Teaching a child to use AI without outsourcing the thinking

The useful question is not whether a child has used AI. It is what part of the thinking remains visible afterward.

UNICEF’s Guidance on AI and Children 3.0 recommends supporting children’s self-sufficiency and critical thinking, using AI to complement their own input and abilities, teaching basic concepts and data literacy, encouraging responsible use, and supporting teachers. Those are policy recommendations. They do not prove that one exercise improves learning, and they do not establish access rules for every tool or setting.

Here is a fully fictional, adult-led exercise. A teacher gives a learner a short description of a fictional pond: three frogs are on a log, two are in the water, and one leaves the pond. The learner’s task is to write a question that can be answered from those facts, then solve it. The learner writes, “How many frogs remain?” and answers 4.

The teacher permits an AI assistant to ask one question at a time, with the learner’s work kept visible. The assistant asks, “Which frogs are included in the starting count, and which one is removed?” The learner writes the calculation: 3 plus 2 minus 1 equals 4. The teacher can see the original facts, the learner’s question, the learner’s arithmetic, and the point at which the assistant intervened.

A different use would ask the assistant to produce the question, calculation, and explanation in one polished response. That might complete the exercise, but the learner’s choices would be harder to inspect. My preference is to ask the tool for a prompt, a request for evidence, or a question about a visible step before asking it for a finished answer. This is a proposed practice, not a guarantee of learning or a claim about how every child will respond.

The adult’s role matters in this fictional setting. The teacher decides the task, chooses the permitted tool, checks the response, and keeps the learner’s own work in view. The learner remains responsible for stating the problem and checking the arithmetic. If the assistant introduces a fact absent from the pond description, that addition becomes visible and can be rejected.

A useful record has four parts: the original facts, the learner’s first attempt, the assistant’s question or suggestion, and the learner’s revision. That record does not make the tool infallible. It makes the division of work inspectable.

The lesson I would keep is simple: use AI to widen the learner’s next question, while preserving the learner’s first question, evidence, and revision. Assistance is most educational when the thinking it supports can still be seen.