# Investigating Cultural Alignment of Large Language Models

**Authors:** Badr AlKhamissi, Muhammad ElNokrashy, Mai AlKhamissi, Mona Diab  
**arXiv:** [2402.13231](https://arxiv.org/abs/2402.13231)  
**Date:** 2024-02-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.13231v1/

## Key points

- The study explores the relationship between language and culture in linguistic anthropology, focusing on Large Language Models (LLMs) as repositories of collective human knowledge.
- It addresses whether LLMs truly capture diverse knowledge adopted by different cultures, finding that cultural alignment is greater when models are prompted with the dominant language of a specific culture and pretrained with a refined mixture of languages used by that culture.
- Cultural alignment is quantified through simulated sociological surveys comparing model responses to actual survey participants, revealing misalignment for underrepresented personas and culturally sensitive topics.
- Anthropological Prompting is introduced as a method to enhance cultural alignment in LLMs using anthropological reasoning.
- Limitations include only considering two languages and data from two countries, suggesting future work could expand to include additional cultures and languages for broader support.
- The need for a more balanced multilingual pretraining dataset to better represent human diversity and cultural plurality in LLMs is highlighted.
- Ethical implications regarding understanding model behavior due to their black box nature are discussed, emphasizing collaboration between computer scientists and social scientists to uncover biases ethically.

## Summary

In this study, the intricate relationship between language and culture is explored within the realm of linguistic anthropology. The focus is on Large Language Models (LLMs) as repositories of collective human knowledge. The central question addressed is whether these models truly encapsulate the diverse knowledge adopted by different cultures. The research reveals that LLMs demonstrate greater cultural alignment when prompted with the dominant language of a specific culture and when pretrained with a refined mixture of languages used by that culture. Cultural alignment is quantified through simulated sociological surveys, comparing model responses to those of actual survey participants. By replicating a survey conducted in various regions of Egypt and the United States, LLMs are prompted with different pretraining data mixtures in Arabic and English to align with real respondents' personas and survey questions. It is found that misalignment becomes more pronounced for underrepresented personas and culturally sensitive topics probing social values. The study introduces Anthropological Prompting as a novel method leveraging anthropological reasoning to enhance cultural alignment in LLMs. However, limitations are acknowledged, such as only considering two languages and data from two countries to keep the analysis manageable. Future work could expand to include additional cultures and languages for broader support. Furthermore, the study highlights the need for a more balanced multilingual pretraining dataset to better represent human diversity and cultural plurality in LLMs. Ethical implications regarding model behavior understanding due to their black box nature are also discussed. In conclusion, the research emphasizes collaboration between computer scientists and social scientists to uncover biases in LLMs ethically. By striving for cultural alignment in AI systems like LLMs, researchers aim to improve people's lives while avoiding harm or misrepresentation of cultural values. This collaborative approach is seen as essential for advancing artificial intelligence ethically while mimicking human language and cultural understanding effectively.