What Should We Teach Before We Teach Young People to Use AI?

If I could teach a young person only one thing before introducing an AI tool, I would not begin with how to phrase a better request. I would not start by comparing which service is fastest or most capable.

I would begin with a habit: when an answer looks polished, complete and confident, keep your own judgement switched on.

More fully, young people need six capabilities before tool fluency becomes the goal. They should learn to judge sources, recognize uncertainty and stopping conditions, disclose what AI contributed, remain responsible for outcomes, protect privacy and other people's rights, and choose not to use AI when it is the wrong tool.

This is not an argument for keeping technology out of childhood. AI is already present in search, recommendations, writing, translation, images and conversational services. Waiting for a perfectly controlled moment is not realistic. The meaningful choice is whether education begins with operational speed or with human agency.

My position is simple: establish judgement first, then accelerate tool use. A new interface can be learned in days. Habits of evidence, responsibility and principled refusal take repeated practice.

Using AI is not the same as being AI-literate

A student can become very good at asking a system for a presentation, rewriting a paragraph or producing an image without knowing where any claim came from. They can deliver polished work without being able to explain which ideas they understand and which decisions the system made for them.

Those are not the same capability.

UNESCO's AI competency framework for students describes twelve competencies across four dimensions: a human-centred mindset, ethics of AI, AI techniques and applications, and AI system design. It develops them through understand, apply and create levels. The OECD and European Commission's 2026 learner framework similarly frames AI literacy as the knowledge, skills and attitudes needed to understand, critically evaluate and use AI ethically and creatively.

Those are facts about the institutions' frameworks. They do not prove that every school must teach the same sequence, and completing a framework does not guarantee learning outcomes.

My educational inference is that a curriculum focused only on operation teaches a temporary interface. A curriculum that asks when to trust, when to verify, when to stop and who remains accountable develops capabilities that can survive the next product cycle.

1. Judge the source, not just the answer

The most misleading AI output is not always obviously false. It is often an uncertain claim delivered with the rhythm and confidence of a settled fact. Fluency lowers our guard. A tidy structure can feel like evidence even when no evidence is present.

The first question, then, is not only “Is this correct?” It is also:

• Where is the original source?
• Is it a study, a public record, a first-person account, a news report or another unattributed summary?
• When was it published, and is it still applicable?
• Do multiple sources independently support the claim, or are they repeating one another?

A family can practise this with an ordinary, low-risk question: an animal's behaviour, a local historical event or the rules of a familiar sport. Read an AI answer, then locate two traceable sources. The goal is not to catch the system making a mistake. It is to make “answer” and “evidence” separate categories in a young person's mind.

Older students can compare the quality of sources. How does a research paper differ from its press release? Does a government statistic answer the same question as an opinion column? What are the sample, denominator and date behind a number?

This work is slower than instant production. It is also one of the most valuable forms of slowness we can preserve.

2. Recognize uncertainty and stopping conditions

Knowing that AI can be wrong is not enough. A learner needs to recognize when they are no longer qualified to keep guessing with it.

I find a three-level model useful.

Low-risk work can be explored freely and checked easily: brainstorming story titles, organizing notes the student already understands or trying different tones. Medium-risk work needs sources or review by a capable adult: historical interpretation, admission requirements or a claim about a classmate's behaviour. High-risk situations require a stop and the right human help: medical symptoms, a safety threat, a signal of self-harm, a legal decision or any recommendation that could directly harm someone.

Literacy does not mean doubting everything equally. It means increasing verification as consequences increase.

The OECD Digital Education Outlook 2026 draws an important distinction from the evidence it reviews: general-purpose generative AI can improve immediate task performance without producing equivalent learning gains, while pedagogical design and guidance matter. That is not evidence that every use of AI harms learning. It is a warning that a better-looking output is not the same measurement as a stronger capability.

A practical stopping question is: “If this answer is wrong, who could be harmed, and would I know?” When the possible consequence is serious and the learner cannot independently verify the answer, the process should stop.

