actually approximate.