Large Language Models (LLMs) are often celebrated as "polyglots," capable of conversing in dozens of languages with near-native fluency. However, a growing body of research suggests that linguistic proficiency is not the same as cultural competence. While a model might speak Tagalog or Hindi perfectly, its underlying values, social norms, and factual associations often remain stubbornly anchored to a Western, English-centric worldview.
The Translation Layer Hypothesis
The core intuition behind this research is that LLMs often treat non-English languages as a mere "translation layer." Instead of accessing a distinct cultural knowledge base when prompted in Chinese or Swahili, the model performs internal reasoning using an English-centric "cognitive bridge."
As Akash Ghosh and colleagues note in The Multilingual Mind, this reliance on high-resource languages as an internal bridge causes models to neglect the context-specific logic inherent in other cultures. The result is a "cultural dominance" effect where, for example, a model might suggest Thanksgiving as a traditional holiday when queried about Chinese celebrations, a phenomenon documented by Wenxuan Wang and Wenxiang Jiao in Not All Countries Celebrate Thanksgiving.
The Paradox of Alignment
One might assume that as models become more "aligned" and "safe" through human feedback, they would become more sensitive to global cultural nuances. Surprisingly, the evidence suggests the opposite.
Wenxuan Wang 0001 and Jen-tse Huang found that as models evolved from GPT-3.5 to GPT-4, cultural bias actually intensified. This is likely because current alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF), often reflect the specific social norms and developer-imposed preferences of the Western teams building them. As Hannah Rose Kirk argues in Multilingual != Multicultural, models eventually reach a threshold where these ingrained biases override the cultural knowledge present in the training data.
Deep vs. Surface Cultural Understanding
To move beyond surface-level facts (like knowing a country's capital), researchers have developed benchmarks to probe "deep" cultural layers—values, social hierarchies, and invisible norms.
- Scenario-based Reasoning: Shiwei Guo and Sihang Jiang introduced SAGE to test if models understand the "why" behind cultural differences. They found that even when models possess latent cultural knowledge, they struggle to apply it in real-world scenarios without explicit guidance (Do Large Language Models Truly Understand Cross-cultural Differences?).
- Dimensional Probing: In CultureScope, Sihang Jiang and Yanghua Xiao evaluated models across 140 cultural dimensions, revealing that training on more languages does not automatically increase cultural intelligence.
- Culture-Specific Items: Mohsinul Kabir and colleagues developed XCR-Bench (XCR-Bench: A Multi-Task Benchmark for Evaluating Cultural Reasoning in LLMs), which shows that models perform significantly better at cultural reasoning when the output is generated in the target culture's native language rather than English, further supporting the "translation layer" theory.
The Failure of Regional Fine-Tuning
A common solution to Western bias is to build "sovereign" LLMs—models specifically fine-tuned for a particular region. However, research shows that even these models struggle to escape their initial training.
In Fluent but Culturally Distant, Dhruv Agarwal 0001 and Aditya Vashistha demonstrated that models fine-tuned for the Indian context still reflect American public opinion more closely than local ground realities. Similarly, Elyanah Aco found that models proficient in Filipino and Cebuano still default to Western perspectives on social nuances (FilBench).
Even in high-resource non-Western languages like Chinese, models often fail to recognize localized social biases or regional prejudices, as shown by Tian Lan and Xiangdong Su in the McBE benchmark (McBE: A Multi-task Chinese Bias Evaluation Benchmark for Large Language Models).
Why It Matters
The disconnect between language and culture is not just a technical curiosity; it is a barrier to the safe and equitable deployment of AI. When a model "flattens" cultural nuances, it risks imposing a monocultural value system on a multicultural world. As Daniel Hershcovich and Laura Cabello conclude in their study of ChatGPT (Assessing Cross-Cultural Alignment between ChatGPT and Human Societies), using English prompts often pulls models toward Western norms, effectively silencing the diverse cultural frameworks they were supposedly trained to understand.