← Back to blog
memorylearningtrainingscreen-free

The failed drawing: learning by seeing

I asked Romane to draw a dog from memory. Then with ten pictures in front of her. Then I tried to explain how an AI learns to recognize something after seeing thousands of examples.

Date29 mars 2026
Target age5-9 yrs
Duration35 minutes
MaterialsWhite paper, Pencils or markers, About ten pictures of dogs (magazines, prints, postcards)

Romane's first dog had four legs, two eyes, and the rough shape of a cloud.

She looked at it, laughed, and said: "It's weird." Then she kept it anyway because she liked it.

How an AI learns to see

An AI that recognizes dogs wasn't born knowing what a dog is. It was shown thousands of images — some labeled "dog," others "not dog." After seeing enough of them, it learned which details matter: ears, muzzle, the general shape of the body, how the legs connect.

This process is called training. Before training, the AI knows nothing. It's a bit like Romane facing a blank page, with just a vague idea of what a dog should look like.

The difference: the AI can look at ten million images in a few days. Romane had ten images and twenty minutes.

The activity

First round: blank paper, no references. "Draw a dog from memory." No pressure, no judgment. The result doesn't matter much.

Second round: put ten dog pictures in front of her. Different breeds, sizes, poses. Look at them together for two minutes. Then flip them face-down. "Now draw again."

Third round: the pictures stay visible. She can look as much as she wants while drawing.

Compare the three drawings. What changed? Why?

For younger kids, no need to do all three rounds — just letting them draw with and without pictures is plenty.

What actually happened

Dog number one: the cloud with legs. Romane was happy with it anyway.

After looking at the ten pictures and flipping them face-down, her second dog had floppy ears and a more defined muzzle. She'd retained the details that showed up most often in the photos.

The third, with the pictures visible, was the most accurate of the three. She was copying, deliberately. She said: "This is too easy, I'm cheating."

I said that's exactly what an AI does during training. It looks, it holds onto what comes up often, it adjusts.

"But it's not cheating for the AI?"

No. Because nobody told it that looking at examples was cheating.

Meryl was drawing alongside. Circles, lots of circles. When I asked what it was, he said: "The dog is sleeping." A curled-up dog, then. The pictures we'd shown didn't include a dog curled in a ball — he'd drawn something he knew from his own experience, not from what he'd seen in the photos.

That's exactly the opposite of what an AI does. It draws what it was shown. Not what it experienced on its own.

To finish

Romane lined up her three drawings side by side and looked at them for a long time.

"So the AI that's seen the most dogs draws the best dogs?"

Not necessarily. It draws the most average dog. The most common one. The dog that doesn't surprise anyone.

Whether that's the best dog is another question.

What the kids said

Romane's first dog looked like a cloud with legs. She found that very funny. After looking at the pictures, her dog had ears. Meryl drew circles and said it was a sleeping dog.