Control or learning?

§01 The case for control

What the committee wants to protect

The note is explicitly preliminary (p. 2), and the committee’s final report is due in November 2026. Its starting point is what it calls the certification function. Universities confirm competence through degrees and rank candidates through grades, and good certification helps share out work and further study fairly and efficiently (p. 6). Two working hypotheses frame the rest. The committee assumes that AI systems will solve written tasks as well as students can be expected to, or better, and that they will match most teachers at planning teaching, explaining the syllabus, assessing answers and giving feedback (p. 7).

Seen from there, cheating is only part of the problem. Norwegian law requires that students are assessed in an impartial and academically sound way. If the examiner can’t separate the candidate’s contribution from the AI system’s, they can’t judge what the candidate knows, even when every rule was followed (pp. 18–19). Better rules won’t close that gap in the committee’s view, since rules on permitted aids can only be checked when an exam is supervised (pp. 19, 31). Institutions should still keep developing their AI guidelines to make them clearer and more predictable, the committee adds, though that won’t solve the problem (p. 3).

The case is made for students too. Without supervision, some get results they didn’t earn, and a student who uses AI exactly as allowed can still be suspected of cheating, which weakens their legal protection (p. 21). Controlled exams, the committee argues, give students more accurate and fairer assessment and better legal protection (p. 3).

§02 Its own premises

What the committee says students must learn

Before the exam chapter, the note makes a strong claim. Knowledge of AI and skill in using and developing it are a condition for society to benefit from the technology and meet its challenges, and higher education has an important role in building that competence (p. 9). The committee describes four kinds of AI competence (general, subject-specific, specialised and enabling) and stresses that universities should help students learn to use generative AI critically and in ways that support their learning (pp. 10–11). The note also warns that uncritical use leads to surface learning, and that central outcomes such as critical thinking need concentrated work on problems over time (pp. 16–17).

In the exam chapter the committee is open about what control can’t do. A supervised sit-down exam (skoleeksamen in Norwegian) is less suited to testing whether students can build well-structured arguments or plan and carry out their own investigations (p. 22). The main argument for open exam forms, it writes, is that independent work on larger tasks over time is needed to train essential skills (p. 23).

It would be a great loss to the value of higher education if such exam forms could not be used.

The committee also sees room for AI inside control. AI tools can be allowed in a controlled exam, and in several cases that is desirable. The candidate’s account and usage log can then let the examiner judge how the tool was used (p. 23). Later in the note it asks how to assess hybrid intelligence, meaning what a student and an AI system achieve together, and says it will keep working on that question (pp. 31–32).

Our addition. Followed through, these premises point somewhere the first recommendation doesn’t go. If AI competence is a condition, higher education has to build it. If assessment should be aligned with the learning outcomes, as the committee itself holds (p. 22), then the summative assessment should be able to show that competence as well. More controlled exams leave less room for students to practise and show their own subject-specific use of AI over weeks of work.

§03 The price

What a controlled part can cost

The note presents the controlled part as a way to keep everything. In our paraphrase, it lets students be tested in every kind of learning outcome without weakening the integrity of the assessment (pp. 23–24). The committee does name costs. A controlled part needs more resources, and Norwegian law fits combined written and oral exams badly. The biggest snag is that an oral performance, which by its nature can’t be checked afterwards, can’t be appealed (p. 24). Corbin, Bearman, Boud and Dawson (2025) see a wider price. In their study, moves towards security gave up something else, such as authenticity or creativity (p. 7).

The open work as raw material

The committee wants the controlled part to carry real weight in the grade and to match the open part in content (p. 23).

Our addition. If the controlled part is where the grade is really decided, the open work can shrink into raw material for the check. Writing over time, which the committee values (p. 23), keeps its worth only while the open work is assessed as the main work.

Who carries the cost

When the note weighs controlled exams, it counts the cost of rooms, equipment and invigilators (p. 23). It notes that surveillance of take-home exams can strain students and hit them unequally (p. 27), but it doesn’t ask whether a shift to sit-down exams would hit some students harder. Dawson, Bearman, Dollinger and Boud (2024) treat assessment that isn’t inclusive as a threat to validity (p. 1008), and they warn that measures against cheating can make validity worse by creating problems for inclusion (p. 1005).

A direction Norway chose before AI

Norway’s 2007 evaluation of its Quality Reform found that assessment had grown more varied, with a strong rise in assignments and tests during the semester and in portfolio assessment, while the traditional final exam was largely kept (Michelsen and Aamodt, 2007, pp. 9, 25). The evaluators argued that more frequent writing with comments from teachers gave better learning than the more traditional forms (p. 8). They also asked whether institutions would return to the traditional sit-down exam when study plans were revised (pp. 26–27). The committee’s note doesn’t refer to that history. Its reference list cites several studies of AI detection and cheating, but none of the TEQSA resources or the work by Dawson, Bearman and Boud (pp. 48–51).

Our addition. More controlled exams would turn part of that direction back. The earlier move was made for the sake of learning, long before generative AI, and reversing it deserves an open account of what is given up.

§04 Rules and the open lane

Where students learn to use AI

On rules, the committee and the assessment research are closer than they may seem. The committee accepts that rules on aids can only be checked under supervision (p. 31), and Dawson and colleagues argue that institutions “should not set rules that will likely be broken” (2024, p. 1011).

Our addition. A rule about AI use in a task nobody supervises can’t work as a security measure, since no one is there to enforce it, though it can still guide responsible use. Students’ real AI use is also hybrid, moving between their own thinking and the tool within one piece of work, as Step 2 described. Open formats are therefore where students can learn to use AI critically over time, with room to revise and get feedback. When more of the grade moves into controlled exams, that learning needs a planned place elsewhere in the programme.

§05 Your course

What you can do whatever the final report says

Our addition. The advice in this section is ours.