When AI Translates a Word, Whose Meaning Does It Choose?

A growing scientific literature examines how large language models may reshape work, productivity, and task performance, from writing to customer support and software development (Acemoglu 2025; Eloundou, Manning, Mishkin, Rock 2024; Noy and Zhang 2023; Peng et al. 2023; Brynjolfsson et al. 2023). Companies like Google, IBM, and Microsoft also use large language models (LLMs) in various ways, including email text suggestions, language classifications, and content recommendations.

A related literature evaluates LLMs as translation systems. Recent studies ask how well Generative Pre-trained Transformer (GPT) models translate across languages, how prompting strategies affect translation quality, and whether document-level or context-aware prompts improve performance (Hendy et al. 2023; Jiao et al. 2023; Vilar et al. 2022; Zhang et al. 2023; Wang et al. 2023). Other work applies LLM translation to scientific research, showing both high accuracy and a more subtle problem: domain experts sometimes prefer not to translate certain technical terms in English when translation makes them less familiar (Kleidermacher and Zou 2026). This blog post builds on that literature but shifts the unit of analysis: instead of asking whether an LLM can translate a sentence or document well, I ask whether it can preserve the contextual meaning of culturally and philosophically dense Chinese terms whose English glosses are useful but incomplete.

Existing work on LLM translation often evaluates accuracy at the sentence, document, or domain level. My experiment focuses on the smaller unit possible: culturally and philosophically dense Chinese terms, which are often made up of only one Chinese character. These Chinese terms are not impossible to translate, but their translations are often partial and don’t capture the full cultural, historical, and philosophical meanings behind the terms. The question that underlines the experiment is therefore not only whether an LLM can translate across languages, but whether it can explain what its own translation leaves out when it happens.

From Overtranslation to Culturally Dense Terms

Kleidermacher and Zou (2026) handle that subjectivity through in-context learning: they give GPT-4o a curated example from a scientific article whose author has reviewed the translation and identified terms that should remain in English. That example then guides how the model translates later texts.

This blog post adapts that idea to a different problem. Instead of asking which scientific terms should remain in English, I ask how LLMs handle culturally and philosophically dense Chinese terms whose English translations are useful but incomplete. A summary of these terms and why they are culturally and philosophically dense is presented in table 1. For example, should 仁 rén simply be translated as "benevolence"? Should 礼 lǐ become "ritual" or "etiquette"? Should 道 dào be translated as "the Way"?

The experiment focuses on two prompt conditions. In the term-only condition, the model sees only the Chinese term and is asked for a common English translation and what may be lost. In the passage-context condition, the model sees the term inside a short sentence. This comparison is inspired by the following question: what changes when the model is given contextual information that guides how a term should be handled?

Terms and Passage Contexts

The experiment uses five classical or philosophical Chinese terms. The common English gloss is not wrong, but it is often too thin to convey the meaning the term carries in context.

Table 1. Five terms used in the analysis.

Experiment Design