What to teach at which age

There is no syllabus yet

Be clear with yourself and your students about this. UNESCO published competency frameworks for students and for teachers in 2024, several countries have added AI to their computing or citizenship curricula since 2023, and none of it is settled. What a 12-year-old “should” know about AI is being decided now, in several places at once, differently. That is not a reason to wait. It is a reason to teach a progression built on the mechanism, which does not change, rather than on a product list, which changes monthly.

The UNESCO student framework is a useful spine if you want one: four strands (a human-centred mindset, the ethics of AI, the techniques and applications, and designing systems) at three levels of depth, understand, apply and create. What follows maps roughly onto it and is organised by what a child of each age can actually hold.

Ages 8 to 10: pattern and sorting

The idea that fits: a machine can learn to sort things by looking at many examples, and it makes mistakes at the edges.

Teach it with objects, not screens. The sorting game in the next lesson is the core activity. Children this age already have the concept (they have sorted animals, shapes, words) and the new step is that a machine can be trained to do it by being shown examples rather than told rules. They can also grasp the edge case: the dog that looks like a cat, and why the machine gets it wrong. That is the whole of machine learning at a level a 9-year-old owns.

What does not fit yet: probability, “predicting the next word”, anything about how a chatbot works. They will ask; “it has learned patterns from lots of writing, like the sorting game but with words” is enough.

Ages 11 to 13: prediction

The offline lesson’s activity (the class as the model, the sentence stem, the tally) is built for this age. They can hold the idea of a spread of likely answers, they can see why “Tokyo” is unanimous and “sea” is rare, and they can draw the consequence: sometimes it is reliable and sometimes it is guessing, and the tone does not change.

Add: where the words came from, and the retelling activity for bias. Add: the confident wrong answer, on paper. That is a complete, honest unit for this age, and a student who has it knows more about the mechanism than most adults.

Ages 14 to 16: training data, bias and evaluation

Now the questions become about people. Whose writing did it learn from? What did it learn about doctors and nurses, about names, about places like ours? What happens when it is used to decide who gets a loan or an interview? How would you test whether it is fair?

This age can run a small evaluation: give the model twenty questions, mark the answers, compute a percentage, and then ask whether the twenty were a fair test. They can compare two tools. They can read a news story about an AI failure and identify which part of the mechanism failed. And the deepfake lesson belongs here, because this is the age at which it happens to them.

Ages 16 to 18: how it is built, and what it is for

The project lesson (building a tiny model by hand) fits here, along with the questions that have no settled answer: what it does to work, what it does to writing, whether it should be used in schools at all, whether it understands anything. Students this age can hold an unsettled question as unsettled, and it is worth showing them that the adults have not resolved it. A debate with real positions, sourced, is a better lesson than a lecture with a conclusion.

For students heading toward technical study, this is where Python, a notebook and a real dataset enter, and the free path is Google Colab or any phone browser with an online Python environment.

Across the ages: three things every year

Whatever the age, three things recur and deepen:

1. It learned from us. Its habits are our habits. Its errors about our town are because we wrote little about our town.
2. It is confident when wrong. Sounding right and being right are different, and telling them apart is a skill.
3. Someone built it and someone profits from it. It is a product, made by a company, with choices inside it. This is the citizenship strand and it is the one most often skipped.

The honest limit

A progression like this assumes time that most schools do not give to a subject with no exam. In practice AI is taught in the gaps: a computing lesson, a PSHE slot, a form period, a science lesson that has a spare twenty minutes. The final lesson of this block is about making every subject carry a piece of it, so that the progression happens across the timetable rather than in a slot that does not exist.