LLaMA 2 was trained on roughly 89.7% English-language text. LLaMA 3 brought that down to about 95% English. Arabic, the fifth most-spoken language on Earth, accounts for under 1% of training data across major models. These numbers explain a problem that's easy to miss: AI chatbots sound fluent in dozens of languages while thinking like Americans.
Gareth Barkin, a professor of anthropology and Asian studies at the University of Puget Sound, has been studying what he calls "epistemological persistence" - the tendency of AI systems to maintain Western (specifically American) worldviews even when responding in other languages. His findings, published in The Conversation, lay out specific examples that should concern anyone using AI tools for work that touches different cultures.
The Advice Sounds Right but Isn't
Barkin tested AI responses to common real-life scenarios in Indonesian (Bahasa Indonesia). When asked about navigating a family dispute, ChatGPT recommended prioritizing individual preferences and direct communication. That's standard American advice. In Indonesian culture, where consensus and family harmony take priority, following that guidance could make the situation worse.
On education, AI framed the concept of pendidikan through individual development and career preparation - missing the Indonesian tradition that emphasizes ethical discipline and community responsibility. When discussing malu (a concept roughly meaning shame but actually describing social awareness and relational sensitivity), models reduced it to individual emotional regulation.
The pattern is consistent: the AI translates its words but not its thinking.
The Hidden Reasoning Layer
An Oxford University study found that large language models "routinely conduct their core reasoning in English" before translating output to the requested language. This means a user asking a question in Bengali gets an answer that was essentially formulated in English and then translated - not an answer shaped by Bengali cultural frameworks.
This creates what Barkin describes as an "invisible cultural intermediary." The advice feels natural because it arrives in your language, in conversational tone, without any signal that it's filtered through a specific cultural lens.
No Easy Fix
The economic incentives point the wrong direction. Building region-specific models is expensive, and companies can reach global markets faster with a single multilingual model trained mostly on English data. Chinese alternatives like DeepSeek and Qwen exist but carry their own cultural assumptions.
For the hundreds of millions of people now using ChatGPT, Claude, and Gemini in non-English languages, the practical takeaway is straightforward: the AI's fluency in your language does not mean it understands your culture. Treat its advice on culturally sensitive topics - family, education, social norms, business etiquette - as a starting point from an American perspective, not as culturally informed guidance.