3. Disclose what AI actually contributed

Disclosure should mean more than adding “AI was used” to an assignment. That statement is too broad to help a teacher, peer or reader understand how the work was made.

A useful disclosure answers four questions:

• At which step was AI used?
• What suggestions, text or analysis did it provide?
• What did the student keep, change or reject?
• Who verified the facts and chose the final position?

For example: “I drafted three arguments myself and used AI to suggest counterarguments. I checked two primary sources, removed one unsupported claim and rewrote the conclusion.” That short account has educational value because the student must describe their own thinking.

Schools do not need to treat every use of AI as cheating. They can define three understandable conditions instead: permitted without special disclosure, permitted with disclosure, and prohibited for this particular task. Specific rules make honest choices more likely than vague warnings do.

Disclosure must also be proportionate. It should illuminate authorship and decision-making, not become continuous surveillance of every click and keystroke.

4. Keep a human responsible for the outcome

“The AI wrote it” cannot be an exemption from responsibility.

If a student submits a fabricated quotation, a false number or language that harms a classmate, the consequence does not disappear because a system produced the words. Responsibility also does not belong to the student alone. Teachers, school leaders and product providers cannot require widespread use and then place every failure on an individual child.

UNESCO's AI competency framework for teachers puts human agency, ethics and pedagogy alongside technical competence. That framing matters. The adult role is not merely to demonstrate operation. Adults must define understandable rules, provide correction and appeal, and keep consequential decisions from becoming fully automated.

One classroom exercise is to attach a small responsibility note to a project: Which facts did I verify? What remains uncertain? What will I do if I discover an error? The point is not to manufacture paperwork. It is to turn accountability from an abstract value into a repairable commitment.

Responsibility should include the right to correct the record. A healthy learning environment rewards a student who identifies and repairs an error instead of teaching them to hide uncertainty to protect a grade.

5. Protect privacy, relationships and other people's rights

AI can feel like a private conversation. A young person may enter information about their health, emotions, location or school. They may also share a friend's secret, a classmate's image or details about a family member who never agreed to participate.

UNESCO's guidance for generative AI in education and research calls for a human-centred, privacy-protecting and age-appropriate approach. UNICEF's third edition of Guidance on AI and Children, published in December 2025, approaches AI through children's rights, including privacy, safety, fairness, transparency, accountability and the best interests of the child.

These principles can become one memorable rule: do not give a system private information that can identify you or someone else without informed permission.

“Don't share your password” is necessary but far too narrow. A face, voice, health concern, school schedule, private conflict or unique combination of details may also identify someone. Young people need examples grounded in relationships, not only cybersecurity vocabulary.

Adults must follow the same standard. A teacher should not paste student writing, learning difficulties or behaviour records into an unapproved service for convenience. A parent should not secretly submit a child's most vulnerable conversations for analysis and assume that care cancels the need for consent.

Protection should not turn into surveillance. If a school monitors everything a student does to determine whether AI was used, it may replace one risk with another. Good governance collects the minimum necessary data, explains the purpose, limits retention and gives students a meaningful way to question a decision.

6. Preserve the ability to work without AI

The most neglected capability may be the ability to say, “Not for this task.”

A learner may decline AI because the purpose is to practise memory, reasoning, handwriting, oral expression or a difficult human conversation. The material may be too sensitive. Or the learner may simply need to hear their own first thought before inviting a system into it.

Refusal is not backwardness. A choice is only meaningful if non-use remains possible. If a student must use one system to complete an assignment, receive feedback or participate in class, the school has created a requirement, not an option.

UNICEF's Adolescent Perspectives on Artificial Intelligence reports on workshops held in 2020 with 245 adolescents across five countries. One enduring lesson is that young people should meaningfully participate in decisions about AI that affect them. This was qualitative work conducted before generative AI became common, so it cannot represent every young person or today's exact product environment. It still challenges adults to design rules with students rather than only for them.

Schools should therefore maintain a workable alternative where possible. The alternative does not have to be identical in every detail, but it should let a student meet the learning objective without surrendering sensitive data or accepting a tool they cannot meaningfully question.

