Research published on April 2, 2026, in the International Review of Modern Sociology found that major artificial intelligence systems retain a Western worldview despite their multilingual fluency.

This phenomenon, which the study terms ‘epistemological persistence,’ suggests that an AI’s ability to translate language does not equate to a translation of cultural context. The immediate consequence, according to one analysis, is that global users may receive advice or information that appears linguistically native but is philosophically rooted in Western, often American, cultural norms. This observation presents a complex challenge to the notion of universally applicable, AI-generated knowledge, raising questions about the cultural appropriateness of information disseminated by these widely used systems.

What We Know So Far

  • A study published on April 2, 2026, found that large language models, including ChatGPT, preserve Western cultural worldviews, according to a report from letsdatascience.com.
  • The study, published in the International Review of Modern Sociology, identifies this as "epistemological persistence," where a model’s core worldview remains unchanged even when communicating in different languages.
  • The underlying worldview of these models is shaped by training data predominantly from English-language sources based in the United States, reports theconversation.com.
  • In one case reported by theconversation.com, AI advice given in Indonesian regarding a family matter was rooted in American cultural assumptions, prioritizing individual autonomy over Indonesian traditions of collective family dynamics.
  • Experiments cited in the same report showed AI models consistently framed the Indonesian concept of 'malu' (a term relating to shame, respect, and social harmony) as an individual emotional experience, diverging from its complex, communal meaning.
  • Data from theconversation.com indicates that Meta’s LLaMA 2 model was trained on approximately 89.7% English-language text, while its successor, LLaMA 3, includes only about 5% non-English data.