TL;DR

Denmark’s education ministry has ordered three measures against AI use in upper secondary schools, effective immediately as pupils return from the summer break. Around 9,000 students a year who write the major SSO assignment at home will now have to defend it orally. Schools must monitor how computers are used during written exams and install firewalls limiting what the network reaches. Third, they are urged to move more written work into supervised classroom time.

A policy that changes what is assessed

The oral defence is the interesting mechanism, because it does not attempt to detect anything. Rather than testing the artefact for machine authorship — which, as the Eisteddfod episode this week demonstrated, is where institutions get into trouble — it tests whether the student can account for the work. A pupil who understands their own argument passes regardless of what tools touched the draft; one who cannot, does not.

Education minister Magnus Heunicke was direct about the trigger, saying upper secondary schools have an AI cheating problem and action is needed now. A broader national strategy for AI across the education sector is being prepared, and Heunicke says he will consult schools, teachers and students to ensure AI does not erode students’ competence, professionalism or capacity for independent thought.

The scale of the problem next door

Sweden provides the number that makes this look proportionate rather than reactive. Suspensions and warnings for unauthorised AI use rose 688% there between 2023 and 2025. The Swedish government has commissioned its own review of how schools are handling it, reporting in April. Upper secondary minister Lotta Edholm’s position in January was that cheating is unacceptable in any form and risks producing grades that do not reflect what a student knows.

Looking forward

For UK readers this arrives as academics warn of a cheating crisis in higher education and as detection tools keep producing contested results. Denmark has picked the opposite approach to detection: change the assessment so that AI use stops being decisive, rather than trying to police it after submission. It costs staff time — 9,000 oral examinations is not free — and it does not scale automatically. But it fails safe, which detection does not, and the person carrying the cost of a wrong call is no longer the student.