5How does it work?

Learned or Told?

So far AI has made you pictures, stories and tunes.

But not every machine is an AI. Build one yourself and you'll see.

With no rules the machine calls all twelve a dog. Only six of them really are, so only six are right.

Now add rules one at a time.

Add It has fur and the toys drop out. Better, right?

But keep going and it gets strange. Add Pointy ears and real dogs drop out too. Plenty of dogs have floppy ears.

The fox and the wolf never drop out. They follow every dog rule.

The machine that drew for you in chapter 1 was given no rules at all. It looked at piles of pictures and found the rules itself. Finding them yourself like that is called .

Now let's see how that works.

A little deeper

For grown-ups, and for children who want more.

Writing the rules by hand is called rule-based computing. Finding them from examples is machine learning.

Rule-based was the mainstream in the 1980s. Expert systems — an expert's knowledge transcribed into if-then rules — carried enormous hope. Then real problems arrived: rules multiplied into the thousands, contradicted each other, and threw up endless exceptions. The score falling as you add rules in this hands-on is exactly that wall.

We measured it. Running all 64 combinations of the six rules, the best scored nine out of twelve, and switching all six on dropped it to six. The fox and the wolf keep every dog rule there is.

Machine learning sidesteps the problem by never writing rules down. In exchange, it becomes hard to explain why it decided what it decided. Every gain costs something.

Ask here — "could one more rule get rid of the fox?" When the child proposes one, test it together. Almost always a real dog gets caught by it too. That moment is the heart of this chapter.

In class — come back to this page when children start calling everything AI. Automatic doors, calculators and traffic lights had every rule written out by a person.