AI-assisted marking: plausible marks, but comments that identify three to five times fewer errors in reasoning than teachers do

Out of 3,041 responses from computer science students, the feedback on three business LLMs remains consistently positive and their sub-ratings are virtually identical, despite the appearance of a detailed analysis.

The major language models are increasingly used to grade and comment on papers in higher education. Their structured outputs (marks by criterion, comments, areas for improvement) give the impression of a thorough evaluation. Xi Zhao, Xinyue Jiao and Zhen Xu tried to find out whether this impression corresponds to reality, based on the JorGPT game: 3,041 student responses to 50 open-ended computer science questions, graded both by teachers and by three commercial LLMs.

Three systematic deviations

First, the scores by sub-criterion are so correlated with each other (r from 0.82 to 0.99) that they hardly provide any distinct information. Second, written comments detect students' misconceptions in only 5 to 7% of cases, compared to 15 to 31% for teachers: they function more as a list of points to be covered than as a diagnosis. Finally, their tone remains uniformly benevolent regardless of the quality of the answer, without the modulated severity of human markers.

The reliability of the score also varies according to the field, with procedural topics being the best evaluated. The authors draw from it benchmarks on what can be delegated to AI and what always requires a human eye.