SESGO:

This study addresses the lack of bias evaluations in large language models (LLMs) for Spanish, especially in Latin American contexts. It proposes a culturally adapted framework that integrates regional expressions and stereotypes across four categories: gender, race, socioeconomic class, and national origin.

Based on a dataset with over 4,000 prompts, it introduces a metric that combines accuracy and error direction to simultaneously evaluate performance and bias in both ambiguous and clear contexts. The results show that mitigation techniques designed for English do not perform as well in Spanish and that bias patterns remain consistent even with different temperature settings.

The framework is flexible, allowing it to be incorporated into new languages, stereotypes, and categories, contributing to fairer and more culturally aware AI evaluations.