Large language models have a hidden obsession with Japanese culture that is accidentally injected during the final stage of training.
April 24, 2026
Original Paper
Why are all LLMs Obsessed with Japanese Culture? On the Hidden Cultural and Regional Biases of LLMs
arXiv · 2604.21751
The Takeaway
This regional bias is not present in the initial pre-training data but appears during supervised fine-tuning. The models exhibit a distinct preference for Japanese aesthetics and cultural references even when they are not prompted for them. This reveals how easily human preferences from a small set of trainers can warp the personality of a global AI system. We often expect bias to lean toward Western or English-speaking cultures, making this specific drift toward Japan highly unexpected. It highlights a critical need to monitor what values are being taught by human annotators during the fine-tuning process. Even small teams of trainers can have a massive impact on the cultural output of an AI.
From the abstract
LLMs have been showing limitations when it comes to cultural coverage and competence, and in some cases show regional biases such as amplifying Western and Anglocentric viewpoints. While there have been works analysing the cultural capabilities of LLMs, there has not been specific work on highlighting LLM regional preferences when it comes to cultural-related questions. In this work, we propose a new dataset based on a comprehensive taxonomy of Culture-Related Open Questions (CROQ). The results