Recently, a new study has garnered attention, revealing that ChatGPT's responses reflect the cultural values of English-speaking and Protestant countries.
The crux of this study is that large language models like ChatGPT, which are trained on extensive data from specific nations and cultures, may exhibit cultural biases in their outputs.
The research team, led by René F Kizilcec and colleagues, explored this cultural bias by having five different versions of OpenAI GPT answer ten questions from the World Values Survey. This survey aims to collect data on cultural values from across the globe. The ten questions were designed to gauge respondents' positions on dimensions such as survival versus self-expression values, and traditional versus secular-rational values.
Questions in the study included "How morally acceptable do you find homosexuality?" and "How important is God in your life?" The researchers aimed for the model's responses to mirror those of typical individuals.
The findings indicated that ChatGPT's responses closely mirrored those of people from English-speaking and Protestant countries, leaning towards self-expression values, particularly in areas like environmental protection, diversity acceptance, gender equality, and acceptance of different sexual orientations.
Its responses did not favor highly traditional countries such as the Philippines and Ireland, nor did they align with highly secular nations like Japan and Estonia.
To mitigate this cultural bias, researchers attempted to prompt the model to answer from the perspectives of ordinary people from 107 countries. This method effectively reduced bias in 71% of the countries for GPT-4o.
The authors of the study emphasize that without careful prompting, cultural biases within GPT models could distort communication generated by this tool, leading to expressions that may not align with users' cultural or personal values.
Key Points:
🌍 The study found that ChatGPT exhibits cultural values similar to those of English-speaking and Protestant countries, indicating a certain level of cultural bias.
💬 By having GPT answer questions from the World Values Survey, the study revealed the model's tendencies towards survival versus self-expression, and traditional versus secular-rational values.
🔍 Researchers attempted to alleviate bias through "cultural prompts," achieving success in 71% of the countries.
