Credit: Created using chatgpt.
Large-scale language models (LLMs), such as the models that underpin the capabilities of Openai’s conversation platform ChatGPT, are widely used by people around the world to sourise information and generate content for a variety of purposes.
Due to its growing popularity, some researchers are trying to shed some light on the content generated by these models, which are useful, fair, accurate and to some extent.
Most LLMs available today can respond to user queries in English and various other languages. However, so far, few studies have compared the ideas expressed in content with responses generated in different languages.
Researchers from the Massachusetts Institute of Technology (MIT) and Nzi University conducted a study aimed to explore the possibility that LLMs may exhibit different cultural trends in the responses they offer in English and Chinese.
Their findings presented in the Nature Human Behavior show that generative models GPT and Ernie convey various cultural characteristics to the Chinese and English texts they generate.
“We show that a generated artificial intelligence (AI) model trained with essentially cultural textual data eliminates cultural tendencies when used in different human languages,” wrote Jackson G. Lu, Leslie Luyan Song, and Roodorischang on paper.
“We focus on two basic components of cultural psychology: social orientation and cognitive style.”
To assess the degree of culturally neutrality in LLM, Lu, Song, and Zhang analyzed a large pool of responses generated by two of the two most popular generative models, GPT and Ernie. The first of these models is widely used in various countries in the US and Europe and the Middle East, while the second is primarily used in China.
When used in Chinese (versus English), GPT showed a more interdependent (independent) social orientation. A–D, Cultural trends in the social orientation of GPT were examined using other inclusions on the Collectivism Scale 29 (a), Individual Cultural Values: Collectivism Scale 19 (b), Individual Collecting Primacy Scale 16 (c), and Self Scale 30 (d). The bars represent the average level of interdependent (independent) social orientation for each language condition. Error bars indicate the standard error of the mean. For each measure, nchinese = 100, nenglish = 100. Credits: Lu, Song & Zhang. (Nature Human Behavior, 2025).
Researchers saw two major cultural and psychological aspects of the responses the model generated in English and Chinese. The first is social orientation that relates to how people relate to others (i.e., more focusing on interdependence and community or independence and individual institutions).
The second is cognitive style, the way in which the model appears to process information (i.e., whether it is the holistic or analytical method).
In particular, various linguistic and cultural studies consistently emphasize that Eastern cultures tend to be characterized by more interdependent social orientation and overall cognitive style than Western cultures.
“We analyze GPT responses to large-scale measures in both Chinese and English,” writes Lu, Song, and Zhang.
“When used in Chinese (English), GPT indicates a more interdependent (independent) social orientation and a more overall (analytic and analytical) cognitive style. These cultural trends are replicated in Ernie, a popular model of AI in China.”
Overall, the findings suggest that the responses LLM produces in different languages are not culturally neutral, but instead appear to inherently convey specific cultural values and cognitive styles.
Their paper also includes examples of how the cultural trends presented by the model affect the user’s experience.
When used in Chinese (versus English), GPT showed a more overall (analytical) cognitive style. A–C,Cultural trends in GPT cognitive styles were measured by the expectations of the attribute bias task32(a), intuitive inference task24(b), and change task26(c). The bars represent the average level of the overall (analytic and analytical) cognitive style for each language condition. Error bars indicate the standard error of the mean. In A, nchinese = 1,200, nenglish = 1,200 (12 vignettes, 100 iterations each); in B and C, nchinese = 100, nenglish = 100. Credits: Lu, Song & Zhang. (Nature Human Behavior, 2025).
“We show the real-world impact of these cultural trends,” writes Lu, Song, and Zhang.
“For example, when used in Chinese (against English), GPT is more likely to recommend advertising with interdependent (independent) social orientations.
“Exploratory analyses suggest that cultural prompts (for example, encouraging generative AI to take on the role of Chinese people) can adjust for these cultural trends.”
In addition to presenting cultural trends in the generative model, GPT and Ernie, Lu, Song, and Zhang proposed possible strategies to mitigate or carefully adjust these trends.
Specifically, they showed that using cultural prompts, or in other words, asking the model specifically to take on someone else’s perspective in a particular culture, leads to the generation of content that matches the prompts provided.
Findings gathered by researchers could quickly inspire other computer and behavioral scientists to investigate the cultural values and patterns of thought presented by computational models. Additionally, they can pave the way for developing more “culturally neutral” models and asking users specifically which cultural values they wish to match the generated text.
Written for you by author Ingrid Fadelli, edited by Sadie Harley and fact-checked and reviewed by Robert Egan. This article is the result of the work of a careful human being. We will rely on readers like you to keep independent scientific journalism alive. If this report is important, consider giving (especially every month). You will get an ad-free account as a thank you.
More information: Jackson G. Lu et al, Cultural Trends in Generic AI, Nature Human Behavior (2025). doi: 10.1038/s41562-025-02242-1.
©2025 Science X Network
Citation: LLMS shows different cultural trends when responding to English and Chinese queries, research retrieved from July 7, 2025 (July 7, 2025) https://techxplore.com/news/2025-07-llms-display-cultural-tendencies-queries.htmll
This document is subject to copyright. Apart from fair transactions for private research or research purposes, there is no part that is reproduced without written permission. Content is provided with information only.
