Study Reveals Major LLMs Are Culturally Biased Towards Japan Due to Alignment

i_dg23 · x · 2026-07-21

A landmark paper by Cardiff and Basque researchers tested 31,680 cultural prompts across 24 languages on frontier models. The findings reveal that six out of eight major LLMs, including ChatGPT, Claude, and Gemini, most frequently reference Japan when asked open-ended cultural questions.

Surprisingly, researchers tracked this bias and found it does not stem from raw pre-training internet data. Instead, the obsession with Japanese culture emerges during the post-training phase, specifically through supervised fine-tuning (SFT) and alignment processes.

Original post →

More from Research

Research channel →