Multilingual Large Language Models and cultural diversity: evidence from civic and moral judgments

Multilingual large language models (LLMs) are increasingly deployed across linguistic and cultural contexts, raising the question of whether multilingual interaction preserves cultural diversity in moral judgments. We compare civic and moral evaluations generated by a multilingual LLM across multiple languages with population-level data from the World Values Survey and the European Values Study. Although the model exhibits meaningful linguistic variability, this does not translate into the preservation of cross-national moral diversity. Alignment with population-level values is highest for WEIRD countries and weaker elsewhere. At the cross-national level, the LLM reshapes the structure of moral distances: differences between WEIRD and non-WEIRD countries are selectively compressed in the model-generated space, while distances within these groups remain largely unchanged. These dynamics are strongly domain-dependent. Anti-civic norms display a pronounced norm-enforcing bias with minimal cross-national variation; personal and bioethical judgments cluster around values typical of WEIRD countries; public and social order norms exhibit systematic directional shifts, particularly outside WEIRD cultural zones; and attitudes toward political violence display increased dispersion rather than convergence. Together, these findings suggest that multilingual LLMs act as normative infrastructures that reshape moral representations in uneven and domain-specific ways, highlighting the limits of multilingual fluency as a guarantee of cultural alignment.