Say it like this

For the child

Imagine you showed a computer a thousand pictures of cats.

Not one. A thousand. After a while the computer starts noticing things that keep showing up. Pointy ears. Whiskers. Fur. Two eyes near the top. It does not know what a cat is. It just knows what keeps appearing in the pictures people called cats.

Then you show it a brand new picture it has never seen before, and it says: I think that is a cat.

That is artificial intelligence. Noticing what things have in common, and then using that to guess about something new.

Here is the part worth knowing. If you only ever showed it orange cats, it might decide a black cat is not a cat at all. The computer only knows what you showed it. So the examples matter just as much as the computer does.

For the grown-up

What is actually going on underneath

The technical version is not far from the crayon version, which is unusual and worth taking advantage of.

A model is trained on a large volume of examples and learns statistical regularities in them, then applies those regularities to inputs it has not seen. There is no representation of a cat in there, no concept, no understanding. There is a very large set of weighted associations that produce the output "cat" with high confidence.

The thing most worth protecting in a child explanation is the absence of understanding. Once a child believes the computer knows what a cat is, every later conversation about what AI can and cannot be trusted with has to start by dismantling that. Keeping it at "noticing" costs nothing now and saves a lot later.

The training data point is also worth keeping, because it is the whole basis of every bias conversation they will have in twenty years, and a five-year-old can follow it completely: it only knows what you showed it.

Do something with it

Teach the computer wrong on purpose

Pick five pictures to show a pretend computer so it learns what a dog is. Then pick one picture that would teach it something wrong, and explain why.

Open the activity

Ask afterward

Four questions that usually go somewhere

If a computer only ever saw pictures of orange cats, what might it get wrong?
Is the computer thinking about cats, or counting what it has seen?
What would you show it a thousand pictures of?
Who picks the pictures? Could they pick badly?

Words in this one

If they ask what a word means

Artificial Intelligence A computer that has looked at so many examples it can start guessing about things it has never seen before. Pattern Something that keeps happening the same way, so you can start guessing what comes next. Data All the little bits of information somebody collected. Facts, numbers, pictures, anything you can write down. Example One of the things you show a computer so it can learn. Show it a thousand and it starts noticing what they share. Computer A machine that follows instructions very fast and never gets bored of doing the same thing again. Algorithm A list of steps for getting something done, written out so carefully that a machine can follow it.

If they want more

The books

When an idea turns out to be worth forty pages instead of two, it becomes a book, and the books live next door at DigitalCrayonBooks.com. Here is how the two sites divide the work.