Why the old question fails

§01 The situation

What AI can do today

Several studies show that a student with a current AI tool can hand in an answer most examiners can’t tell from a good student’s.

These studies used models from 2023 and 2024, and today’s tools are stronger.

Assumptions about what AI can’t do often hold for only a few months. Norway’s government committee on AI in higher education advises against chasing AI-proof tasks for the same reason: what AI can do is hard to predict, and fewer and fewer tasks will be immune.

§02 A different question

Validity, not cheating

It is tempting to see the problem as cheating. Dawson, Bearman, Dollinger and Boud (2024) suggest another question. Instead of asking whether the student did this themselves, ask whether the assessment gives valid grounds for what the grade claims about the student’s competence.

AI has broken the inference from a good product to a competent student, even for students who follow every rule, and that makes it a validity problem. On its own, an open exam now gives weak evidence of what the student can do. Its strength lies elsewhere. It gives students time to write and to work things through in depth, and it points them towards deep learning.

A sit-down exam is often the opposite. It gives safer evidence, but of a narrow slice of competence and in a situation that looks little like how students will work later. No form of assessment is good at everything, so what do we actually base our inferences on?

Four kinds of evidence

All assessment rests on indirect evidence. A learning outcome describes a broad competence, and an assessment captures only a slice of it. Fawns, Boud and Dawson (2026) distinguish four kinds of evidence, each with strengths and weaknesses.

Validity belongs to the inferences we draw from the evidence, so it isn’t a property of the task itself. If an assessment rests on products alone, the inference about what the student can do is weak, and that is exactly the inference AI has weakened most. The designs that hold up best combine several kinds of evidence.

§03 Rules

Talk or structure?

Most institutions have answered AI with new rules: guidelines for permitted use, AI declarations and scales for how much AI is allowed. Corbin, Dawson and Liu (2025) call these discursive changes. They change what is said about an assessment and leave the way it works as it was.

Structural changes alter what students actually have to do, for example a conversation where the student explains their own work, or a part of the task solved on the spot. They work whether the student follows the rules or not.

In an unsupervised assessment, discursive rules can’t be enforced. The authors call this an enforcement illusion. A study from King’s College London shows how it plays out: 74 % of the students left out their AI use, although the declaration was mandatory (Gonsalves, 2025).

Activity: sort the measures

For each measure, ask one question: does it change what the student has to do?

§04 Detection

Why not detect it?

If the rules can’t be enforced, can we at least detect when they are broken? In short, no, for four reasons.

Detectors have improved, but they lag behind new models. Jisc’s national centre for AI in UK education points out that better detection of AI contributions doesn’t necessarily mean better evidence of misuse. Norway’s committee on AI in higher education advises institutions not to use AI detectors at all: they are unreliable, their results can’t be checked, and they weaken students’ legal protection.

Activity: the detector’s dilemma

Here are 300 exam texts. Some were handed over to AI, the rest the students wrote themselves. Set how good the detector is and see who gets flagged.

§05 What follows

Two lanes

When rules can’t be enforced and detection doesn’t work, there is one consequence left to draw. The University of Sydney has drawn it by sorting all assessments into two lanes (Liu and Bridgeman, 2023).

A take-home exam with an AI ban is still in the open lane, because nobody can see whether the ban holds. So there is no restricted lane without supervision. In the secured lane, you can still decide that AI is excluded, limited or allowed.

Sydney also counts a live conversation about submitted work as secured. A take-home task followed by a conversation is then a pair, with one open and one secured part. Step 4 comes back to this.

None of this makes the open lane second-rate. It is where students get time to go deep, and where they can learn to use AI in a way their field would accept.

Activity: which lane?

A critique, and an answer

The model is contested. Curtis (2025) argues that students need a third lane with stated limits on AI use, partly because novices have to learn to write on their own. The concern is real, but a restricted lane without supervision leads straight back to the enforcement illusion.

The more realistic answer is to combine the lanes. If writing on your own is a learning outcome, it has to show where it can be observed, for example in a writing session or a conversation about the student’s own text, and then the open work can be truly open. The novice’s problem isn’t only about rules, either. Students have to learn the subject and how to work with AI in it, critically, at the same time. That is where Step 2 begins.