Most schools assume teaching AI properly needs a specialist they cannot hire. Building the AI Ninjas curriculum taught us the opposite: if the course is designed right, the specialist knowledge lives in the materials, and a good teacher of any subject can lead the room. Here is what that design takes.
1. Decide what “teaching AI” means at your school
There are two very different courses hiding under the same name:
- AI use: how to prompt a chatbot, when not to trust it, and the school’s rules on using it for homework. Every student needs this, and it fits in a few lessons.
- AI understanding: how a model actually learns from data, why it is sometimes confidently wrong, and how to build one. This is a subject, and it takes years, like any other.
Most products on offer are the first kind sold as the second. If you want students who can explain a model rather than operate one, they have to build models.
2. Make every unit end in something that runs
Students remember what they build. We structure the path so each stage produces working code:
| Year | Focus | Student builds |
|---|---|---|
| 1 | Data as numbers, first Python, first prediction | First working AI programs |
| 2 | Finding patterns automatically | A first predictive model |
| 3 | Which inputs matter; layers and weights | A two-layer neural network |
| 4 | Optimization; error down, accuracy up | A three-layer network with backpropagation |
| 5 | Production practice | A model in Keras and TensorFlow |
The math is taught where it earns its place (vectors when you need them for an image, gradients when you need them to learn) rather than front-loaded or skipped.
3. Put the expertise in the materials, not the teacher
This is the decision that makes the rest possible. Two things carry it:
- Lessons that explain themselves. Every concept is taught in the student-facing material, in short lessons of two to seven minutes. The teacher leads the room; they do not lecture from scratch.
- A mentor guide for every session. With the answers, the misconceptions students commonly hold, and the points where a class usually gets stuck. Planning is done before term starts.
With those in place, training is short. Ours is a 45-minute get-started session before the first lesson, a program onboarding before the first full cohort, and optional co-teaching after that. The skill we train most is not machine learning; it is unblocking a stuck student by asking the right question rather than handing over the answer.
4. Fit the schedule you already have
A new period is the fastest way to kill a pilot. Short lessons and around two modules a year (10 to 25 hours of student time) fit inside computing, math or STEM electives, or an after-school club. U.S. schools often start around Grade 8; schools elsewhere start earlier. Make the last year lighter: a capstone needs time more than pace.
5. Assess with evidence you can show
Department heads, parents and inspectors all ask the same question: what can the students actually do? Build the answer in from the start:
- A short diagnostic at the start, so progress is measured from a baseline.
- A mastery check at the end of every module, with concept questions and a small piece of working code.
- A rubric-graded artifact at the end of every level, against a published rubric, not a private opinion.
- Credentials that verify, so a certificate means the work was done, not that the videos were watched.
6. Ask hard questions about student data
Any platform your students use for AI will hold their names and their work. Before you sign, ask the vendor:
- Will you sign a data processing agreement with the school as the controller?
- What exactly do you collect? (The answer should be short: name, school email, grade, class, progress.)
- Is any student data used for advertising or to train AI models?
- Are leaderboards optional, and off by default for students under 13?
- How do parents see their child’s record and request deletion, and how fast do you act?
- What happens to the data when we leave?
U.S. schools and education service providers have responsibilities under student-privacy laws, including FERPA and COPPA, that depend on the context. Request written answers to these questions and have the appropriate school staff review them before selecting a vendor.
7. Pilot small, then decide
One class, one teacher, one semester. At the end you have student work, mastery data and a teacher’s honest view, which is a far better basis for a department decision than any demo. That is how we recommend every school starts with AI Ninjas, and it is what we would recommend with any provider.