Prompt tips are easy to teach. Add context. Name the audience. Ask for a table. Be specific.

Children can learn these moves quickly. That is precisely the problem. A pupil can become good at getting a fluent answer before understanding what kind of thing produced it.

Prompt fluency looks like competence. It may only be button fluency with better grammar.

Before children learn how to ask generative AI for better output, they need a small but sturdy model of what happens after they press enter. Not a computer science lecture. Four ideas will do: prediction, examples, uncertainty and purpose.

1. It predicts

A generative AI system does not retrieve a fully formed answer from a cupboard. It produces an output by predicting what should come next based on patterns it learned.

Kennisnet describes generative AI systems as working through probability. The result can vary even when a user repeats the same request. That makes a simple classroom demonstration possible.

Ask the teacher's approved AI tool the same harmless question twice:

Invent a name for a tiny boat built by mice.

If the answers differ, ask pupils why. Did the machine change its mind? Was one answer stored and the other invented later?

The useful explanation is less human: the system generated two likely continuations. It did not have a private opinion about mouse boats.

This distinction matters because conversational language invites children to imagine a person behind the screen. The interface says "I think" and "I understand." Those phrases make interaction smoother, but they are poor technical explanations.

In class, prefer "the system generated" over "the AI knew."

2. Its patterns came from examples

Prediction needs prior patterns. Generative systems learn from large collections of examples, although the exact data and training process differ between tools and are often not fully visible to users.

Children do not need to understand model weights to grasp the educational point: the output depends on what the system has encountered and how people built it.

Try an unplugged activity. Give groups a set of six imaginary book covers. Five show a dragon, a castle and a knight. One shows a dragon running a bakery. Ask pupils to invent the seventh cover based only on the set.

Most groups will repeat the castle pattern. A few may choose the bakery because it is more interesting. Now ask:

• Which examples influenced your choice?
• What kind of story became "normal" in this collection?
• Whose stories might be missing?

This is not a simulation of machine learning. It is a way to make patterned prediction discussable. It also opens the door to bias without turning the lesson into an abstract warning.

Klasse's work on computational thinking makes a similar case for beginning with concepts rather than software. Children can explore algorithms through paper instructions and physical activities. AI literacy can also start unplugged.

3. Fluency is not certainty

Generative AI is exceptionally good at sounding like an answer.

That sentence deserves a pause. In ordinary classroom conversation, confidence and coherence are useful clues. A pupil who explains an idea clearly may understand it. A reference book written in polished prose has usually passed through editing and review. A chatbot can produce the surface signs of confidence without checking whether the claim is true.

Kennisnet calls the resulting false but plausible output "hallucination." Its guidance on AI tutors also notes that generated information may look useful while still being wrong, and that hallucinations cannot simply be assumed away.

Give pupils three possible labels for any AI answer:

• we can check this now;
• we need another source;
• this is an idea, not a fact.

The labels are intentionally ordinary. "Check the model's epistemic status" will not improve a Year 5 lesson.

For a creative task, uncertainty is not always a defect. A strange boat name can be fun. For a claim about medicine, a historical event or a classmate, it is a serious issue. The same tool can move between those situations without changing its friendly tone.

Children need to notice that the purpose changes the standard.

4. The purpose comes from us

A better prompt can improve output. It cannot decide whether the task should be delegated in the first place.

If the learning goal is to practise writing a convincing opening, asking AI to write the opening may remove the work. If the goal is to compare three openings and explain which one fits an audience, generated examples may be useful material.

The difference sits outside the prompt box. It comes from the teacher's purpose and the pupil's intention.

This is why AI literacy should begin with a question that sounds almost too simple:

What are we trying to learn or make?

Then ask:

Which part should a person practise, and which part could a tool support?

Kennisnet's ethical guidance warns that systems designed to optimise learning can suppress uncertainty and creative struggle. Those awkward moments are not always inefficiencies. Sometimes they are the lesson.

A four-question routine

Before the next classroom AI demonstration, put these questions on the board:

1. What is the system predicting or generating?
2. What kinds of examples may have shaped its patterns?
3. What could be uncertain, missing or wrong?
4. What is our purpose, and which decisions stay with us?

Use them before writing the prompt, then return to them when the output appears.

For younger pupils, shorten the language:

1. What will it make?
2. What has it seen before?
3. What might it get wrong?
4. What do we decide?

The second question is necessarily approximate. Users rarely know the exact training data behind a commercial model. Say so. The goal is not to invite children to guess a secret dataset. It is to break the illusion that the answer appeared from nowhere.

Prompting comes fifth

Once pupils understand the four ideas, prompt technique becomes more meaningful.

"Add context" now means giving the system information it could not infer safely. "Ask for alternatives" creates material for comparison, not an automatic shortlist. "Check the answer" is connected to uncertainty, not added as a small-print warning after the exciting part.

The European Commission defines AI literacy broadly enough to include the knowledge and understanding needed for informed use and awareness of opportunities, risks and possible harm. That is a much larger educational aim than learning to coax a chatbot into producing a nicer worksheet.

So teach the prompt tips. They are useful. Just do not mistake them for the foundation.

Spend one lesson on prediction, examples, uncertainty and purpose first. Then open the prompt box.

Sources

• Kennisnet, What is generative AI?.
• European Commission, AI Literacy: Questions & Answers.
• Klasse, Computational thinking: especially in the age of AI.
• Kennisnet, Opportunities and limits of AI tutors.
• Kennisnet, Ethical dilemmas and AI in education.