Practical exercises for families and schools

You do not need a new platform to begin. Five small practices can make the six capabilities visible.

1. Deconstruct an answer. Mark each sentence as a verifiable fact, an inference or a recommendation. Find sources for the facts.
2. Use red, yellow and green conditions. Green tasks are safe to explore. Yellow tasks require verification or capable adult review. Red tasks require stopping and finding an appropriate person.
3. Write a contribution note. In three sentences, describe what AI did, what the learner changed and what remains uncertain.
4. Run a no-AI round. Begin part of the same task without AI, then compare what the tool added and what it removed from the learning process.
5. Reverse the roles. Let students propose the classroom rules. Adults must explain why a piece of data is collected, how a decision can be appealed and when the data is deleted.

Assessment should change with the exercise. Do not look only at whether the final product is polished. Ask whether the learner can explain the evidence, identify limitations, correct an error, protect someone else's information and stop when the system is inappropriate.

Age matters. A younger child can begin with “Who says this?” and “May I upload my friend's photo?” A teenager can work with conflicting sources, model bias, consent, competing interests and institutional appeal. The same expectations should not be imposed on every developmental stage.

Capacity matters too. Teacher workload, device access, language support and professional development shape what a school can implement. A new framework that quietly transfers every cost to classroom teachers is unlikely to become good practice.

What to watch over the next 12 to 24 months

I will watch four signals.

First, do school policies move beyond total bans or unrestricted use toward task- and risk-specific rules? Second, does assessment begin to recognize process evidence instead of relying only on the final product? Third, can students understand how their data is used and choose a realistic alternative? Fourth, do teachers receive time, training and an appeal process, rather than another list of responsibilities?

These signals matter more than adoption counts. A high percentage of students using AI says very little about whether they are learning, becoming more independent or sharing data under acceptable conditions.

What would cause me to revise the “judgement before fluency” position? Strong longitudinal evidence would matter: if beginning with intensive tool use and adding ethics and judgement later consistently produced better understanding, transfer and equitable outcomes across ages and resource settings—without increasing privacy harms, dependency or confused accountability—I would reconsider the sequence.

The opposite outcome should also be treated honestly. If usage rises and projects look more impressive, but students cannot explain the core idea without the tool, locate the source of a claim or refuse an unreasonable system requirement, we should not call that successful AI education.

Frequently asked questions

At what age should AI literacy begin?

There is no single threshold. Once a child encounters recommendations, generated content or a conversational assistant, they can practise age-appropriate habits around sources, privacy and asking for help. The depth should grow with development; it does not need to arrive as one complete course.

Will teaching children to question AI make them afraid of technology?

It should not. Good judgement education teaches both when a tool is useful and when it carries risk. The goal is not automatic distrust. It is trust supported by evidence, limited to an appropriate scope and open to revision.

Should schools ban AI entirely?

A total ban rarely reflects the different purposes of different assignments. A clearer approach identifies when AI is permitted, when it is permitted with disclosure and when the learning objective requires independent work. A meaningful alternative should remain available to students who cannot or do not wish to use a particular service.

How can we tell whether a student understands rather than memorizes rules?

Watch what they do in a new situation. Can they ask where a claim came from, what happens if it is wrong, whose data is involved, who needs to know, and when to stop? Literacy appears in choices, explanations and repairs, not only in a quiz answer.

Sources and further reading

• UNESCO: AI competency framework for students
• UNESCO: AI competency framework for teachers
• UNESCO: Guidance for generative AI in education and research
• UNICEF: Guidance on AI and children
• OECD and European Commission: Empowering Learners for the Age of AI
• OECD: Digital Education Outlook 2026
• UNICEF: Adolescent Perspectives on Artificial Intelligence

Teaching young people to use AI matters. Teaching them where it belongs matters more.

A tool can assist thought without replacing judgement. It can contribute to a project without absorbing responsibility. It can offer options without taking away the right to refuse.

When a young person can choose deliberately between convenience and dignity, speed and understanding, an answer and its evidence, we are no longer teaching the interface of the moment. We are helping the next generation hold its place in a world full of intelligent systems.