Not a mirror, a caricature: How LLMs reproduce cultural identity?
摘要
We examine whether large language models reproduce human-like cultural identity or instead produce deterministic cultural prototypes. Using the 30-item Self-Construal Scale (SCS), we asked two GPT models to answer as if they were American, Polish, Japanese, or with no specified identity, in English or Polish, under weak vs. strong identity cues (2 × 4 × 2; 30 runs per cell). Identity prompts yielded almost perfectly separable cultural profiles with near-zero within culture variance. In neutral conditions, outputs systematically skewed toward a U.S. profile in 75% of tests, consistent with an anglocentric default. A simple nearest-neighbor classifier achieved 99.8% leave-one-out accuracy in predicting the assigned cultural identity from item-level responses, confirming near-perfect profile separability. These findings indicate that, for cultural identity reproduction, current models behave as caricaturists rather than mirrors of human cultural variation.