What the open lane is for

§01 Why keep it

Time to think

Take-home exams, assignments and projects give students time to write and to work things through in depth. That pushes them towards deep learning. It is also only here that students can learn to use AI well in a longer piece of work, judging and checking what it gives them as they go.

A sit-down exam can test whether a student writes a good prompt on the spot, but it can hardly show whether they use AI well across weeks of work. Even the Norwegian committee that recommends more controlled exams writes that we learn to think largely by learning to write (Malthe-Sørenssen committee, 2025, p. 23, our translation).

So the open lane is worth keeping, although the thinking it was meant to force is now up to the student.

§02 A shift

From cognitive necessity to cognitive choice

Matthew Montebello, at the University of Malta’s Department of AI, describes what has changed (Montebello, 2026). Assessment always did more than measure. An essay forced students to read, organise and argue, and a problem set forced them to understand, so even students who took shortcuts had to think. Montebello calls this productive friction. A task’s cognitive necessity, in Montebello’s terms, is the degree to which it structurally requires understanding, analysis and evaluation.

Generative AI changes the order of the work. A student can now start from a finished, coherent product and only then choose whether to do the drafting, structuring and arguing. Thinking that used to be forced becomes voluntary, which is the move from cognitive necessity to cognitive choice. Students can still go deep, but nothing requires it any more.

The gap between performance and learning

Montebello calls the distance between completing an assessment and actually learning the cognitive engagement gap. Learning doesn’t stop, but the result no longer shows whether it happened. The article points to three consequences.

The article’s answer is a set of five principles called cognitive engagement-centred assessment (CECA): say which thinking you assess, make the process visible, design so the assessment stays meaningful whether AI is used or not, assess how students manage their own AI use, and assure learning through design and dialogue rather than surveillance. The framework is conceptual and still untested, as the article itself says.

§03 Hybrid practice

Learning two things at once

In practice, students move back and forth between their own thinking and the tool, often within the same paragraph. Their work with AI is hybrid. Sorting tasks into “do it yourself” and “let AI do it” doesn’t describe that well, and any fixed line goes out of date as the tools change.

Students need to learn the subject itself and, alongside it, how to work with AI in that subject, with a critical eye for where it fails. The two depend on each other. Judging what AI gives you takes knowledge of the field, because only someone who knows it can see when AI is wrong (Bearman et al., 2024). And knowing a field now includes knowing how AI handles it: where it helps and simplifies, and where it sounds right but isn’t.

AI can play different roles in the same task, and each can be the right one at a certain point.

The order isn’t fixed, and no role is right everywhere. What has to run through all of them is that the student develops their own understanding.

In a field experiment with high school students in Turkey, practice grades in mathematics rose 48 % with ChatGPT, but on the exam without AI those students scored 17 % lower than students who never had access. With a version built as a tutor that gave hints instead of answers, the negative effect largely disappeared (Bastani et al., 2025). So how AI is used matters more than whether it is used. In a study from Aalto University, AI made people better at the tasks and, at the same time, worse at judging their own performance (Fernandes et al., 2026).

That is why we put metacognition before rules: students thinking about their own thinking before, during and after they use AI.

Activity: two strands

Follow one student through the exam question from Start, one stage at a time. For each stage you choose the role AI plays and watch two strands grow or thin out: their grasp of the subject, and their ability to work with AI critically. Here the help role is split in two, AI as a coach and AI as a sparring partner, because the two teach different things.

Keep the productive struggle

Research on productive failure shows that students often understand better when they struggle with a problem first and get guidance afterwards (Kapur, 2008). AI can remove that struggle in seconds. Montebello suggests designs that keep it, and that turn AI from a shortcut into something to examine:

• timed work on a problem before AI is allowed
• comparing an AI answer with the student’s own
• finding, criticising and fixing the errors in a flawed AI answer
• asking students to explain which AI suggestions they kept, changed and discarded, and why

So make AI use, and the judgement of it, part of what is assessed. Ask for choices between alternatives, with reasons and checked sources, and have the student explain what role AI played and why that was sound. A student who set up a search agent, for instance, can hand in the instructions along with a test of what the agent found and missed.