We were in New Delhi participating in the India AI Impact Summit 2026, an international gathering that brought together researchers and organizations to discuss the social, political, and educational impact of artificial intelligence.
On February 18, as part of the session “Research Symposium on AI and its Impact,” Guido Ivetta represented our team and presented the poster: “HESEIA: A Community-Based Dataset to Evaluate Social Biases in Large Language Models, Co-Designed in Real School Contexts in Latin America.”
A Dataset Built from the Classroom
HESEIA emerged from a teacher extension course in which a dataset of 46,499 sentences was collaboratively developed. The dataset captures intersectional biases across multiple demographic axes and school subjects, reflecting local contexts through the lived experiences and pedagogical practices of educators from Latin America.
Unlike other datasets designed exclusively in technical environments, HESEIA was co-designed in real school settings, incorporating the situated perspectives of teachers. Its goal is to strengthen bias evaluations in large language models from a community-based and context-aware perspective.
HESEIA is part of the work carried out by the AI team at Fundación Vía Libre: Guido Ivetta, Marcos Javier Gómez, Sofía Martinelli, Pietro Palombini, Emilia Echeveste, Nair Carolina Mazzeo, Beatriz Busaniche, and Luciana Benotti.
The paper was published by the Association for Computational Linguistics and is available at the following link:
🔗 https://aclanthology.org/2025.emnlp-main.1275/
We continue working to advance AI evaluation frameworks that integrate pedagogical knowledge, local perspectives, and critical approaches to social bias in large language models.