Before AI Enters a Classroom, Five Questions Schools Should Answer
Schools are under pressure to make decisions about artificial intelligence while evidence of learning gains, data practices and costs remain uneven. A ministerial statement presented at UNESCO’s Digital Learning Week in September 2026 called for educational benefit to be demonstrated before deployment and for systems to be explainable, auditable and accountable. That principle can be made practical. Before adopting a classroom AI tool, a school should be able to answer five concrete questions.
- What learning problem does it solve?
A product demonstration can show that a tool generates fluent answers; it cannot show that pupils learn more. A school should define a specific task, such as feedback on a draft or translation support, and decide what improvement would count. A small trial with comparable assignments, teacher observations and student work can be more informative than adoption figures alone. The test should check whether learners still practise reasoning rather than simply receive finished work.
- What happens to learner data?
Children’s work may include personal details, mistakes, disabilities or sensitive family context. Schools should know what information is collected, where it is stored, how long it is kept, whether it trains a model, and who can remove it. A free product still has data and administrative costs. UNESCO’s statement treats learners and teachers as rights-holders over their data and urges public oversight even for free tiers or donations.
- Who can challenge the output?
An AI system can be useful while also producing incorrect or biased responses. Teachers need a clear way to review suggestions, override them and report recurring problems. Pupils need to know when a response came from a system and how to contest a consequential mistake. The school should identify who is responsible for decisions; a vendor’s explanation should not substitute for a teacher’s professional judgement.
- Does it work for all learners?
A single impressive English-language example says little about local languages, accessibility, age suitability or low-connectivity use. A trial should include the actual range of pupils who will use the tool and look for unequal outcomes. UNESCO’s ministers called for multilingual and inclusive evaluation as well as approaches appropriate to children’s development. A tool that saves time for some pupils but excludes others needs redesign or a different procurement choice.
- Can the school change course?
Contracts, training, device needs and data export all affect the real price of adoption. Schools should calculate ongoing costs and ask whether student work can be moved if a service changes terms or closes. An exit plan is especially important when a tool becomes part of assessment or teaching routines. The ministerial statement explicitly asks systems to count total ownership costs and favour portability and interoperability.
How to make the questions useful
These are decision tests, not a universal ban or endorsement. A school might approve a limited teacher-controlled use for one age group while declining automated assessment elsewhere. It can publish the purpose, trial period, safeguards, evaluation criteria and point at which the trial will be reviewed. That makes the decision accountable to families and staff without promising that every risk is eliminated.
UNESCO also urges investment in teacher preparation and AI competencies for both teachers and students. Training matters because a policy on paper will not reveal whether pupils are delegating the thinking a task was meant to develop. The most useful measure is whether people can explain, critique and improve the work with the tool, rather than whether they can produce more text faster.
The September statement is a global call for governance, not a binding rule for every school. Its practical lesson is that adoption should be reversible and evidence-led. The five questions turn a broad debate into a set of decisions that a school, teacher and family can actually inspect.
