Romane asked me if AI dreams.
We were at the table. I'd mentioned something about work — a tool I was using, an automatically generated response. She'd listened without saying anything, then asked the question between two bites, like she'd been thinking about it for a while.
Meryl kept eating his pasta.
I didn't know what to say.
Not because the question is hard (it is, but not in that way). Because I didn't have the words. Not the right words — not for her, not at that age, not without turning the answer into a lecture or a dodge.
I said something vague. She nodded and we moved on.
But it stayed with me.
What I realized that evening
I've been working with AI tools for a while. I read about it, I'm interested in it, I use it. And yet, faced with my six-year-old's question, I'd improvised a fuzzy answer.
It's not a knowledge problem. It's a translation problem.
There's a gap between understanding something for yourself and knowing how to explain it to someone who shares none of the same references. Kids don't have the same technical metaphors available. They don't have "neural network," "training data," "language model." They have spoons, pebbles, playground jokes, and a much sharper ability to spot a dodge than we give them credit for.
This blog grew out of that. Not from a teaching plan. From a failed answer.
The first activity (unplanned)
A few days later, I tried something. No clear objective, no materials, no structure. I'd printed four photos (three cats, one dog) and put them on the table after dinner.
"Which one is different from the others?"
Romane pointed to the dog immediately. "Him. He doesn't look like the others."
Meryl looked at the photos. He pointed to the same one. "He has different ears." Then he wanted to keep the photo of the ginger cat because he liked it.
I said: "That's exactly what an AI does — it looks for the one that doesn't fit."
Romane frowned. "But does it have eyes?"
And off we went.
That conversation lasted twenty minutes. Nothing structured. We talked about how computers "see" images, why an AI might get confused if it was shown a fox but mostly trained on dogs. Meryl asked if the dog in the photo had a name. We said no. He found that a shame.
No preparation beyond four photos. Just questions that led to more questions.
At the end, Romane asked: "Can we do it again tomorrow?"
What this blog is, and what it isn't
It's not a teaching guide. I'm not a teacher, not an AI researcher, not a child development expert.
I'm a father who thinks the topic matters and believes kids are capable — far more capable than we imagine — of grasping the core ideas, as long as we present them differently than with slides and definitions.
Every article here documents a real attempt. Some work well. Some drift sideways. Meryl runs off with a photo at the wrong moment. Romane invents categories no engineer would have thought to test.
That's what I want to keep: not the pedagogical success, but the honesty of the fumbling.
To finish
Romane asked the question again a few days ago. Whether AI dreams.
This time I said: "I think no. But I'm not sure. And the people who build it aren't sure either."
She thought about it. "It's strange to build something without knowing everything it does."
Yes. That's exactly it.