Prompt and Prejudice: Implicit and Explicit Cultural Responsiveness in Large Language Models

Abstract

This thesis investigates the cultural responsiveness of large language models using the Inglehart–Welzel cultural framework. Evaluating eight models across eight benchmark countries, it asks not only where models sit in cultural value space by default, but whether explicit prompting moves them toward the relevant human benchmark, whether that movement occurs in the correct direction, and whether weaker identity cues such as names can induce similar shifts. The results show that models do not begin from culturally neutral positions and that their baseline outputs occupy a much narrower region of cultural value space than the human benchmark countries. Explicit country prompting often reduces benchmark distance, but its effectiveness is highly uneven across both countries and models. Moreover, distance reduction and directional accuracy do not always move together, showing that cultural responsiveness cannot be evaluated by movement alone. A further name-inference experiment finds that implicit cues influence outputs, but are unrelated and less reliable than explicit prompting. Together, these findings argue that cultural alignment in large language models is not a single trait, but the interaction of multiple capacities: baseline positioning, responsiveness to explicit instruction, directional control, and sensitivity to social cues.