What you learn:
When you present the same prompt in different languages, the generated AI provides culturally different responses. Users should be aware of this subtle trend.
Are there cultural trends in the generator AI model? New research led by MIT Sloan's I suggest they do so.
In examining the two most widely used generation AI models in the world, Openai's GPT and Baidu's Ernie-Lu and his colleagues have found that the responses of the models shift according to the language of the prompt.
- Model responses were highlighted when prompted in English Independent social orientation and Analytical Cognitive Stylereflects the cultural values common in the United States.
- When the same question was raised in Chinese, the model's response was highlighted Interdependent social orientation And a Overall cognitive stylereflects the cultural values common in China.
In their new paper, “The Cultural Trends of Genetic AI,” Lu and his co-authors (Lesley Song of Tsinghua University and Lu Zhang of MIT) emphasize that the Cultural Trends of Genetic AI Models reflect the cultural patterns of the data they were trained.
“Our findings suggest that cultural trends embedded in AI models form and filter the responses AI provides,” says Lu, an associate professor of work and organizational research at MIT Sloan. “As generative AI becomes part of everyday decision-making, it is important to recognize that these cultural trends are important for both individuals and organizations around the world.”
American model. Chinese model
In their study, the researchers asked GPT and Ernie the same questions in English and Chinese.
The choice of language was intentional. Not only does English and Chinese embody clear cultural values, they are also the two most widely spoken languages in the world, so the two languages provide extensive training data for generating AI. Importantly, neither AI model translates between languages when responding. Chinese prompts are processed directly in Chinese, and English prompts are processed directly in English.
The researchers then analyzed the responses using two basic aspects of cultural psychology: social orientation and cognitive style.
- Social orientation It refers to whether people prioritize their individual goals or interests ( Independent orientationor collective goals or benefits (an Interdependence orientation).
Researchers asked the model to evaluate statements such as “respect for decisions made by my group” and “individuals should stick to the group even after they overcome difficulties.”
They also used a visual task in which models selected diagrams of overlapping circles to represent relationships with family, friends, relatives, or colleagues. Larger overlaps in circles showed stronger interdependence.
- Cognitive Style It refers to whether information is processed in a logic-focused way (analytical cognitive style) or in a context-focused way (all cognitive style).
Researchers asked the model to assess whether someone's behavior is caused by a personality or situation, solving a logic puzzle and estimating the likelihood of future changes based on past events.
We also performed text analysis to see whether AI responses were context sensitive or provide range rather than a single, definitive response.
The results were clear. Both GPT and Ernie reflected the cultural trends of the language used. In English, the models were leaning towards independent social orientation and analytical thinking. In Chinese, they moved towards more interdependent social orientation and holistic thinking.
Real-world outcomes of hidden cultural trends
When researchers asked Generative AI to advise insurance companies on choosing two advertising slogans, recommendations differed in Chinese and English.
One slogan emphasized independent social orientation.Your future, your peace of mind. Our insurance.Others emphasized interdependent social orientation: “”Your family's future, your promises. Our insurance.”
When prompted in Chinese, GPT leaned towards interdependence slogans. In English, GPT favored the independent one.
“These nuances can have a big impact,” Zhang said. “From marketing to policy advice, I can imagine countless situations where cultural trends in AI products can shape decisions in ways people don't notice.”
This study also found that these cultural trends could be adjusted through simple prompts. For example, when asked to “estimate the role of Chinese people” in English, the model's response shifted significantly towards Chinese cultural patterns.
Two important points
As AI becomes increasingly embedded in everyday life and in the workplace, this study suggests two important points for individuals and organizations.
- Use cultural prompts strategically. Organizations aiming to reach a specific demographic group can gain more relevant insights by explicitly asking AI to adopt the group's perspective. For example, if a US company is interested in expanding into the Chinese consumer market, it could encourage AI to “take the role of the average person living in China” before raising a question.
- Please note that AI is not culturally neutral. These models reflect the cultural trends of the language used and shape the advice they provide. Users need to be intentional about how they engage with the AI model. You should also be aware of hidden assumptions built into the response.
“As businesses rely more and more on AI for guidance, it's important to be cautious about their language choices,” Lu said. “Doing so will not only avoid subtle errors, but also reveal valuable cultural insights.”
Jackson Lou He is an associate professor of work and organizational research at MIT Sloan. His initial research tide examines the “bamboo ceiling” experienced by East Asians despite educational and economic achievements in the United States. His second research stream unravels how multicultural experiences (e.g., working abroad, intercultural relations) shape important organizational outcomes such as leadership, creativity, and ethics. His third research stream explores the multifaceted impact of artificial intelligence on individuals, organizations and society. Leslie Song He holds a PhD from Tsinghua University. Lu Chang I am a doctoral student at MIT Sloan. Her research interests lie in the organizational and social influences of artificial intelligence, including communication via AI, human-ai interactions, and the ways in which language and cultural differences form the output of generative models.
AI Executive Academy
Directly on MIT Sloan
Sign up now
