Lost in Translation: Visualizing Cultural Bias in AI Semantic Space

Abstract

Large language models (LLMs) encode language as high-dimensional vectors—a process that is both powerful and opaque. We ask whether the embeddings of abstract value concepts diverge across languages, and whether such divergence can be made visible. Using text-embedding-3-small, we embed 20 culturally loaded value concepts in English, Chinese, Spanish, and Russian, and convert the cosine similarity of each translation to its English baseline into an angular displacement we call Cross-lingual Semantic Embedding Divergence (CSED). We map CSED onto the branch angles of L-System trees rendered in TouchDesigner, with branch thickness encoding each language’s training-data share. Russian diverges most (mean 65.2◦), Spanish least (42.1◦), and emotionally loaded concepts drift farther than institutional ones. The work makes algorithmic bias perceptible to non-technical audiences and frames it as a property of skewed training data rather than of “true” meaning.