11How should I use it?

Real or Fake?

One of the two is real. The other one was made by AI.

Pick the one you think is real.

Round 1 was easy. The dog face was not even there.

How about round 2? You probably missed one or two.

Both machines do the very same thing. But how many pictures they saw is different. The round 1 machine saw two thousand. The round 2 machine saw eighteen thousand.

That is why your eyes alone are not enough. Today's clue is gone tomorrow.

When you see a picture, ask this. Where did it come from? Who put it there?

When you show someone a picture AI made, say so. Then they can ask too.

A little deeper

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

How convincing a fake is depends on model size and how much data it saw. The two rounds here show that difference directly.

Round 1's machine ran 12,000 steps on 2,059 dog photographs from Pexels. The results do not even read as dogs. Round 2's machine is the one from chapter 1, which saw 18,000 character pictures. They do the same job. But the amount is not the only difference — photographs are far harder material than simple drawings, and the round 1 machine was also smaller and ran fewer steps. Pinning the whole gap on data volume would be an overstatement, so with the child we go only as far as "the more it sees, the closer it gets to real."

The real pictures in round 2 are crisp vector drawings while the fakes keep some grain, so a sharp-eyed child can tell them apart by texture. All the more reason not to leave blurry means fake as the rule — that clue vanishes as soon as the machine gets a little better.

Watch out here — if the child gets a few right in round 2 they may finish with "I'm good at spotting fakes." That is the opposite of the lesson. Missing one or two of the twelve is the point.

The rule to leave them with is not blurry means fake. That flips within a few years. What has to survive is the habit of asking where it came from — "where did you see this? who posted it?"

Try together — write down one clue the child found today, then ask "will this clue still work next year?"