What the official Swiss texts require of schools on generative AI

Why this page

When a school leader asks what they are allowed to do, the answer sits in three documents that are not read together and do not address the same audience. This page brings together what each one requires, with the exact reference, so it can be checked.

It is maintained rather than published once. The texts move, and the update date at the foot of the page marks the state of my reading.

The texts

1. CIIP (2025) — Intelligence artificielle générative – IAG : recommandations pour une utilisation réfléchie et mesurée dans l’enseignement. The intercantonal conference for public education in French-speaking Switzerland and Ticino. PDF
2. Eduvaud portal (2024) — IA et protection des données [information sheet], Direction générale de l’enseignement postobligatoire. Addressed to upper secondary. PDF
3. HEP Vaud (2025) — Directive 05_05 : Évaluations certificatives (as at 21 November 2025). It applies to the university of teacher education, not to compulsory schooling, but it shows a transferable mechanism. PDF

[TO BE COMPLETED: CDIP texts — exact reference and scope, to verify before publication.]

[TO BE COMPLETED: the Vaud cantonal directive applying to compulsory schooling, if one exists — to confirm with the DGEO.]

Student data

This is where the texts are clearest, and it is the first question school leaders ask.

The Swiss recommendations forbid putting student work into a generative AI tool for as long as the canton does not provide one whose reliability is guaranteed (CIIP, 2025).

The Eduvaud sheet for upper secondary puts it in one sentence: never transmit personal or sensitive data through online AI tools, and in case of doubt use the tools the canton provides (IA et protection des données, 2024).

The practical consequence is broad: marking a pile of scripts in a consumer chatbot falls outside the framework, even seemingly anonymised, for as long as the cantonal tool is unavailable.

Detection

The Swiss recommendations say what the literature says, in one sentence: detection tools, themselves based on probabilistic methods, are not able to assess reliably whether generative AI has been used (CIIP, 2025).

In other words, a detector score does not ground a decision about a student. The article that develops this explains why reliability is not the right question.

Assessment

The CIIP's explicit recommendation on assessment is to refocus on the process rather than on the finished product alone, and to accept that in homework, banning the tool is impossible.

The HEP Vaud directive on certificative assessment provides a concrete mechanism of that kind: the panel may call the student to an oral on a short extract of their own work, chosen by the panel, and concludes in principle to failure if the student does not demonstrate the coherence and accuracy of that extract (HEP Vaud, 2025).

It targets higher education, not compulsory schooling. It is worth knowing because it moves the burden of proof: you do not try to prove AI was used, you ask the author to account for their own text.

What this page does not yet say

Three questions come up repeatedly and have no sourced answer here. They will be added once the texts are verified, not before.

• [TO BE COMPLETED: what the texts say about a minimum age and about use by students themselves.]
• [TO BE COMPLETED: what is decided intercantonally and what is left to each canton.]
• [TO BE COMPLETED: obligations to inform parents, if any exist.]

What this reading rests on

I teach in Vaud classrooms and I am working on an analysis of the texts that frame generative AI in French-speaking Switzerland, as part of a paper in preparation. This page is the short, checkable version of that work.

• AI detectors and decisions
• Marking with AI
• Judging an assessment item
• Research project
• Training for teaching teams

Text written with the help of AI, reviewed and approved by the author.