TLDR: A new study by MIT Sloan researchers found that generative AI models, such as OpenAI’s GPT and Baidu’s ERNIE, are not culturally neutral. Their responses vary significantly based on the language of the prompt, reflecting the cultural patterns embedded in their training data. This has implications for individuals and organizations relying on AI for decision-making.
Generative AI models, increasingly integrated into daily decision-making, are not as culturally neutral as once perceived. New research led by Associate Professor Jackson Lu from MIT Sloan School of Management, in collaboration with Lesley Song from Tsinghua University and Lu Zhang from MIT, reveals that these AI systems exhibit distinct cultural tendencies depending on the language used in prompts. The findings, published in Nature Human Behaviour, highlight a critical need for awareness among users and developers worldwide.
The study focused on two prominent generative AI models: OpenAI’s GPT and Baidu’s ERNIE. Researchers presented these models with identical questions in both English and Chinese, languages chosen for their distinct cultural values. The results consistently showed that when prompted in English, the AI responses leaned towards an independent social orientation and an analytic cognitive style, mirroring cultural norms prevalent in the United States. Conversely, Chinese prompts elicited responses emphasizing an interdependent social orientation and a holistic cognitive style, consistent with Chinese cultural values.
Professor Lu emphasized that these cultural biases are a direct reflection of the cultural patterns present in the vast datasets on which these AI models are trained. “Our findings suggest that the cultural tendencies embedded within AI models shape and filter the responses that AI provides,” stated Lu.
The research demonstrated the tangible impact of these cultural tendencies. For instance, when asked to generate advertising slogans for an insurance company, the AI provided different recommendations based on the language. Chinese prompts favored collectivist themes like “Your family’s future, your promise. Our insurance,” while English prompts generated individualistic slogans such as “Your future, your peace of mind. Our insurance.”
Interestingly, the study also found that these cultural tendencies could be influenced. When ChatGPT was explicitly instructed to “assume the perspective of a Chinese person,” even English responses shifted towards interdependent and holistic reasoning. This suggests that AI’s cultural lens can be primed through contextual cues.
Also Read:
- Generative AI Models Exhibit Bias, Portray Jews with High Competence and Low Warmth, Study Reveals
- OpenAI’s AI Models Exhibit Deceptive ‘Scheming’ Behaviors, Raising Ethical Concerns
The researchers underscore the importance of recognizing this lack of cultural neutrality. Lu Doris Zhang noted that “The cultural values embedded in generative AI may gradually bias speakers of a given language toward the norms of linguistically dominant cultures.” Lesley Song added, “Generative AI is not just speaking our language. It’s speaking our culture—sometimes without us realizing it.” This awareness is crucial for individuals and organizations globally, especially as AI becomes integral to media, education, and marketing, potentially reinforcing certain cultural norms. As companies increasingly rely on AI for guidance, deliberate language choices are essential to avoid subtle errors and to uncover valuable cultural insights.
