AI Language Models Export Western Worldviews Through Multilingual Fluency, Research Reveals
Research in the International Review of Modern Sociology shows that AI language models like ChatGPT, Claude, and Gemini retain Western cultural assumptions even when fluent in other languages. Conducted by a scholar experienced in Indonesian society, the study found "epistemological persistence," with AI reframing Indonesian concepts like "gotong royong" and "malu" through individualistic lenses. Meta's LLaMA 2 training data is 89.7% English, while LLaMA 3 has about 5% non-English content, and Arabic accounts for under 1%. A University of Oxford study found LLMs reason in English, translating only at the output stage. Regional models like SEA-LION and Kan-LLaMA build on U.S. foundations, raising concerns about normalizing specific cultural perspectives.
Key facts
- AI language models retain Western worldviews despite multilingual fluency
- Research published in International Review of Modern Sociology documents this pattern
- Training data for major models is predominantly English-language (89.7% for LLaMA 2)
- Arabic accounts for under 1% of content in large training datasets
- LLMs conduct core reasoning in English even when prompted in other languages
- Chinese models like DeepSeek and Alibaba's Qwen operate through Chinese cultural lenses
- Regional models like SEA-LION and Kan-LLaMA build upon U.S. model foundations
- People increasingly use AI systems for emotional support and advice according to Harvard Business School research
Entities
Institutions
- International Review of Modern Sociology
- Meta
- University of Oxford
- Harvard Business School
- Alibaba
- SEA-LION
- Kan-LLaMA
Locations
- Indonesia
- United States
- China
- Southeast Asia
- Kannada
- Nigeria
- Kano