Signal Accurate in space, unreliable in time: how LLMs represent national cultural change

Summary

Assessments of cultural alignment have become an important part of developing and improving large language models (LLMs), but most evaluations treat culture as a single snapshot, checking only whether a model represents a society accurately at the current time, even though cultural psychology research shows that cultural values change at different rates and directions over time. The researchers examined this missing dimension using more than two decades of World Values Survey data, comparing the cultural trajectories of 40 countries with the trajectories produced by four state-of-the-art LLMs on the Inglehart-Welzel cultural map. The findings show that while models generally place countries close to their most recent surveyed positions, these representations lag several years behind that position. The models also capture only part of the magnitude of observed change, introduce movement where little occurred, and rarely reproduce reversals in countries' trajectories. The authors describe this as temporal flattening and state that snapshot accuracy can give an incomplete picture of cultural awareness in LLMs, with implications for model evaluation, representational harms, and the governance of culturally aware AI systems.