Teachers have been using AI for a while now. For a school board the question is rarely whether, but which tool, under what conditions and for whom. And it tends to arrive at the wrong moment: a teacher has found something, a vendor has made an offer, and the board has to say yes or no in a single meeting.
The conversation then turns to the tool. Is it good, is it handy, what do the teachers think? But the board isn't signing for the tool. It's signing for what happens with the data of students and colleagues, for what reaches a classroom under the school's name, and for being able to explain all that when a parent asks. A demo doesn't show any of it. It's in a document, or it's nowhere.
Why the answer so often goes missing
It's rarely unwillingness. A teacher can judge whether a tool is useful for a lesson. An IT coordinator can judge whether it runs properly. What falls between the two is the question of who carries the risk and where that is written down. Without a fixed set of questions, everyone answers the one that fits their own role, and the rest stays unanswered.
Kennisnet, the Dutch public knowledge centre for ICT in education, puts it plainly for AI in schools: before you can deploy AI applications, your information security and privacy have to be in order. That is a condition up front, not a task for after the purchase.
Six questions, one per layer
VergeLabs assesses every AI system with TRACES.
TRACES are the six layers on which you assess an AI system: Trail, Retrieval, Approval, Constraints, Expertise, Storage. Where a layer is missing, there is a risk.
Below, each layer has a question the board can put to the vendor in writing, and a document the school itself needs to have. The first half is for the vendor. The second half is the board's own homework, and it's the part most often forgotten.
T · Trail. What does the tool record of what teachers do with it, and can the school see it? The school's side of that is a question: who can read those logs, and when? A trail nobody opens helps nobody.
R · Retrieval. Which sources does the tool answer from, and what does it do when the answer isn't in them? A teacher having lesson material generated wants to be able to check a reference. The school's side is an agreement on what a teacher does with an answer that has no source: don't use it until it has been checked.
A · Approval. Where does a person confirm the content before it reaches a student? In a school that is often the teacher, and that's fine, as long as it's agreed. Without an agreement, everyone assumes someone else checked.
C · Constraints. What can the tool technically not do, even if a teacher asks? Think of processing data outside the agreed environment. The school's side is central account management: Kennisnet advises managing accounts centrally, so that staff or students don't create accounts of their own outside the school's view.
E · Expertise. Which decision stays with the teacher? The grade, the assessment, the advice about a student. Write it down, because the tool makes the work lighter and doesn't take over the judgement.
S · Storage. Where is the data kept, may it be used to train models, and does the data processing agreement say so? For a school this is where most is fixed, because the GDPR has rules for it.
What the law and the education sector already ask
A board doesn't have to think up three things itself.
The data processing agreement. Under Article 28 of the GDPR, anyone who has a supplier process personal data may only use a processor that provides sufficient guarantees, and has to set the processing out in a binding contract. Kennisnet advises using only AI services whose agreements on personal data are laid down in a data processing agreement. For education there are model agreements from the Privacyconvenant Onderwijs (the education privacy covenant), which Kennisnet points to in its framework for information security and privacy.
The DPIA. Article 35 of the GDPR requires a data protection impact assessment before processing that is likely to result in a high risk for the people concerned, in particular when new technologies are used. Kennisnet's practical version: carry out a DPIA when an AI application processes personal data.
The teachers themselves. Article 4 of the AI Act requires providers and deployers of AI systems to take measures for the AI literacy of their staff. Rolling out a tool without explaining it to the people who'll use it is, in other words, unfinished.
Whether a given application falls into a higher risk category depends on the application. A tool that prepares lessons is something else than a tool that assesses students' results. Look that up per tool, and don't count on a blanket pass.
How to read the answers
Give each answer one of three ratings: present, partial or absent. Present is specific and findable in a document. Partial was promised verbally, or the feature exists but the explanation is missing. Absent is an evasive answer, such as "we take privacy very seriously", or no answer at all.
One absent on Storage or Constraints is enough for a school to stop. Those are the layers where student data leaves the building, and a good conversation won't change that. An absent on Retrieval or Trail is more likely a condition for a pilot: in a small group, with agreements, and a date on which the board looks again.
If you want to see where the chain behind your own teaching material is weakest, the Ketencheck on this site goes through the four phases of SEAL: Source, Edit, Approve and Launch. That tells you which of the six layers weighs most for you.
What you can do this week
- List the AI tools teachers use right now, official or not. Just ask them.
- Pick one, and send the vendor the six questions in a single email. Ask for answers on paper, with a reference to documentation or the contract.
- Find out whether there's a data processing agreement, and whether a DPIA has been done. If either is missing, that's the first item on the agenda.
- Put the answers in a table with three columns: layer, answer, rating.
- Record the decision on one sheet: who may use the tool, for what, under which agreements, and when the board looks back.
Sources
- Kennisnet, Privacy en AI: hier moeten scholen op letten (updated 22 September 2025). kennisnet.nl
- Kennisnet, Kansen en risico's van generatieve AI in het onderwijs. kennisnet.nl
- Normenkader informatiebeveiliging en privacy voor het onderwijs, sub-project 2.8, pointing to the model data processing agreements of the Privacyconvenant Onderwijs. normenkaderibp.kennisnet.nl
- Regulation (EU) 2016/679 (GDPR), Articles 28 and 35. eur-lex.europa.eu
- Regulation (EU) 2024/1689 (AI Act), Article 4. eur-lex.europa.eu