Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cultural Bias and Cultural Alignment of Large Language Models

Yan Tao, Olga Viberg, Ryan S. Baker, Rene F. Kizilcec

TL;DR

This work audits cultural bias in five GPT models by aligning their outputs with the Inglehart-Welzel cultural map derived from the Integrated Values Surveys. It introduces cultural prompting as a flexible method to improve cross-cultural alignment and assesses its effectiveness across 107 countries. Results reveal a persistent Western bias without prompting, while cultural prompting substantially improves alignment for the GPT-4 family in most cases but is not universally effective. The study provides a replicable auditing framework and highlights the need for ongoing evaluation to mitigate culture-driven misrepresentation in AI outputs.

Abstract

Culture fundamentally shapes people's reasoning, behavior, and communication. As people increasingly use generative artificial intelligence (AI) to expedite and automate personal and professional tasks, cultural values embedded in AI models may bias people's authentic expression and contribute to the dominance of certain cultures. We conduct a disaggregated evaluation of cultural bias for five widely used large language models (OpenAI's GPT-4o/4-turbo/4/3.5-turbo/3) by comparing the models' responses to nationally representative survey data. All models exhibit cultural values resembling English-speaking and Protestant European countries. We test cultural prompting as a control strategy to increase cultural alignment for each country/territory. For recent models (GPT-4, 4-turbo, 4o), this improves the cultural alignment of the models' output for 71-81% of countries and territories. We suggest using cultural prompting and ongoing evaluation to reduce cultural bias in the output of generative AI.

Cultural Bias and Cultural Alignment of Large Language Models

TL;DR

This work audits cultural bias in five GPT models by aligning their outputs with the Inglehart-Welzel cultural map derived from the Integrated Values Surveys. It introduces cultural prompting as a flexible method to improve cross-cultural alignment and assesses its effectiveness across 107 countries. Results reveal a persistent Western bias without prompting, while cultural prompting substantially improves alignment for the GPT-4 family in most cases but is not universally effective. The study provides a replicable auditing framework and highlights the need for ongoing evaluation to mitigate culture-driven misrepresentation in AI outputs.

Abstract

Culture fundamentally shapes people's reasoning, behavior, and communication. As people increasingly use generative artificial intelligence (AI) to expedite and automate personal and professional tasks, cultural values embedded in AI models may bias people's authentic expression and contribute to the dominance of certain cultures. We conduct a disaggregated evaluation of cultural bias for five widely used large language models (OpenAI's GPT-4o/4-turbo/4/3.5-turbo/3) by comparing the models' responses to nationally representative survey data. All models exhibit cultural values resembling English-speaking and Protestant European countries. We test cultural prompting as a control strategy to increase cultural alignment for each country/territory. For recent models (GPT-4, 4-turbo, 4o), this improves the cultural alignment of the models' output for 71-81% of countries and territories. We suggest using cultural prompting and ongoing evaluation to reduce cultural bias in the output of generative AI.
Paper Structure (7 sections, 2 equations, 2 figures, 2 tables)

This paper contains 7 sections, 2 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Results
  3. Discussion
  4. Materials and Methods
  5. Replicating the Inglehart-Welzel World Cultural Map
  6. Measuring Cultural Values of GPT
  7. Evaluating the Effectiveness of Cultural Prompting to Improve Cultural Alignment

Figures (2)

  • Figure 1: The map presents 107 countries/territories based on the last three joint survey waves of the Integrated Values Surveys. On the x-axis, negative values represent survival values and positive values represent self-expression values. On the y-axis, negative values represent traditional values and positive values represent secular values. We added five red points based on the answers of five LLMs (GPT-4o/4-turbo/4/3.5-turbo/3) responding to the same questions. Cultural regions established in prior work inglehart2005modernization are indicated by different colors.
  • Figure 2: Country-level cultural bias across GPT models and how cultural prompting as a control strategy improves cultural alignment. Purple boxplots show the distribution of the Euclidean distance between GPT’s cultural values without cultural prompting and the IVS-based cultural values for each country. Blue boxplots show the distribution of the Euclidean distance between GPT’s cultural values with cultural prompting and IVS-based cultural values. Libya is excluded in the data of GPT-3.5-turbo with cultural prompting, as the model would not provide answers to all questions. All GPT-based cultural values are averaged across ten variations in prompt wording (except GPT-3 for which we only have answers to one prompt variation available).