【公开学术报告】Cultural Tendencies in Generative Artificial Intelligence

Cultural Tendencies in Generative Artificial Intelligence

Speaker: Lesley Luyang Song (Tsinghua University)

Date & Time: Wed. 2^nd, April 2025, from 10:00 AM to 11:30 AM (Beijing Time)

Location: Room 2101, Tongji Building A

ABSTRACT

This paper shows that responses of generative AI models—trained on textual data that are inherently cultural—exhibit cultural tendencies when used in different human languages. We examine cultural tendencies in OpenAI’s GPT, Baidu’s ERNIE, and DeepSeek, three widely used generative AI models. First, we analyze GPT’s responses to a large set of measures in both Chinese and English—without any cultural prompts (e.g., “for an average Chinese person…”). Our measures focus on two foundational constructs in cultural psychology that underlie everyday life: social orientation and cognitive style. When used in Chinese (vs. English), GPT’s responses exhibit a more interdependent (vs. independent) social orientation and a more holistic (vs. analytic) cognitive style; the effect sizes (medium to large) are meaningful. These cultural tendencies are robust across different measures—including Likert scales, vignette tasks, imagery tasks (e.g., selecting one pair of circles that best represents a social relationship), and text analysis—as well as different model parameters (e.g., temperature) and prompt formats. Second, we replicate these cultural tendencies in ERNIE, the most popular generative AI model in China. Third, we demonstrate the real-world impact of these cultural tendencies in GPT and DeepSeek. For example, when used in Chinese (vs. English), GPT is more likely to recommend an advertisement slogan with an interdependent (vs. independent) social orientation. Fourth, exploratory analyses suggest that cultural prompts can adjust these cultural tendencies: Prompting generative AI to assume the role of a Chinese person makes its responses in English more interdependent and holistic. This research has broad implications for AI developers, individual users, organizational users, and non-users.

Keywords: artificial intelligence; large language models; culture; psychology; social science