2026 ASEE Annual Conference & Exposition

The integration of Artificial Intelligence (AI) as a standardized assessment tool in engineering education poses critical challenges related to algorithmic fairness. This study investigated biases related to gender, indigenous heritage, and socioeconomic level in an AI system used for grading tasks in an Information Systems course with 55 undergraduate students of Industrial Engineering at a higher education institution.
The tool was built by fine-tuning an open-source language model (Mixtral-8x7B), which was run locally to ensure student data privacy. Through an anonymous self-identification form, it was identified that 11% of students (n = 6) identified themselves as belonging to indigenous peoples, a proportion close to the national average of 10% in higher education in the region. The methodological design employed a mixed-methods approach, combining statistical analysis of grades (ANOVA and t-tests) with qualitative analysis of student perceptions through semi-structured surveys (n = 42 responses).
The quantitative results revealed significant disparities in the performance of the AI system. Female students received, on average, grades 12% lower than their male peers (p = 0.032, d = 0.48). Students belonging to indigenous peoples received significantly less detailed feedback and less alignment with the established rubric (p = 0.041, d = 0.53). In addition, students who worked more than 20 hours per week, an indicator of socioeconomic vulnerability, received more generic comments with less technical specificity (p = 0.038).
The qualitative analysis complemented these findings, with 68% of female students and 83% of Indigenous students expressing distrust in the objectivity of AI. In comparison, 45% of all participants reported concerns about cultural biases in automated assessment. The identified patterns suggest that the model reproduces the intersectional inequities present in the training data.
These findings show that, without rigorous pre-implementation auditing, AI tools can perpetuate and amplify systemic inequities in engineering education. As a limitation, the study is limited to a single course in a specific context; however, the observed patterns are consistent with the international literature on algorithmic bias.
Practical implications include: (1) the need for mandatory equity audit protocols before implementing AI systems in educational evaluation, (2) the importance of culturally diverse and representative training datasets, and (3) the recommendation to maintain human oversight in high-impact evaluation processes. This research provides empirical evidence from Latin America on algorithmic justice in contexts of ethnic diversity, responding to the global need to develop more equitable AI systems in engineering education.