← Back to blog
memorylearningtrainingdrawingscreen-free

The failed drawing: learning by seeing

I asked the kids to draw a dog. Romane drew a dog that looked a bit like a cat. Meryl drew a dog that existed mostly in his imagination and named him Charlie. I tried too. I wasn't much better.

The failed drawing: learning by seeing
DateMarch 29, 2026
Target age3-7 yrs
Duration35 minutes
MaterialsWhite paper, Pencils or markers, An adult who's willing to fail too

Meryl named his dog Charlie.

On the page, you had to take it on faith. The shape wasn't obvious. But Meryl had no doubts. It was Charlie. And once a child gives their drawing a name, explaining that the dog is a failure becomes a bad idea.

The three dog drawings on the table: Romane's, Meryl's (Charlie), and mine

The three dogs. Hard to say which one failed.

How an AI learns to see

An AI that recognises dogs wasn't born knowing what a dog is. It was shown a lot of images. Over time it learned which details come up most often: ears, muzzle, the shape of the body. That process is called training.

But recognising and producing aren't the same thing. I can recognise a dog perfectly well. That doesn't mean I can draw one. I'd underestimated that gap.

It's true for AI. It's true for us. The brain that identifies and the brain that makes things aren't the same circuits. You can have seen ten thousand dogs without ever learning to build one.

The activity

Blank paper, pencils, no reference. The instruction is simple: draw a dog. No pressure, no judgment. The adult draws too. That part matters.

Then look at all the drawings together. Not to pick the best one. To ask: what makes us say this is a dog? The ears? The tail? Or just someone stating it with complete confidence?

What actually happened

Romane's dog looked like a cat. Not entirely. But enough to hesitate. She knew what she wanted to draw. Between the image in your head and the line on the page, there's a gap: ears that drift, a body that can't decide if it barks or meows.

I drew too. I thought it would be easy. On paper, my dog looked like it had escaped from a tired old cartoon. Not much more convincing than the kids'. Which was reassuring for them.

Meryl wasn't in that headspace at all. He drew. He held his idea. He gave it a name.

"Charlie."

Meryl names his dog0:00 / 0:00

Three and a half. No hesitation.

I have a recording of that moment.

Romane was looking for the right shape. I was trying not to embarrass myself. Meryl already had the dog. He even had its name.

To finish

At the end there were three dogs on the table. Romane's, which looked a bit like a cat. Mine, which didn't make me want to change careers. And Charlie.

An AI trained on millions of images can recognise a dog far better than I can. It can even generate one that looks exactly like a dog.

But I wonder if it would understand what had just happened at that table.

Or if it would simply say: "dog, medium confidence."

What the kids said

Romane could see her dog wasn't quite a dog. Meryl had no doubts at all. His dog was called Charlie. I discovered that recognising a dog is much easier than drawing one.