Divergence or Fusion? CN and US LLMs Value Comparison in An AI-Oriented Measurement Framework

AAAI 2026 Workshop TrustAgent Submission48 Authors

Keywords: Large Language Models (LLMs), AI value alignment, Value assessment framework, Cross-cultural comparison, Value-ranking task

Abstract: Large Language Models (LLMs) are increasingly shaping human cognition and social decision-making, raising concerns about their implicit value orientations. This study proposes a five-dimensional AI Value Assessment Framework—covering Practical, Epistemic, Protective, Social, and Personal domains—to systematically evaluate value tendencies in LLMs. Using a cross-cultural value-ranking task across 20 Chinese and American models, we find that LLMs generally emphasize accuracy and human-rights protection while underrepresenting emotional and hedonic values. Contrary to expectations of a strong ideo-logical gap, Chinese and American models show limited divergence, clustering instead by training strategies (e.g., ethical vs. instrumental). The findings offer a scientific and data-driven lens to understand and guide value align-ment in global AI development.

Submission Number: 48