💬 NLP

Language Models Entangle Language and Culture

This paper demonstrates that large language models systematically provide lower-quality responses and exhibit altered cultural contexts for users querying in low-resource languages, thereby creating systemic disadvantages based on language choice.

Original authors: Shourya Jain, Paras Chopra

Published 2026-01-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shourya Jain, Paras Chopra

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a giant, magical library where a super-smart robot librarian (the Large Language Model) helps you with any question you have. You might ask about how to fix your sleep, how to start a business, or how to invest your savings.

This paper is like a report card for that robot librarian, but with a twist: the researchers wanted to see if the librarian treats you differently depending on what language you speak.

Here is the story of what they found, broken down simply:

1. The "Magic Mirror" Test

The researchers created a list of 20 everyday questions, like "How do I get better at sleeping?" or "What's a good business plan for selling shirts?" They didn't ask these questions in just English. They translated them into Hindi, Chinese, Swahili, Hebrew, and Brazilian Portuguese.

They then asked the robot librarian these same questions in all these different languages.

The Result: The librarian wasn't fair. When people asked in English, Chinese, or Portuguese, the answers were high-quality, detailed, and helpful. But when people asked in languages like Hindi, Swahili, or Hebrew, the answers were often shorter, less detailed, and just plain worse. It's as if the librarian has a "favorite language" and puts less effort into the others.

2. The "Cultural Costume" Change

Here is where it gets really interesting. The researchers discovered that the language you speak doesn't just change how well the robot answers; it changes what kind of world the robot imagines.

Think of the robot as an actor.

If you ask the question in English, the actor puts on a "Western" costume. They might give advice based on American or European habits, values, and examples.
If you ask the exact same question in Hindi, the actor swaps the costume for an "Indian" one. They suddenly start thinking about Indian family structures, local food, or different social norms.
If you ask in Swahili, they put on an "African" costume.

The researchers proved this by taking the answers given in Hindi, translating them back to English, and asking a second robot (a "Judge") to guess the culture. The Judge could still tell, "This answer was originally written in Hindi," because the flavor of the advice was different. The language didn't just change the words; it changed the cultural lens the robot used to think.

3. The "Factual Map" Test

To double-check this, they used a map of cultural facts (a benchmark called CulturalBench). They asked the robot questions like "What is a famous dish in Brazil?" or "Who is a historical figure in Nigeria?"

They found that the robot's accuracy depended heavily on the language used. The robot knew more facts about a country when asked in that country's language (or a major language associated with it) compared to when asked in a different language. It's like the robot has different "fact books" for different languages, and some of those books are much more complete than others.

4. Why Does This Happen?

The paper suggests that for these AI models, language and culture are tangled together, like two vines growing around the same tree. You can't easily separate them.

Because the AI was trained on a massive amount of internet text, it learned that "English" is often linked with "Western culture," and "Hindi" is linked with "Indian culture." When you switch the language, the AI automatically switches the cultural context it pulls from its memory.

The Bottom Line

The main takeaway is simple: If you ask an AI a question in a "low-resource" language (one with less data on the internet), you might get a worse answer than if you asked in English.

Even worse, the answer you get might be culturally different, not just in quality, but in the very assumptions it makes about your life. The researchers argue that this is unfair. Just like a human should be able to ask for advice in their native tongue and get the same good help as someone speaking English, these AI models need to be trained better so they don't leave users behind based on the language they speak.

Technical Summary: Language Models Entangle Language and Culture

Problem Statement

Current evaluations of Large Language Models (LLMs) in multilingual contexts predominantly focus on general knowledge, instruction-following, mathematical reasoning, or programming capabilities. These benchmarks often rely on multiple-choice questions (MCQs) or short-form answers, failing to capture the stylistic and cultural variations introduced when users interact with models in different languages. Furthermore, existing bias studies often rely on explicit cultural cues (e.g., names, nationalities) within prompts, which do not reflect how users naturally frame generic, open-ended queries.

The core problem addressed is whether users are systemically disadvantaged by the language they use. Specifically, the authors investigate if answer quality and the cultural context embedded in responses vary significantly based on the query language, potentially impacting decision-making for users of low-resource languages.

Methodology

1. Dataset Construction (WildChat Analysis)

To address the gap in evaluating generic advice-seeking queries, the authors analyzed the WildChat dataset (Zhao et al. [2024]).

Filtering: They retained only English queries, removed programming/error-fix queries (to avoid niche skew), and filtered for lengths between 40–400 characters.
Clustering: Using the fuzzywuzzy library for deduplication and Qwen3-0.6b for embeddings, they clustered queries via the HDBSCAN algorithm.
Selection: From these clusters, they curated 20 culture-neutral, open-ended advice-seeking questions covering domains like Healthcare, Business, Education, and Job Interviews.
Translation: These queries were translated into Chinese, Hindi, Brazilian Portuguese, Swahili, and Hebrew using the Gemini-2.5-Flash model (temperature 0).

