Abstract
Large Language Models (LLMs) are increasingly deployed across diverse linguistic and cultural contexts, yet it remains unclear whether cultural values embedded in language influence how these systems behave in social interactions. This paper examines how language and culture shape LLM decision-making in the context of economic negotiations. We prompt agents instantiated from multiple versions of LLMs to play a modified ultimatum game. In this game, one agent proposes how to divide resources, while the other specifies the minimum acceptable offer; the responder’s veto power is systematically varied across experimental conditions. Prompts are presented in 17 natural languages representing diverse cultural backgrounds. We report three main findings. First, agents’ offers vary significantly across languages, suggesting that the linguistic context of a prompt influences fairness-related behavior. Second, consistent with patterns observed in human participants, agents make higher offers and set higher acceptance thresholds when responders possess greater veto power. Third, both offers and acceptance thresholds are significantly associated with the Power Distance Index (PDI) of the cultures linked to each language, connecting LLM behavior to real-world cultural attitudes toward hierarchy and inequality. The positive relationship between Power Distance and offer size is consistently observed across all evaluated models. For acceptance thresholds, the overall association remains significant, although the strength of the effect varies across model families. Together, these findings provide evidence that culturally embedded biases are reflected in LLM behavior and may affect fairness in multilingual and cross-cultural applications. We discuss implications for the development and deployment of AI systems that operate equitably across linguistic and cultural settings.

This publication has 8 references indexed in Scilit: