Sector guidance
Education extension
Verified orientation snapshot: 19 August 2026. Education law and data-protection authority are often national, cantonal, regional, or institutional. Confirm the applicable mandate locally.
The educational outcome is learning, access, and learner agency—not the amount of content generated. Efficiency cannot compensate for exclusion, surveillance, invalid assessment, or a decision that a learner cannot understand or challenge.
Separate support from allocation
| Use | Example | Default boundary |
|---|---|---|
| Teacher support | Draft a lesson outline or summarize approved material | A0–A1; teacher owns pedagogy and accuracy |
| Learner support | Practice, feedback, translation, or accessibility aid | A0–A1; disclose limits and preserve non-AI access |
| Administration | Route requests or prepare records | A1–A2; approval before changing an official record |
| Assessment | Mark, proctor, predict, or detect prohibited behavior | R3 review; no sole automated adverse outcome |
| Access or pathway | Admission, placement, progression, discipline, or support eligibility | R3; formal legal and fundamental-rights review |
Education gates
E0 — Pedagogical mandate
- define the learning or service problem before selecting the tool;
- name the teacher, academic, or administrative authority that owns the outcome;
- identify non-AI options and the educational theory behind the proposed use;
- state which admissions, grades, discipline, support, and safeguarding decisions remain human.
Stop: the system has no defensible pedagogical purpose or is being used to avoid staffing, accessibility, or due-process duties.
E1 — Learners, access, and data
- identify age, language, disability, socioeconomic, cultural, and device-access segments;
- map student records, special-category data, minors’ data, intellectual property, and secondary use;
- establish the legal and institutional authority for collection, profiling, retention, cloud processing, and research;
- provide an equivalent channel for learners who cannot or should not use the system.
E2 — Classification and procurement
- determine whether the use concerns access, admission, learning-outcome evaluation, level assignment, or exam monitoring;
- require supplier disclosure of data flows, model changes, training reuse, moderation, age safeguards, and deletion;
- prohibit unannounced behavior changes during an assessment period;
- verify accessibility, language coverage, interoperability, export, and exit before purchase.
E3 — Evaluation by educational outcome
- preregister learning, workload, quality, integrity, and access measures separately;
- compare with current teaching or administration using the full learner denominator;
- test fabricated content, feedback quality, overreliance, disparate error, accessibility, and teacher correction burden;
- involve educators and representative learners in structured evaluation without exposing individual grades or sensitive records.
E4 — Bounded pilot
- start with staff-only or shadow use, then a limited voluntary learner pilot where appropriate;
- keep official grading, admission, discipline, and safeguarding outside the pilot unless formally authorized and independently reviewed;
- give learners a clear notice, human contact, correction route, and non-retaliatory opt-out where required;
- stop if educational value is not demonstrated or access gaps widen.
E5 — Accountable operation
- monitor learning outcomes, teacher workload, complaints, opt-outs, corrections, accessibility, and subgroup effects;
- re-evaluate at each model, curriculum, policy, population, or assessment change;
- preserve records needed to explain consequential decisions;
- retire uses that replace pedagogical judgment without evidence of better learner outcomes.
Minimum evidence pack
- pedagogical mandate and decisions reserved for people;
- learner/data map and applicable local authority;
- accessibility and fundamental-rights assessments;
- supplier and change-control record;
- preregistered results by relevant learner segment;
- notice, challenge, correction, and non-AI access procedures.
Transfer boundary
Evidence from one course, age group, language, institution, curriculum, or assessment format does not establish effectiveness elsewhere. Separate improved teacher throughput from improved learning, and never infer either from model benchmark scores alone.
See the dated primary-source register, the accessibility assessment, and the fundamental-rights impact assessment.
To print or save as PDF: Ctrl+P (⌘P on Mac).
Source and history · GitHub