Journal Article Published 2026

Exploring Cultural Variations in Moral Judgments with Large Language Models

Hadi Mohammadi, Robert A. Bagheri

Utrecht University, The Netherlands

Computational Linguistics in the Netherlands Journal, Vol. 15 (2026), pp. 59–77

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

Survey Items as Ground Truth

Moral-justifiability items drawn from the World Values Survey and the Pew Research Center Global Attitudes Survey, so the reference is existing cross-national survey data rather than newly written scenarios.

Model Scoring

Eight decoder-only models scored the same items: GPT-2, OPT, BLOOMZ and Qwen among the smaller models, and GPT-4o, GPT-4o-mini, Gemma-2-9b-it and Llama-3.3-70B-Instruct among the instruction-tuned ones.

Log-Probability Elicitation

Each model's judgement is read from the log-probabilities it assigns to the survey response options, so no human annotation step is involved.

Cross-Cultural Comparison

Model scores are correlated against each country's survey responses, which is what exposes the gap between W.E.I.R.D. and non-W.E.I.R.D. nations.

Key Findings

8
Models evaluated
2
Cross-national value surveys
W.E.I.R.D.
Where alignment is strongest
Smaller models
GPT-2, OPT, BLOOMZ and Qwen often produce near-zero or negative correlations with the survey data.
Instruction-tuned models
GPT-4o, GPT-4o-mini, Gemma-2-9b-it and Llama-3.3-70B-Instruct reach substantially higher positive correlations.
Regional coverage
Alignment is strongest for W.E.I.R.D. nations and weaker elsewhere, across every model tested.

Implications

Citation

@article{mohammadi2024cultural,
  title={Exploring Cultural Variations in Moral Judgments with Large Language Models},
  author={Mohammadi, Hadi and Bagheri, Robert A.},
  journal={Computational Linguistics in the Netherlands Journal},
  volume={15},
  pages={59--77},
  year={2026}
}