Why are all LLMs Obsessed with Japanese Culture? On the Hidden Cultural and Regional Biases of LLMs

2026-04-23 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, quick

Summary

A new study introduces the Culture-Related Open Questions (CROQ) dataset to analyze regional and cultural biases in Large Language Models (LLMs). Contrary to prior research suggesting Western and Anglocentric biases, this work reveals a distinct preference for Japanese culture in LLM responses. The research also indicates that LLMs generate more diverse outputs when prompted in high-resource languages like English, and exhibit less bias towards countries where the input language is official. Furthermore, the study pinpoints the emergence of this cultural bias, finding that it becomes evident after supervised fine-tuning, rather than during the initial pre-training phase of the LLMs.

Key takeaway

For research scientists evaluating LLM fairness and cultural representation, you should specifically investigate biases that emerge post-supervised fine-tuning. Your evaluations must extend beyond Anglocentric biases to detect unexpected regional preferences, such as the observed tendency towards Japanese culture, and consider how input language choice influences output diversity.

Key insights

LLMs exhibit a surprising bias towards Japanese culture, emerging post-supervised fine-tuning.

Principles

• LLM cultural bias is not solely Anglocentric.
• Input language diversity impacts LLM output diversity.

Method

The CROQ dataset, a taxonomy of Culture-Related Open Questions, was used to probe LLM cultural preferences and identify bias emergence points.

In practice

• Use high-resource languages for diverse LLM outputs.
• Evaluate LLMs for post-finetuning cultural biases.

Topics

• LLM Cultural Bias
• Regional Bias
• Culture-Related Open Questions
• Supervised Fine-tuning
• Pre-training

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• Location Not Found: Exposing Implicit Local and Global Biases in Multilingual LLMs — Multilingual LLMs exhibit implicit US-centric and population-driven biases, exacerbated by instruction tuning.
• Extending the Contact Hypothesis: Cross-Linguistic Evaluation of Religion and Nationality Bias When Prompting LLMs in German and Icelandic — LLMs exhibit religion and nationality biases, requiring cross-linguistic evaluation for fairness.
• Bias and Justice in Language Models: Evidence from a Linguistically, Socially, and Culturally Asymmetric Literature — LLM bias research is skewed towards English and binary gender, lacking intersectional and multilingual real-world analyses.
• Korean Culture into LLM Alignment: Toward Cultural Coherence — LLM cultural alignment requires defining culturally coherent responses, not just suppressing harmful ones.
• IndoBias: A Dual Track Culturally Grounded Benchmark for LLMs Bias Evaluation in Indonesian Languages — Culturally-grounded benchmarks are crucial for evaluating LLM bias in diverse, multilingual contexts like Indonesia.
• Location Not Found: Exposing Implicit Local and Global Biases in Multilingual LLMs — Multilingual LLMs exhibit implicit US-centric and demographic biases when answering locale-ambiguous questions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.