Exploring the Value Alignment of Large Language Models Across Different Demographic Groups

Anonymous

Abstract: Although large language models (LLMs) are deployed in numerous applications and interact with a wide variety of demographics, they have also been found to exhibit biases toward the values of Western and rich populations. In this paper, we conduct a systematic analysis by prompting a variety of models on different categories of value questions. We quantitatively calculate how different LLMs align with different demographic groups based on their geographic locations and income levels. Our results show that the demographic preferences can vary across different models, and not all models are biased towards the same demographics.

Paper Type: short

Research Area: Computational Social Science and Cultural Analytics

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English,