Why are some AI models obsessed with "Japanese culture"? European team presents research on the reality that outputs from "4o-mini" and others are skewed toward Japan: Innovative Tech

A European-led study found that several large language models skew culture-related answers toward Japan and the U.S.

When the user’s country was removed from prompts, models still tended to default to those two regions.

The researchers tested eight LLMs with a custom multilingual set and observed the pattern across systems including GPT-4o-mini and Gemini 2.5 Flash.

They also found the bias became stronger after supervised fine-tuning, raising concerns about hidden geographic bias in multilingual AI.