AI courses for kids

What ACARA says

The curriculum's own example is "a model with many examples of horses and no zebras in its training data is likely to classify all zebras as horses". Our sorter makes the same mistake with wallabies.

Why it matters

The sorter does not say "I don't know". It gives an answer, sometimes a confident one, because nothing in its training taught it that something else exists.

The fix

Add examples of the missing category, add an "unsure" answer for low confidence, and always test on things the model was not trained on.

Where children see it in real life

A song recommender that only knows what you have already played will keep suggesting more of the same. A voice assistant may struggle with words it rarely heard in training. Each is a version of the zero problem.

Once children recognise the pattern, they start asking the right question of any AI tool: what examples was this trained on, and what might it never have seen?

The fairness link

The curriculum also asks students to consider "ensuring that the data used to train an AI model minimises any potential biases in its output and is representative of the target audience". Missing or thin categories are a common way for unfair results to creep in.

In our lessons, children train a model with a category left out, watch it fail, then add examples and watch it improve. Seeing it happen is far more convincing than being told.