Copies of handwritten math corrected by GPT-5.1: Exercises deemed excellent, an effect on examination still impossible to isolate

Tested with 579 Belgian, English and Dutch students, automatic return produces very well rated exercises, but often too vague diagnoses, especially for the weakest copies.

A team from the Vrije Universiteit Brussel had GPT-5.1 analyse the handwritten answers and grades of mathematics students at the university, to generate personalised feedback and training exercises. The system accompanied six optional weekly tests in an English-speaking class (343 students) and a Dutch-speaking class (236); two of them followed a cross-protocol where only some of the students received feedback.

What the correctors think

Six reviewers scored 157 messages according to six criteria. The proposed exercises were rated as very good in 95.5% of cases, but the accuracy of the diagnosis (41.1%) and the clarity of the tone (33.4%) were much less likely to receive the maximum mark. For the weakest papers, the feedback was less usable: these students would need more targeted support from the teachers.

In terms of results, comparisons on related exam questions show no significant difference after statistical correction. The most diligent students who are most exposed to feedback do better, without separating the effect of feedback from that of attendance and training. The authors argue for a hybrid model, where the teacher guides and AI personalizes on a large scale.