title: Cultural Homogenization
aliases: [cultural homogenization, homogenization, cultural flattening]
tags: [term]
category: Ethics & governance
summary: The tendency of AI systems to flatten diverse cultures toward the dominant (largely Western, English-language) patterns in their training data.

Cultural Homogenization

Generative AI systems learn from whatever they are trained on, and that material is overwhelmingly English-language and Western. As a result, their default outputs tend to pull toward those patterns: the imagery, the storytelling, the assumptions about what is normal or neutral. When a model is used to generate, translate, or summarize across the world's cultures, it can quietly replace locally-specific knowledge and narratives with a more uniform, Americanized version, eroding the diversity that makes those cultures distinct.

This matters for language work because language carries culture, not just grammar. A tool that produces fluent but culturally-flattened text can push a community toward someone else's ways of expressing things while appearing helpful, a subtler cousin of unwanted standardization and closely tied to misrepresentation. It is also part of why community control over inputs and outputs matters: the "average" models encode are rarely the community's own.