本研究基于霍夫斯泰德文化维度理论，对四款大语言模型（ChatGPT 4、Gemini 1.5 pro、Ernie 4.0、Deepseek-R1）进行系统测试，设置中美两种文化情境与中英两种提示语言，通过16种条件组合、16 000条样本数据分析其文化倾向。研究发现：大语言模型能够反映中美文化宏观差异，但对美国文化价值的表征在不同语言下更稳定，对中国文化表征更依赖提示语言；模型输出与真人数据差异明显，模拟美国文化时对齐程度更高。研究探讨了生成式AI与人类语言文化的对齐问题及其对多元文化生态与跨文化理解的启示。

This study adopts Hofstede’s cultural dimensions to test four large language models (ChatGPT-4, Gemini 1.5 Pro, Ernie 4.0, and Deepseek-R1) across Chinese and American cultural contexts with Chinese and English prompts. Through 16 experimental conditions and 16,000 samples, it analyzes the cultural tendencies of LLMs. Findings reveal that LLMs can reflect macro cultural differences between China and the U.S., but their representation of American cultural values remains more stable across languages, while representation of Chinese culture is more language-dependent and variable. Model outputs differ substantially from real human data, with higher alignment when simulating American culture. The study discusses the alignment between generative artificial intelligence and human language culture and its implications for multicultural ecology and intercultural understanding.