Why Are All LLMs Obsessed with Japanese Culture? (arxiv.org)

Recent research highlights a surprising trend in large language models (LLMs) exhibiting a significant bias towards Japanese culture. This study introduces a novel dataset centered on Culture-Related Open Questions (CROQ) to explore the regional preferences of LLMs when responding to cultural inquiries. Findings reveal that LLMs, especially when prompted in high-resource languages like English, gravitate towards Japanese contexts, diverging from an earlier understanding that these models primarily reflect Western viewpoints. This tendency raises questions about the cultural competence of LLMs and suggests a more complex interplay of biases in AI systems. The significance of this research lies in addressing the under-explored aspect of cultural biases within LLMs, particularly the emergence of these biases post-supervised fine-tuning rather than during initial pre-training phases. By mapping out how LLMs engage with diverse cultural elements and the language in which they are prompted, the work calls for a reevaluation of training data and methodologies in AI development. The results could inform future enhancements in LLMs to produce more balanced and inclusive outputs, ultimately advancing the AI/ML community's understanding of cultural representation in artificial intelligence.