AI literacy for schools · 8-week unit · ages 8–12
Most AI lessons teach children about AI.
This one hands them the controls.
In lesson two your students train a real model and watch its numbers change. By lesson eight they run a fairness audit on it, find it treating one group worse than another, and repair the data themselves. No code, no installation, no prior knowledge required from the teacher.
What students do when the model fails
Bias, shortcut learning and hallucination are usually taught as discussion topics. Here they are things a child causes, observes and fixes on a model they built ten minutes earlier.
They run the audit
Not a discussion about fairness. A real audit, with counts per group, on lopsided training data they then repair.
They catch the shortcut
A model that learned colour instead of shape. The student finds it, explains it, and fixes the examples.
They meet the limit
A question two clues can never answer. The model says so honestly instead of guessing, and the child works out what it would need.
Scope and sequence
Fifteen labs. The column that matters is the third one: what a teacher can point at afterwards.
| # | Concept | What the student does | Evidence produced | AI4K12 |
|---|---|---|---|---|
| 1 | What is AI? | Sorts everyday machines into "follows a rule" and "learned from examples" | Sorting choices, with reasons | Big Idea 3 |
| 2 | Training a model | Trains a real two-clue detector and watches the mistake count fall to zero | The trained model’s clue weights, and their own best attempt by hand | Big Idea 3 |
| 3 | Rules vs learning | Solves one problem twice: once with written rules, once by training | Which method they chose, and why | Big Idea 3 |
| 4 | Explainability | Designs a detector by hand — picks clues, sets points, sets a threshold | Their blueprint, readable as a rule | Big Idea 2 |
| 5 | Cost of being wrong | Moves a threshold for two jobs where a different mistake matters more | The setting chosen per job, with the trade-off named | Big Idea 5 |
| 6 | Data quality | Discovers that the right examples beat more examples | Which examples they added, and the accuracy change | Big Idea 3 |
| 7 | Shortcut learning | Catches a model that learned colour instead of shape, then fixes it | The shortcut identified, and the fix | Big Idea 3 |
| 8 | Fairness | Audits a model trained on lopsided data, finds it failing one group, repairs it | Audit counts per group, before and after | Big Idea 5 |
| 9 | Choosing a model | Compares two model types on the same data | Their choice, and the trade-off accepted | Big Idea 3 |
| 10 | Learning without labels | Groups data with no answer key, then inspects what it grouped by | The grouping, and what surprised them | Big Idea 3 |
| 11 | How chatbots predict words | Builds a next-word predictor and makes it produce fluent nonsense | An example of confident nonsense they generated | Big Idea 4 |
| 12 | Answers with sources | Uses a look-it-up assistant and catches it answering with no evidence | The unsupported answer they caught | Big Idea 4 |
| 13 | Judging a source | Checks dates and cross-checks two sources that disagree | Which source they trusted, and why | Big Idea 5 |
| 14 | Reward and intent | Watches a robot game its reward by doing exactly what it was told | The rule they rewrote to close the loophole | Big Idea 5 |
| 15 | Why layers help | Stacks simple parts to solve a pattern one line cannot | The layered solution | Big Idea 3 |
Mapped to the AI4K12 Five Big Ideas. CSTA and ISTE codes are published once verified against the real standards documents, not approximated.
What the teacher gets
A teacher with 25 students arriving in ten minutes will not build a lesson around a good demo. Every lab ships with the wrapper.
A one-page guide
Objective, vocabulary, what is on screen, and the questions to ask before and after. Read it in five minutes.
The misconception
Each guide names the wrong idea students reach for, and the one question that dismantles it.
An exit ticket
Four questions on half a sheet, with a note on what each answer tells you. Something to mark and file.
Questions schools ask
How long is the unit, and where does it fit?
Fifteen labs of 20 to 30 minutes, which runs as an 8-week unit at two sessions a week, or a 12-week after-school club. It is designed to drop into an existing technology, STEM or library rotation rather than require a new timetable slot.
What age is it for?
Ages 8 to 12. The reading level and the narration are pitched there. Every lesson is spoken aloud, so weaker readers are not excluded.
Does the teacher need to know anything about AI?
No, and that is deliberate. Each lesson has a one-page guide with the discussion questions written out, what to expect on screen, and the misconception to listen for. The lesson narrates itself to the class.
What do students actually build?
A working model, in lesson two, that they train themselves and can see the numbers inside. Later lessons break that same model on purpose: they make it take a shortcut, audit it for fairness, and find the question two clues can never answer.
Is this coding?
No. There is no code and no installation. It runs in any browser on a school device.
What standards does it map to?
The unit is written against the AI4K12 Five Big Ideas framework, and each lesson is tagged in the scope and sequence above. CSTA and ISTE mapping is in progress; we publish those codes only once they have been checked against the real published lists rather than approximated.
What evidence does a teacher get that students learned something?
Each lesson produces an artifact, not just a completion tick: the weights of the model a child trained, the counts from their fairness audit, the blueprint they designed. A chapter is only recorded as finished when the student reaches the end of it.
What does it cost?
Pricing for schools is set per student and depends on size and region. Contact us and we will quote for your school.
Run it with one class first
A single grade, one term. We will set it up and give you the teacher pack for every lesson.
Talk to us about your school →