Researchers identify cultural biases in large language models

A research team from the University of the Basque Country and Cardiff University has reported a significant tendency for large language models (LLMs) to favor certain countries when asked general cultural questions. In a study involving eight models across 24 languages, researchers found that when no specific region was requested, models frequently defaulted to mentioning Japan and the United States.

The study analyzed over 31,000 questions across 11 fields, including food, arts, and social life. Findings indicated that models often select countries that correspond to the language used in the prompt, with such mentions accounting for 43% to 78% of responses. Even when excluding language-specific biases, Japan was the most frequently mentioned country in five out of eight models tested, particularly in categories like food and arts.

Researchers noted that these biases appear to intensify during the instruction-tuning phase of model development. While English-language prompts yielded a broader range of countries, prompts in languages with less training data resulted in highly concentrated responses. The study highlights that these patterns do not necessarily reflect the accuracy of cultural knowledge or the frequency of a country's appearance in training datasets, but rather an inherent regional bias in model outputs.