[Paper Review] Cultural Bias and Cultural Alignment of Large Language Models

Yan Tao, Olga Viberg+2 more|arXiv (Cornell University)|Nov 23, 2023

Computational and Text Analysis MethodsSocial Sciences12 citations

TL;DR

This study evaluates cultural bias in five widely used LLMs by comparing their outputs to nationally representative survey data, and shows that cultural prompting can improve alignment for a majority of countries in recent models.

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.

Motivation & Objective

• Motivate understanding of how culture shapes LLM reasoning, communication, and outputs.
• Quantify cultural bias in five widely used LLMs against nationally representative survey data.
• Evaluate a control strategy, cultural prompting, to improve cultural alignment across countries/territories.
• Provide actionable guidance on reducing cultural bias in generative AI outputs.

Proposed method

• Disaggregate evaluation across five LLMs: GPT-4o, GPT-4-turbo, GPT-4, GPT-3.5-turbo, GPT-3.
• Compare model responses to nationally representative survey data to assess cultural values reflected in outputs.
• Apply cultural prompting as a control strategy to influence model outputs toward target cultural norms.
• Measure cultural alignment of outputs by country/territory before and after prompting.
• Report improvement percentages (71-81%) for recent models across countries/territories.

Experimental results

Research questions

• RQ1 To what extent do popular LLMs encode and reflect cultural values from English-speaking and Protestant European contexts?
• RQ2 Can cultural prompting improve the cultural alignment of LLM outputs across diverse countries and territories?
• RQ3 How does cultural alignment vary across recent versus earlier GPT-ffamily models?
• RQ4 What proportion of countries/territories show improved alignment under cultural prompting for modern LLMs?

Key findings

• All evaluated models exhibit cultural values resembling English-speaking and Protestant European countries.
• Cultural prompting improves cultural alignment for 71-81% of countries/territories in recent models (GPT-4, 4-turbo, 4o).
• Cultural prompting can be an effective control strategy to reduce cultural bias in LLM outputs.
• The study demonstrates the feasibility of disaggregated, country-level evaluation of cultural bias in LLMs.