Multilingual Large Language Models and Cultural Diversity: An Empirical Study on Civic and Moral Judgments

Core Overview

This article was published by Eugenio Vicario on GitHub (original link: https://github.com/EugenioVicario/multilingual_llm, published on June 15, 2026). It focuses on the performance of multilingual large language models in civic and moral judgment scenarios involving cultural diversity. Through experiments, it reveals that models have a Western-centric tendency and cultural understanding biases, providing empirical evidence for AI fairness and cross-cultural applications.

Research Value

Addressing the issue of cultural fairness in the global application of LLMs, it emphasizes that technology needs to balance capability and cultural sensitivity to avoid sacrificing diversity.

With the global popularization of LLMs, a core question emerges: Can models fairly understand inputs from different cultural backgrounds?

Existing mainstream LLMs are mainly based on English corpora, which easily lead to systematic biases in non-Western cultural contexts. Cultural diversity is not only at the language level but also reflected in values, moral judgments, and social norms (such as the balance between individual rights and collective obligations). If models cannot capture these differences, unfair or harmful outputs may occur in global applications.

The study evaluates the cultural sensitivity of multilingual LLMs through systematic experiments:

1. Dataset Construction: Create cross-lingual and cross-cultural test datasets to quantify the alignment between model outputs and human cultural values;
2. Comparative Analysis: Present prompts of civic obligations and moral dilemmas to the models, collect results, and compare them with human respondents from different cultural backgrounds to reveal the strengths and weaknesses of the models.

1. Western-Centric Tendency: When handling cultural value issues, models tend to reflect Western liberal values and have insufficient understanding of cultural perspectives on collective harmony and social order;
2. Superficial Multilingualism: Even if a model uses a language fluently, it may not understand the cultural connotations behind it, leading to the phenomenon of "superficial multilingualism, deep monoculture", which restricts global applications.

1. Enterprise Applications: Provide a basis for enterprises deploying AI products globally to address cultural biases;
2. Academic Value: Open-source datasets and code provide standardized evaluation tools for subsequent research;
3. Macro Implications: Technological globalization should not sacrifice cultural diversity, and responsible AI needs to balance technical capability and cultural sensitivity.

1. Framework Expansion: Expand the evaluation framework to more languages and value judgment tasks;
2. Developer Insights: Multilingual capability should be regarded as cross-cultural understanding ability, not just language translation, and intelligent systems serving global users need to be developed.