Why Cito deliberately chooses not to use generative AI for marking exams

Within the Digitalisation of Central Exams programme, Cito is investigating how AI can support teachers in marking central exams. Researcher Wouter Jansen is clear about this: the consequences for exam candidates are too significant to take risks with a system that does not explain how it arrives at its assessment.

That is why Cito deliberately opts for NLP techniques instead of generative AI. Not a language model that proposes a grade without clear justification, but techniques that make it transparent why answers are grouped together.

In practice, this means that the system stacks answers and clusters similar responses, so that teachers can see at a glance which students have answered in a comparable way. Peer feedback becomes easier, and the teacher retains full control over the final judgement.

The aim is not to automate the marking process. The aim is optimal collaboration between teacher and AI: time savings for the teacher, and higher quality marking because similar answers are assessed consistently.

This example is relevant for anyone considering AI in assessment. The temptation to choose the most advanced technology is great, but when much is at stake for students, explainability outweighs language proficiency. Cito demonstrates that a conscious, less flashy choice is often the wiser one.

At EduGPT, we recognise this principle. AI that supports education professionals should empower them, not replace them, and certainly not be opaque at moments when much is at stake for students.

Source: Cito, "Teacher and AI on the road to optimal collaboration"