2. Models Evaluated

The study evaluated five distinct model families to ensure provider diversity:

Qwen3-14B (Chinese provider)
Cohere-Aya-32B and Cohere-Aya-8B (Canadian provider, multilingual-focused)
Magistral-small (French provider, Mistral-AI)
Sarvam-m (Indian provider, fine-tuned Mistral-Small)

3. Evaluation Framework (LLM-as-a-Judge)

The authors employed an LLM-as-a-Judge approach to score responses on a scale of 1–5.

Judge Model: Cohere Command-A was selected for its advanced multilingual capabilities.
Ablation for Judge Robustness: To ensure the judge itself was not biased, the authors tested six configurations. The optimal configuration involved providing the original query and response along with 8 randomly chosen reference responses (without evaluation), which yielded the highest alignment (Pearson correlation and Cohen's Kappa) with ground truth scores.
Scoring Rubrics: Responses were evaluated on Detail/Completeness, Linguistic Quality, Factual Correctness, Actionability, and Relevance.

4. Cultural Entanglement Analysis

To determine if language and culture are entangled:

Cultural Classification: All non-English responses were translated to English and classified by the LLM-as-a-Judge into specific cultural contexts (English/Western, Indian, Chinese, African, Latin American, Jewish) based on tone, values, and worldview.
CulturalBench Evaluation: A translated subset of the CulturalBench (Chiu et al. [2025]) benchmark (750+ questions across 29 regions) was used to test factual cultural knowledge across languages.
Control Experiments: The authors tested for random perturbations by appending random strings to queries, comparing the variance against language-based variance.

Key Results

1. Quality Disparities Across Languages

The evaluation revealed that LLMs consistently provide lower-quality answers in low-resource languages compared to English.

Statistical Significance: Kruskal–Wallis tests confirmed statistically significant differences in response quality across languages for all models ( $p < 0.05$ ).
Specific Findings: Responses in Hindi, Swahili, and Hebrew were consistently worse than those in English, Chinese, and Brazilian Portuguese. Even models specifically trained for multilingual use (Cohere-Aya) showed degraded performance in Swahili.
Model Size Impact: Larger models (Aya-32B) showed smaller performance gaps between languages compared to smaller variants (Aya-8B), suggesting model scale improves cross-lingual consistency.

2. Language-Culture Entanglement

The study found that language choice fundamentally alters the cultural context of the response, even for generic queries.

Cultural Classification: When non-English responses were translated to English, the LLM-as-a-Judge could still classify them according to the cultural context of the original query language (e.g., a query in Hindi generated responses with Indian cultural markers, while a query in Hebrew generated responses with Jewish cultural markers).
CulturalBench Performance: Accuracy on cultural factual questions varied significantly across languages for the same model (Qwen3-14B). The Kruskal-Wallis test on this data yielded $H = 45.5158$ and $p = 1.1395 \times 10^{-8}$ , confirming that performance differences are not random but tied to the language-culture pairing.
Control: Adding random strings to queries reduced performance but did not replicate the specific, structured variations seen when changing languages, ruling out simple noise as the cause.

Key Contributions

Generic Advice Dataset: Creation of a set of 20 culture-neutral, open-ended advice-seeking queries derived from real-world WildChat data.
Qualitative Evaluation: The first qualitative evaluation of LLM responses across languages for generic, advice-seeking queries, moving beyond MCQs and niche domains.
Translated CulturalBench: Creation and release of a translated version of the CulturalBench dataset to facilitate cross-lingual cultural evaluation.
Entanglement Demonstration: Empirical evidence demonstrating that language and culture are entangled in LLMs; the choice of query language dictates the cultural context used in the response, which in turn impacts answer quality.

Significance and Claims

The paper claims that users interacting with LLMs in low-resource languages are systemically disadvantaged not only by lower answer quality but also by the imposition of different cultural contexts that may not align with their needs or expectations.

The authors argue that current multilingual benchmarks fail to capture these nuances. They conclude that the relationship between language and culture in LLMs is not merely a translation issue but a fundamental architectural and training data phenomenon. The paper calls for:

Improved multilingual training data and training methods to increase uniformity in response quality.
Further research to identify and mitigate biases that negatively affect groups based on language.
The development of evaluation methods that account for cultural context, not just factual accuracy.

The work is modest in its scope, limited to small-to-moderate open-source models and LLM-based evaluation, but it establishes a critical baseline for understanding how language choice shapes the utility and cultural alignment of AI assistants.