AI in Danish schools 2026: rules and what teachers may use

25. That is how many Danish public schools sent at least one kid home from an exam for using AI this summer (Folkeskolen, 3 September 2026). Behind it sit the rules for AI in Danish schools, four documents: seven recommendations (the Danish Agency for Education and Quality, STUK, 26 June 2025), a mini-guide on lawful use (the Danish Agency for IT and Learning, STIL, 6 January 2026), the exam order that bans generative AI at the exit exams, and an emergency package against AI cheating in upper secondary schools (the Danish Ministry of Children and Education, 6 August 2026). The short version: a teacher with a chatbot is one problem. A kid with a chatbot is another.

Seven recommendations, and not one of them is a rule

On 26 June 2025 STUK published seven recommendations on generative AI in public-school teaching. They are advice, not law. The full document says there are no national rules for generative AI in public-school teaching, and that under the Danish Public Schools Act (section 44, subsection 4) the school board at each school approves teaching materials.

The seven points: a shared direction; GDPR-compliant solutions; dialogue among staff; staff competence; AI only where it makes pedagogical sense; talk with kids about how, when and why; tell parents.

Two lines matter more than the list. It repeats the exam expert group's advice that AI should feature significantly less in the youngest classes (grades 0 to 3) than in the oldest classes (grades 7 to 9). And it says kids must be able to sit their final exams without generative AI, so the skills still have to be taught. Mattias Tesfaye, minister at the time, put it in one line: "It must never become a cheating tool, but it may be a learning tool."

The mini-guide says what is legal, not what is wise

On 6 January 2026 STIL published a mini-guide on lawful use of AI for educational institutions, written with STUK and the LISSI network of information-security leads. The full guidance of 28 May 2025 was for lawyers; this one is for teachers and school leaders, a seven-step checklist to run before an AI tool enters a classroom: map the tool and its data, establish the school's role, assess lawfulness under GDPR and the EU AI Act, assess the risk level, inform kids, parents and staff, set policies, train staff.

Three parts are usable on a Monday. First, adopting an AI tool is a leadership decision, not one teacher's. Second, it lists what is high risk under the AI Act in education: systems that monitor kids' computers for cheating, assess learning outcomes, decide on access to education or steer a kid's learning process. Third, its first worked case (AI feedback on kids' essays) concludes it can be done if a data processing agreement rules out model training, no sensitive data is entered, kids and parents are informed, and the teacher remains the final assessor. Exam scripts stay off limits.

AI in Danish schools: the teacher's laptop and the kid's laptop are different problems

When I taught in a Danish public school, preparation happened at a kitchen table after 20:00 with a laptop and last year's copies. A teacher who now uses AI at that table to draft a differentiated text or ten practice questions is using an ordinary work tool. The mini-guide's advice for that side: work from a list of approved tools, never enter personal data, exam answers or grades, treat the output as a draft, keep the judgement with the teacher. No kid has touched AI.

The kid's laptop is the other question. Once a 12-year-old logs in, the school is processing a child's personal data, a data processing agreement is normally required (STUK, June 2025) and the AI Act's transparency rules apply. Kids may use generative AI in lessons if the school board has approved it. The exam rule does not move: under the exam order (the order on public-school exit exams, section 43, subsection 2) pupils may not use generative AI at the public-school written or oral exit exams, including any preparation time that counts towards the mark.

Denmark's rule of thumb is simple: AI may help a teacher teach and a kid learn, but it may not do the work that is being marked.

This summer the exam ban was enforced for the first time

On 7 August 2026 Folkeskolen reported unpublished figures from the ministry's school panel survey, shown to the minister at the Sorø meeting: 44 per cent of kids in grades 6 and 8 use a chatbot for help with schoolwork, and 23 per cent use it to get the answer instead of solving the task. A kid in grade 6 who asks the chatbot to explain fractions again. A kid in grade 8 who pastes the whole assignment in at 21:30. A kid who does both and cannot tell the difference.

On 3 September 2026 Folkeskolen published its own survey. It asked 1,067 schools with lower secondary classes; 203 answered. Of those, 25 had sent at least one kid home from a summer exam for AI use, and 25 school leaders had been contacted by an external examiner who suspected AI in a written paper. Nineteen per cent is a thin response rate, and schools with a problem were probably more likely to answer. It is what we have. We talk about this as a cheating problem, a phone problem, a ChatGPT problem. It is an assessment problem: the kids did not change, the tool on the desk did, and the exam formats mostly did not.

The government's first response (6 August 2026) is aimed at upper secondary schools, not Danish public schools. The emergency package against AI cheating has three parts: the roughly 9,000 HF pupils who write the larger written assignment each year must defend it orally; upper secondary schools are encouraged to use screen monitoring and firewalls at written exams; more written work is to be done at school. Education minister Magnus Heunicke said: "There is a need for action now, and we are starting with these three initiatives." I think the oral defence is the right tool and screen monitoring is the wrong one, and I would not want it near a 12-year-old. The same announcement promises a national strategy for AI across the school and education sector. No date is given, and as of 18 September 2026 none has been published.

What a school can do this term

1. Put the decision on the leadership agenda. Write down which AI tools are approved, for whom (staff only, or kids too) and for which tasks. The mini-guide's template is three lines: tool A for general text, no personal data; tool B for anonymised data under a data processing agreement; tool C not for kids' work or grades. About 45 minutes in one leadership meeting.

2. Ask every vendor for the data processing agreement first. The mini-guide says a tool cannot be risk-assessed without one. One email per vendor, about 10 minutes; if nothing comes back, the checklist stops there.

3. Run the seven steps on each tool kids log into. If it assesses learning outcomes or steers a kid's learning path, it may be high risk under the AI Act. Two to three hours per tool for whoever owns GDPR, once.

4. Say the exam rule out loud in every subject at the start of term: what is allowed in lessons, and that nothing generative is allowed on exam day. Five minutes per class, plus one marked piece of writing per term done in class.

5. Send parents one written note a year on the school's approach. About 30 minutes, then it goes out on the school's parent platform (Aula) with the autumn letters.

6. Give staff one afternoon: prompts, what must never be typed into a tool, how to check output. Step 7 of the mini-guide calls this a requirement. Three hours once, then 15 minutes a term.

Where GradeAid fits

GradeAid is a Danish platform for teachers. The teacher types in the topic and the level and gets a lesson plan, a worksheet, a quiz, a presentation and screen-free activities in seconds, which you can also differentiate. The material can be printed. That puts it on the teacher side of the Danish line, but it does not do the school's checklist: ask us for the data processing agreement first, as you would anyone else.

If you work with AI in a Danish or Nordic school, we would like to hear from you. Write to us at founders@gradeaid.ai.