How I Use AI in Assessment — and What I Never Let It Decide

Assessment is one of the heaviest parts of a teacher's job. Not because it's hard, but because there's so much of it: thirty students' assignments, feedback and grades in the same week, on top of everything else.

When I started using AI to support assessment, I made one rule right at the start: I decide the assessment. AI doesn't give grades, doesn't pass or fail anyone, and doesn't draw conclusions about students. The same rule applies to all my tools, and it's the one point where I don't bend.

With that rule in place, AI became genuinely useful in assessment. Here's what I use it for in my own work.

Drafting feedback

Feedback is the most laborious part of assessment. I know what to say. But verbalizing the same things for thirty students takes an evening.

I give the AI my assessment criteria and my own notes on the student's work. It drafts the feedback; I check and edit. The criteria always go in with the prompt — I don't let a language model invent its own idea of good work.

The feedback didn't improve because a machine writes better than I do. It improved because there's now actually time to give it.

Criteria in the student's language

Competence requirements and assessment criteria are written in administrative language. Students don't always make sense of them — and on first reading, neither do I.

I ask the AI to say the same thing in the student's language: what you actually need to be able to do here, and what an accepted performance looks like. I check that the translation matches the original. It's faster than writing it myself — but unchecked, it can't be used.

Differentiation and accessibility

As a special education teacher, I consider this the most important use of AI in assessment. The same assignment at different difficulty levels, the same instructions in plain language, the same competence demonstrated in a different way.

This work used to go undone because there was no time. Now it goes undone less often.

What the AI doesn't get

Student personal data doesn't go into AI tools. Feedback and drafts are made without names or identifiers. My rule of thumb is simple: if the text could be pinned to a bulletin board, it can go to an AI. If it couldn't, it can't.

The other principle is transparency. I tell my students how I use AI in assessment. I expect the same transparency from them.

Try your own assignments on a language model

One thing surprised me. When I tried my own assignments on a language model, a large share of them were solved in seconds. A task that measures recall and repetition no longer measures the student's competence — it measures whether the student has access to a language model.

The design framework didn't change, though. I use the same tools as before: Bloom's taxonomy and constructive alignment, where goals, teaching and assessment tell the same story. AI just made visible how many of my tasks lived on the lowest levels of thinking. And the same tool that exposed the problem helps fix it: lifting tasks toward analysis and justification goes faster with it than alone.

Where to start

If you want to try AI as an assessment aid, four steps are enough to begin:

1. Decide the line first. What can AI take part in, and what does only the teacher decide? Write it down before the first experiment.
2. Include the criteria. Feedback is based on your criteria, not the language model's guess.
3. Keep personal data out. The bulletin board rule works well here.
4. Try your own assignments on a language model. You'll quickly find out which of them still measure anything.

What I've learned

AI does not assess the student. It helps the teacher assess.

The benefit didn't come from a machine knowing how to assess. It came from time: verbalizing and drafting got faster, and the freed-up time goes to the part of assessment a machine cannot do — understanding the student's situation and making the decision.

That part wasn't a problem before AI, and it isn't a problem now. It's the teacher's job.