Large language models (LLMs), such because the model underpinning the functioning of OpenAI’s conversational platform ChatGPT, are actually extensively utilized by folks worldwide to supply info and generate content material for numerous functions.
Due to their rising recognition, some researchers have been making an attempt to make clear the extent to which the content material generated by these models is beneficial, unbiased, and correct.
Most LLMs accessible immediately can reply to consumer queries in English and numerous different languages. Yet only a few research to date have in contrast the concepts expressed within the responses and content material they generate in several languages.
Researchers at Massachusetts Institute of Technology (MIT) and Tongji University carried out a research geared toward investigating the likelihood that LLMs exhibit totally different cultural tendencies within the responses they supply in English and Chinese.
Their findings, published in Nature Human Behavior, present that the generative models GPT and ERNIE convey totally different cultural traits within the Chinese and English texts they generate.
“We present that generative synthetic intelligence (AI) models—skilled on textual knowledge which might be inherently cultural—exhibit cultural tendencies when utilized in totally different human languages,” wrote Jackson G. Lu, Lesley Luyang Song and Lu Doris Zhang of their paper.
“We concentrate on two foundational constructs in cultural psychology: social orientation and cognitive model.”
To assess the extent to which LLMs are culturally impartial, Lu, Song and Zhang analyzed a big pool of responses generated by GPT and ERNIE, two of the most well-liked generative models. The first of those models is extensively used within the U.S. and in numerous international locations throughout Europe and the Middle East, whereas the second is primarily utilized in China.
The researchers checked out two fundamental cultural and psychological points of the responses that the models generated in English and Chinese. The first is social orientation, which pertains to how folks relate to others (i.e., focusing extra on interdependence and group or independence and particular person company).
The second is cognitive model, or, in different phrases, how the models seem to course of info (i.e., whether or not in a holistic or analytic means).
Notably, numerous linguistic and cultural research persistently highlighted that Eastern cultures are usually characterised by a extra interdependent social orientation than Western ones, in addition to a holistic cognitive model.
“We analyze GPT’s responses to a big set of measures in each Chinese and English,” wrote Lu, Song and Zhang.
“When utilized in Chinese (versus English), GPT displays a extra interdependent (versus impartial) social orientation and a extra holistic (versus analytic) cognitive model. We replicate these cultural tendencies in ERNIE, a well-liked generative AI model in China.”
Overall, the findings counsel that the responses that LLMs produce in totally different languages usually are not culturally impartial, however as an alternative they seem to inherently convey particular cultural values and cognitive types.
In their paper, the researchers additionally embody examples of how the cultural tendencies exhibited by the models may have an effect on the {experience} of customers.
“We display the real-world affect of those cultural tendencies,” wrote Lu, Song and Zhang.
“For instance, when utilized in Chinese (versus English), GPT is extra prone to suggest commercials with an interdependent (versus impartial) social orientation.
“Exploratory analyses counsel that cultural prompts (for instance, prompting generative AI to imagine the function of a Chinese particular person) can alter these cultural tendencies.”
In addition to unveiling the cultural tendencies of the generative models GPT and ERNIE, Lu, Song and Zhang proposed a potential technique to mitigate these tendencies or rigorously alter them.
Specifically, they confirmed that utilizing cultural prompts, or, in different phrases, particularly asking a model to tackle the attitude of somebody from a particular tradition, led to the era of content material that was aligned with the prompts offered.
The findings gathered by the researchers may quickly encourage different pc scientists and behavioral scientists to analyze the cultural values and pondering patterns exhibited by computational models. In addition, they may pave the best way for the event of models which might be extra ‘culturally impartial’ or that particularly ask customers what cultural values they want a generated textual content to be aligned with.
Written for you by our writer Ingrid Fadelli,
edited by Sadie Harley, —this text is the results of cautious human work. We depend on readers such as you to maintain impartial science journalism alive.
If this reporting issues to you,
please contemplate a donation (particularly month-to-month).
You’ll get an ad-free account as a thank-you.
More info:
Jackson G. Lu et al, Cultural tendencies in generative AI, Nature Human Behaviour (2025). DOI: 10.1038/s41562-025-02242-1.
© 2025 Science X Network
Citation:
LLMs show totally different cultural tendencies when responding to queries in English and Chinese, study finds ( 7)
15
llms-display-cultural-tendencies-queries.html
The content material is offered for info functions solely.