A year-group leader at a secondary school gets two questions within an hour on a Wednesday. A Dutch teacher wants to know whether her class may practise writing an argument with a chatbot. A colleague in the office asks whether he may have a letter to parents rewritten by Copilot. The school's AI protocol is about cheating in assessments, and says nothing about either.
An AI policy for a school answers two kinds of questions, and they differ from a company's. For staff: which tools may you use, with which account, and which pupil data never goes in. For pupils: what may you do with AI on an assignment, how do you show it, and what happens if you do not. Two policy lines, joined at one point: the teacher who explains the rules in class also uses AI, and meets the same questions.
Why two policy lines rather than one?
Because the risks differ. With a pupil the question is about learning and assessment: what still counts as their own work, and what someone learns when the chatbot writes the argument. That is a teaching and examination question, and schools have their own instruments for it. With staff the question is about data: a report comment, a care note, the notes of a parent meeting. That is data about minors, often about health or home circumstances, and it sits in exactly the texts someone wants rewritten at the end of the day.
One document that tries to cover both usually turns into a cheating protocol with a paragraph on privacy attached. The teacher finishing a letter at four o'clock finds no answer in it.
What do you agree with staff?
The same questions as in any AI policy, with a school's answers.
| Question | What a usable answer looks like for a school |
|---|---|
| Which tools are allowed? | A list of names with the account attached: Copilot through the school account yes, ChatGPT on a personal account no |
| Which data never? | A pupil's name and class combined with a concern, a diagnosis, a home situation or a grade; a citizen service number always |
| What is fine? | Lesson material, rewriting an assignment, a letter to parents without names, feedback on an anonymised essay |
| What do you do with a tool that has no decision? | Public information only, and here is where you report it |
| Who assesses a new tool? | A role and a deadline, with the data protection officer involved whenever pupil data goes in |
| How do you hear about a change? | Not only in the newsletter; the decision sits where the tool is opened |
Two points are specific to education. First, the data processing agreement: Dutch schools sign processing agreements under the education privacy covenant, and a free AI account has none. A tool without a processing agreement can be on the list, but only without pupil data. Second, the account: a teacher who logs in at home on a personal account falls outside every agreement the school has with a vendor. That is not bad faith, it is the school laptop that was already shut. The policy has to address it, and "use the school account, at home too" works better than a ban. The difference between a free and a business account is, in education, the difference between having an agreement and not.
What do you agree with pupils?
Here the policy answers different questions: is AI allowed on this assignment, how far, and how do you show what you did. Kennisnet, the Dutch public organisation for ICT in education, advises schools to make clear agreements on how pupils and staff use generative AI, and for assessment there are instruments that set a level per assignment, from no AI to writing together with AI.
What is usually missing is the data side. A pupil who asks a chatbot for feedback on a personal story sometimes shares more than they realise: a name, a school, something from home. Pupils are minors, and the terms of most free tools were not written for them. So a pupil agreement also says which tool the school has set up for use in class, and what you do not put in a chatbot about yourself or your classmates.
Where do the two lines meet?
At the teacher. The teacher explains the agreement in class, judges whether a pupil kept to it, and uses AI to prepare the lesson. A school where staff may use any tool and pupils none cannot keep that up. A school with sharp agreements on data for staff and clear agreements on assignments for pupils can stand in front of a class with one story.
There is a second meeting point: the tools themselves. A learning tool the school has used for years gains an AI feature, and that feature then applies to pupils and teachers at once. That belongs in the AI register, not in the cheating protocol.
What does the school not have to arrange?
More than you think. The school does not have to predict every tool; a status of "no decision yet" with an instruction attached covers the tools you do not know yet. The school does not have to write its own risk assessment of every model; the question is which data goes in and under which agreement, not how the model works. And the school does not have to obtain a certificate in AI literacy. Article 4 of the AI Act asks organisations that use AI systems to ensure a sufficient level of AI literacy among their staff, and what is sufficient differs by role. How to meet that without a compulsory course is in AI literacy in 2026.
What the school does have to show is that it decided something per tool, that the decision has a date and a reason, and that staff knew about it. The accountability principle in the GDPR does not ask for a thick document. It asks for a decision you can find again.
How does the agreement reach the teacher?
A board with seven schools decides that Copilot through the school account is allowed and ChatGPT only without pupil data. An inventory six months later finds Gamma, MagicSchool, a transcription tool for parent meetings and two lesson-material tools in use across the schools as well. Three of them appear nowhere. Nobody did anything wrong; nothing had been decided about those tools. It does show that a policy set once a year is, six months on, about different tools from the ones in use.
The decision works when it appears where the tool is opened, not only in the document on SharePoint. BeeSensible records per AI tool what the board has decided and shows that decision when someone opens the tool in the browser, with the approved alternative as the first button. Continuing is always possible; the extension blocks nothing and records no user id. This is what that overview looks like, tools next to their status:
