AI arrives in the classroom: the points to settle for the 2026 school year

Published 14 July 2026 · updated 16 August 2026

In the 2026 school year, artificial intelligence officially enters several French-speaking Swiss school systems. Fribourg gives lower-secondary pupils access to Copilot, ages 12 to 15, including from home. Vaud has chosen a more careful route: a dedicated tool, usable in class only, for supervised exercises. These two cantons don’t represent every policy in French-speaking Switzerland. But they show that classroom AI has moved past a few pioneering schools.

The private schools aren’t starting from nothing. The International Baccalaureate addresses AI in learning, teaching and assessment: student work has to follow its academic-integrity policy and acknowledge the use of generated content appropriately. Schools across the region are rethinking their examinations, with oral assessment coming back wherever written homework no longer proves much. So the question in front of school leadership is no longer “should we?” but “how, and with what safeguards?”.

The work is both pedagogical and organisational

Pedagogy belongs to the teachers and can’t be delegated away. In parallel, leadership has to answer another set of questions: where the entered text goes, which pupil data circulates, who signed what with which supplier, and who stays accountable for a decision.

All of that counts double in a school. The data concerns minors, and the legal frame is specific: a private school generally falls under the federal FADP for its private processing, while a state school depends on cantonal law. A delegated public task, or the school’s status, may complicate that boundary; see the dedicated analysis. The processing register, the information to parents, the processing agreements: all of that already exists on paper, and AI arrives on top of it.

Some faculties have set a clear rule. The University of Geneva’s Faculty of Science prohibits entrusting personal or confidential data to public AI tools; that source isn’t a university-wide policy. A school that provides neither an approved tool nor a clear rule gets the worst of both worlds: teachers pasting assessments, coursework and pupil names into free accounts anyway, with no frame behind them and no trace. A governance vacuum raises the risk. The design of the tool and its supplier matter, and so does the way people use it.

School tools are already the weak link

Even before AI, the school digital services showed how exposed they are. PowerSchool was compromised in late 2024: its investigation describes exfiltration after a support credential was compromised. The precise scale varies by district and by notification, and it shouldn’t be compressed into one unexplained global number. In 2024, Singapore’s government confirmed remote wiping through Mobile Guardian. In May 2026 an attacker also claimed data from thousands of schools and universities using Canvas, without complete operator confirmation in that source. A criminal claim stays an allegation until the investigation establishes it.

Switzerland isn’t spared: in 2023 the Basel school network had 1.2 terabytes of data stolen, including pupil data later published on the darknet.

The conclusion isn’t to be afraid. It’s to classify every tool, AI included, by the decisions it actually makes. The supplier may be a processor, an independent controller or a joint controller. Use a contract that fits that role, check the processing locations and the subprocessors, protect the administrator accounts with strong authentication, and keep a plan to leave if the supplier disappoints. That’s ordinary diligence, not advanced expertise. But somebody has to do it.

The points an establishment needs to have settled

A usage policy

For pupils, teachers and administration. The IB approach gives you a useful starting point on academic integrity, not a complete data-protection policy. State what is allowed, with which tools, how assistance has to be acknowledged, and which data must never enter an unapproved service.

A list of approved tools

Few tools, but verified ones: a processing agreement, known hosting, accounts the school manages. A list like this removes the reason to open a free personal account.

Role-based training

A short first session can set the rules and answer the immediate question: which data may go in, and which may not? One session isn’t enough over time. Teachers, administrators, leadership and support staff need different scenarios, exercises, refreshers and a channel to report mistakes.

Information for parents

Proactive, and preferably before the first question at a parents’ evening. In an international school, families used to the GDPR will ask anyway, and they’ll ask in English, so the documents need to exist in both languages.

A sovereign option for the administration

For internal uses (meeting summaries, correspondence, preparing documents), there are open models hosted in Switzerland or on site, and frontier models served from Zurich cloud regions. I set out what is realistic in a dedicated article. A Swiss or local option exists, but the access, the logs, the subprocessors, the keys and the quality still need testing. A Swiss region doesn’t erase the operator’s jurisdiction.

The register that connects pedagogy and governance

For every use, write down the educational purpose, the age group, the users, the permitted and prohibited inputs, the provider’s role, whether the model trains on the data, the processing locations and duration, the human who decides, the accessibility, the integrity rule, the incident contact, the review date and the way out. The register forces pedagogy and data governance into the same conversation.

The cantonal sources cited

Cantonal policies and product features change. Review the register before each school year and after any material change at a supplier.

Cantonal policies as they stood on 14 August 2026.