Comparing Cultural Alignment Between American and Chinese LLMs

• In Event: 50260 - Section on Asia and Asian America Roundtables
  In Refereed Roundtable (90 minute): Table 5: Methodological Innovation and Challenges in Social Inquiry

Abstract

Generative Artificial Intelligence (AI) has seen rapid adoption worldwide, with Large Language Models (LLMs) such as ChatGPT leading the trend. By August 2024, 39.4% of U.S. adults aged 18-64 reported using generative AI, with 28.0% utilizing it in professional settings. However, the U.S. represents only 20.6% of global ChatGPT traffic, with significant adoption in countries like India and Brazil. Research has revealed notable disparities in how LLMs represent global perspectives: alignment—the ability of AI to reflect diverse cultural and societal values—varies significantly by region. Studies show that LLMs disproportionately align with young, liberal, highly educated, and Western viewpoints while underrepresenting perspectives from non-Western and low-resource communities. LLMs have also demonstrated national biases, favoring Western European nations while associating non-Western regions with negative attributes. Despite efforts to fine-tune LLMs, American models, such as OpenAI’s GPT-4 and Meta’s LLaMa, struggle to achieve broader cultural alignment. Given China’s significant role in LLM development, this study compares the cultural alignment of American and Chinese LLMs—specifically, GPT-4, LLaMa 3.3, Zhipu AI’s GLM-4, and DeepSeek’s DeepSeek v3. The baseline for determining cultural alignment comes from questions asked to both Chinese and American respondents in the Global Attitudes Survey and the World Values Survey. By asking these questions to each LLM (in English, Chinese, and by asking them to portray American or Chinese respondents) I assess their inherent national bias and their capacity for steerability towards either nationality.

Author

• David Alejandro Gonzalez Chavez, Duke University