Teach the Child Before Assisting the Task
AI can widen access to tutoring. Schools must still protect the unaided effort through which knowledge and judgement become the child's own.
Generative AI can improve a child's homework while weakening the skill the homework was supposed to build. That is not a contradiction. It is the central design problem for AI in education.
A polished answer measures the joint performance of student and tool. Education must also measure what remains when the tool is removed: memory, comprehension, judgement and the capacity to begin.
Schools should therefore adopt an effort-before-assistance rule. For tasks intended to build core knowledge, the child first attempts the explanation, solution or structure. AI can enter afterward to question, hint and critique. Assistance should reveal thinking rather than make it unnecessary.
Better work can conceal weaker learning
In a study of nearly 1,000 secondary mathematics students, access to a general GPT-4 assistant improved practice performance but was followed by worse unaided test performance. A tutor designed with safeguards largely avoided the harm. The experiment appeared in the Proceedings of the National Academy of Sciences.
No single experiment applies to every age, subject or system. The finding nevertheless exposes a measurement error. If a school evaluates only AI-assisted output, it may reward a programme that makes assignments look better while students become less capable independently.
Design changes outcomes. A system that immediately supplies the completed solution encourages delegation. One that asks the learner to attempt a step, offers a limited hint and waits can support productive effort.
Human mediation matters too. A field experiment involving 700 tutors and about 1,000 students found that an AI assistant for tutors increased topic mastery by roughly four percentage points. The working paper studied technology expanding an instructor's capacity, with larger benefits among lower-rated tutors.
The serious comparison is not AI versus no AI. It is one learning design versus another.
Make the mode explicit
Every assignment should state one of three modes.
In human-only mode, pupils build foundational recall, reading, writing, calculation and first-pass reasoning. Devices are absent, and the work shows what the learner can do alone.
In AI-assisted mode, the student records an initial attempt before requesting help. The final work notes the important suggestion, what changed and which advice was rejected. The objective is revision with judgement, not invisible substitution.
In AI-native mode, using the system is part of the subject. Students compare outputs, test bias, verify sources, design prompts or build applications. Assessment rewards responsible operation and independent checking.
Visible modes protect honest students. Vague rules create a classroom where those who disclose assistance are punished while concealed use succeeds.
Literacy is not permanent dependence
Young people need AI literacy. They should understand that models predict rather than know, can invent confident claims, reflect data bias and may expose information entered into them. They should learn to verify and declare assistance.
UNESCO's AI competency framework places human-centred judgement and ethics alongside technique. Its guidance on generative AI stresses age-appropriate use, data protection and validation of educational systems.
That does not imply that a model should accompany every lesson. Schools teach calculators while preserving mental arithmetic and teach search while still requiring knowledge. AI can be both an object of study and a tool whose use is deliberately limited.
Protect children as users, not data sources
UNICEF reported in June 2026 that children were adopting AI services more than three times faster than adults in its ten-country analysis. At least 20 million children used the systems, and about 13 million reported homework use. UNICEF also cautioned that evidence about effects remains incomplete.
Schools should not respond to uncertain evidence by turning pupils into an uncontrolled experiment. Identifiable records, health information, family details, photographs and unpublished work should not enter consumer tools without an approved data agreement. Vendors should not train on children's interactions by default or construct advertising profiles from them.
Younger children need restricted, school-managed systems used for a defined lesson. Teachers and parents need plain information about retention, review and deletion.
The strongest countercase
First effort cannot mean one rigid format. A learner with dyslexia, a disability or limited command of the classroom language may need assistance before producing conventional text. A child in an overcrowded classroom may receive more useful feedback from a guarded tutor than from no tutor at all.
Accessibility support is not misconduct. The first attempt may be spoken, visual or made with assistive technology. Schools should design accommodations with the learner and invest first where AI can extend scarce teaching capacity.
The governing test is agency: does the support help the child acquire a capability, or merely hide its absence?
Assess what remains
Every assisted unit should include a short unaided act of retrieval, explanation or application. It can be an oral defence, an in-class paragraph or one worked problem rather than another high-stakes examination.
Schools should track both assisted and independent progress. If only the first rises, the programme is producing improved artefacts, not improved education.
Children will work alongside AI. Education should prepare them for that world without surrendering the difficult process through which a thought becomes theirs. Teach the child first. Then assist the task.
The Global Federation treats education as the formation of agency. A useful tool expands what a learner can eventually do without surrendering judgement to it.