Cultural tendencies in generative AI

Cultural Bias in AI Models

Generative artificial intelligence (AI) models, trained using inherently cultural textual data, display cultural tendencies when used across different human languages. This is evident in two cultural psychology constructs: social orientation and cognitive style.

GPT Analysis

An analysis of GPT’s responses to a set of measures in both Chinese and English reveals that when used in Chinese, GPT shows a more interdependent social orientation and a more holistic cognitive style, versus a more independent social orientation and analytic cognitive style in English.

ERNIE Replication

These cultural tendencies were also found in ERNIE, a Chinese generative AI model.

Real-World Impact

The impact of these cultural tendencies is demonstrated by GPT's recommendations. When used in Chinese, GPT is more likely to recommend advertisements with an interdependent social orientation, compared to an independent one in English.

Cultural Prompts

Exploratory analyses suggest that cultural prompts (such as prompting generative AI to act as a Chinese person) can adjust these cultural tendencies.

Data are publicly available. Data were analysed using R (version 4.3.1) in RStudio (version 2024.04.2+764). Analysis code is publicly available.