Grading an essay-based examination involves tasks that are difficult to scale: reading carefully, interpreting arguments, applying criteria consistently, and providing feedback that is useful to the student. As the number of responses increases, as occurs in massive courses and some university degree programmes, this task competes with other teaching responsibilities and becomes vulnerable to fatigue, variations in judgment, and lack of time.
The article by Adrián Reina, Manuel Bermúdez, Enrique García-Salcines, Juan Alfonso Lara, and Cristóbal Romero addresses this problem through Exam Grader, a web application that uses generative artificial intelligence to assist in evaluating open-ended responses according to rubrics defined by teachers. The tool does more than produce a grade: it allows users to create and distribute examinations, collect responses, generate comments, compare AI and teacher evaluations, and review each result before communicating it to students.
The validation was conducted using 91 examinations from an online Introduction to Practical Philosophy course, and its findings offer a less reassuring picture than the simple promise of automation. The grades generated by the AI showed a strong correlation with those assigned by the teacher, indicating that both evaluators tended to rank student performance in a similar way.
However, following the same general trend does not mean assigning the same score: on average, the AI graded students 2.21 points lower than the teacher, and around 15% of the evaluations differed by more than three points. Variations were also observed among the system’s flexible, moderate, and strict modes. The results show that a tool can identify certain assessment patterns while simultaneously introducing a bias significant enough to make delegating the final decision to it unacceptable.
This is where the study makes its most important contribution. Exam Grader is designed around human oversight: the AI makes suggestions, performs comparisons, and helps identify possible discrepancies, while the teacher retains the ability to modify the grade and comments. This division of responsibilities allows automation to be understood not as a replacement for teacher judgment, but as a way of making that judgment more explicit through rubrics, agreement analyses, and review processes.
The evidence still comes from a limited sample and a single disciplinary field, so further validation, calibration mechanisms, and more transparent explanations for each suggested grade will be needed. Even so, the study places the debate on productive ground: before asking how many examinations an AI can grade, we should determine which decisions it can assist with, which it must justify, and which must remain under the teacher’s non-delegable responsibility.
---
How to Cite: Reina, A., Bermúdez, M., García-Salcines, E., Lara, J. A., & Romero, C. (2026). Web application for automated correction assistance in written development exams using generative AI. RIED-Revista Iberoamericana de Educación a Distancia, 29(2), 369–392. https://doi.org/10.5944/ried.47116
