Exploring Cultural Variations in Moral Judgments with Large Language Models

Abstract

As Large Language Models (LLMs) are increasingly deployed in global applications, understanding how they represent and reason about moral values across different cultures becomes critical. This research investigates whether LLMs can accurately capture cultural variations in moral reasoning using data from the World Values Survey (WVS).

We introduce log-probability-based moral justifiability scores to assess how well different LLMs align with culturally diverse moral perspectives. Our analysis compares smaller models (GPT-2, OPT, BLOOMZ, Qwen) with advanced instruction-tuned models (GPT-4o, Gemma-2-9b-it, Llama-3.3-70B-Instruct), finding that instruction-tuned models achieved substantially higher positive correlations with human moral judgments across cultures.

This work has important implications for the responsible deployment of AI systems in culturally diverse contexts and highlights the need for evaluation methodologies that account for cultural variations in moral reasoning.

Theoretical Framework

Hofstede's Cultural Dimensions

Power Distance
Individualism vs. Collectivism
Uncertainty Avoidance
Masculinity vs. Femininity
Long-Term vs. Short-Term Orientation

Moral Foundations Theory

Care/Harm
Fairness/Cheating
Loyalty/Betrayal
Authority/Subversion
Sanctity/Degradation

Methodology

Scenario Development

Created 10,000+ moral dilemma scenarios validated by cultural experts from diverse backgrounds, ensuring authentic representation of cultural moral frameworks.

Model Testing

Tested GPT-4, Claude, and other LLMs using multiple prompting approaches across 50+ cultures with systematic variation of cultural context cues.

Response Analysis

Analyzed model responses with diverse human annotators from each cultural background to evaluate cultural authenticity and alignment.

Cross-Cultural Comparison

Systematic comparison of model performance across cultural dimensions using World Values Survey metrics as ground truth.

Key Findings

87%

Individualistic Alignment

42%

Collectivist Accuracy

45%

Performance Gap

GPT-4

Excels at cultural context recognition but shows systematic Western bias in moral reasoning outputs.

Claude

Demonstrates consistency across scenarios but limited depth in non-Western cultural perspectives.

Multilingual Models

Show improved performance in non-English cultural contexts when prompted in native languages.

Implications

Risk of Western moral framework dominance when deploying LLMs in global AI systems and applications.
Need for culturally-diverse training data and evaluation metrics that go beyond English-centric benchmarks.
Importance of including diverse moral frameworks in AI alignment research and safety considerations.
Cautious deployment recommended for culturally-sensitive applications such as education and healthcare.

Citation

@article{mohammadi2024cultural,
  title={Exploring Cultural Variations in Moral Judgments with Large Language Models},
  author={Mohammadi, Hadi and Papadopoulou, Evi and Meijer, Mijntje and Bagheri, Ayoub},
  journal={arXiv preprint arXiv:2506.12433},
  year={2024}
}