WorksExperimentsAbout Me

Lost in Translation



If you ask an AI what "Freedom" means in English, then ask again in Chinese—does it give the same answer?
YearApril 2026RoleResearch, Design, DevelopmentTechnologiesPython, OpenAI API, TouchDesigner
This project investigates whether AI semantic space is culturally neutral by measuring how the same abstract social value concept shifts when processed by a large language model across four languages: English, Chinese, Spanish, and Russian. Using cosine similarity analysis, angular displacement values are computed for each concept-language pair and mapped to L-System tree structures rendered in real time via TouchDesigner. The resulting forms make invisible computational differences in meaning visible as organic, branching geometries. The trees do not represent what these words mean to people; they represent what the model has learned to associate with them.#Data Visualization
#Critical AI Studies
#Computational Art
#Multilingual NLP
“Melancholy”: melancholy (English) · 忧郁 (Chinese) · Melancolía (Spanish) · Тоска (Russian)

Research Question

When the same word is processed by a LLM in different languages, does it mean the same?


This project investigates a deceptively simple idea: that AI semantic space is not culturally neutral. Abstract social value concepts like "Shame," "Freedom," or "Loyalty" carry different cultural weights across languages, and when encoded by a large language model, the resulting high-dimensional vectors diverge in measurable ways. This divergence, which we term Cross-lingual Semantic Embedding Divergence (CSED), is not a technical error. It is a trace of whose language, whose texts, and whose conceptual frameworks were most present when the model learned to understand the world.

Technical Solutions

A four-stage pipeline links natural language input to real-time visual output:
  • Concept selection: 20 abstract social value concepts, each translated into English, Chinese, Spanish, and Russian
  • Embedding extraction: OpenAI text-embedding-3-small, called directly within TouchDesigner's Python environment
  • CSED computation: cosine similarity between English and each language → converted to angular offset θ
  • L-System rendering: θ drives branch angle; training data share (Common Crawl, log-normalized) drives branch thickness


Findings

The study measured CSED values for 20 concepts across three languages. Several patterns emerged.

Overall, Russian shows the highest average divergence (≈65.2°), followed by Chinese (≈55.6°), and Spanish the lowest (≈42.1°) — likely reflecting typological distance from English. Emotionally charged, introspective concepts consistently diverge more than institutional or behavioural ones.

Three data points stand out:
  • Тоска (Melancholy, Russian) — 108.10° · A culturally untranslatable term for diffuse longing; nearly perpendicular to the English branch
  • 羞耻 (Shame, Chinese) — 72.85° · In Chinese cultural context, shame is collective and face-related rather than individual — a divergence that shows up clearly in the model's encoding
  • Obediencia (Obedience, Spanish) — 17.25° · Shared Latin etymology with English produces near-perfect semantic alignment
melancholy · 忧郁 · Melancolía · Тоска
shame · 羞耻 · Vergüenza · Стыд
obedience · 服从 · Obediencia · Послушание

Why it matters?

AI semantic space is not a neutral map of human meaning; it is shaped by training data, reflecting whose language, whose texts, and whose cultural frameworks were most present when the model learned. This project makes that bias measurable and visible: the angular divergence between branches is not a technical error, but a geometric trace of how differently the same concept has been encoded across languages and cultures.

Next steps

  • Introduce color rendering to the tree structures: mapping the semantic attributes of each concept to a color palette, so that the visual output encodes not just divergence but also the affective register of the concept itself
  • Expand the concept set to test whether the pattern of higher divergence in emotionally charged concepts holds more broadly
  • Test alternative models trained on non-English-dominant corpora (e.g. DeepSeek) to observe how semantic divergence shifts when the dominant training language changes

Reference

Vromen, E. (2024). Language Models as Semiotic Machines. arXiv:2410.13065
Bateman, J. A. (2024). What are Large Language Models Doing? University of Helsinki Language Technology Seminar.
Harris, Z. S. (1954). Distributional Structure. Word, 10(2–3), 146–162.
Mikolov, T. et al. (2013). Efficient Estimation of Word Representations in Vector Space. arXiv:1301.3781

©2026 by Han <3