Published on March 22, 2026
Last week I went through a pretty stressful experience: I taught AI to my daughter's 2nd grade class — twenty 8-year-olds, 45 minutes, a laptop and a webcam. The only time I've been this stressed in front of kids was when I visited her preschool class to talk about Kung Fu.
I had an AI agent running on my phone that could update a class website in real-time — that turned out to be the secret weapon. I'd send voice notes to my agent via Telegram, and it would update the projected site a few seconds later.
It went pretty well. The kids were engaged, they had fun, and I think they actually learned something. I'm sharing how it went so that other parents who want to do something similar have a starting point.
I started with a question: Can a piece of paper be intelligent?
Then I pulled out a folded piece of paper with a tic-tac-toe algorithm on it — a printout from CS Unplugged's "Intelligent Paper" activity (PDF). I challenged one of the kids to play against it. My daughter Janka was my assistant — she followed the algorithm on the paper, making the moves it dictated. The paper never loses — against a strong opponent it can draw, and this time the kid played well enough that it ended in a draw.
The point landed: intelligence doesn't require a brain — it requires the right set of rules. A computer program that follows rules can be intelligent too. That's AI.
I sent a voice note to my agent summarizing what the kids said. Seconds later, their thoughts about AI appeared on the projected website.
The paper tic-tac-toe works because someone wrote clever rules. But what about something more complex — like recognizing whether a hand is showing rock, paper, or scissors? Try writing rules for that. It's basically impossible.
So I told the kids: this is where we need a different kind of AI. Instead of writing rules by hand, we show examples and let the computer figure out the rules on its own. That's machine learning.
I opened Google Teachable Machine on the projector. We created three classes: Rock, Paper, Scissors. Kids lined up to show their gestures to the webcam. We trained the model and tested it.
And then something great happened: the model kept getting Paper wrong. The kids figured out why on their own — there weren't enough examples of Paper. So we added more samples, retrained, and it worked. They had just done their first iteration of the ML development cycle without anyone telling them what that was.
I told them: "This is actually what I do at work every day. I train models, I see what's not working, I figure out what data to add, and I train a better model. Exactly what you just did."
I exported the model, sent the URL to my agent, and it deployed the classifier to our class website. The kids could show it to their parents at home — open the site and play Rock-Paper-Scissors against the model they'd built together.
I asked the kids: what is a picture for a computer? They came up with pixels — tiny dots that make up the image. I showed them a necklace with colored beads as an analogy: each bead is a pixel, each pixel is just a number.
Then: a neuron takes those numbers and does multiplication and addition. That's it. Math they already know. One neuron isn't smart enough, but connect thousands of them and you get a neural network — that's what just recognized their gestures.
Then I asked ChatGPT live: "Is this how you work? Multiplication and addition?" It confirmed: "Yes, essentially — a lot of multiplications and additions, but that's fundamentally how I work."
I asked the kids where AI helps people. We tried to come up with good use cases together. Janka shared that she'd used ChatGPT to help her bake cookies — it gave her a recipe. I also tried to show how AI can be used for work and for education.
But the answer the kids gravitated to naturally: for fun. AI can draw things, make music, create games. For eight-year-olds, that's the killer app. Honestly? That's a perfectly valid answer.
I'd prepared three images — not generic "AI risks" slides, but metaphors the kids already knew:
Pinocchio. AI can lie. Not on purpose, but it can confidently say things that aren't true. Just like Pinocchio — the words sound right, but they might be completely made up. Not everything AI tells you is real.
The Mirror of Erised from Harry Potter. The mirror shows you what you want to see, not what's true. AI tends to tell you what you want to hear. If you ask it "is my idea great?", it'll probably say yes. That doesn't mean it is. This is called sycophancy, and it's one of the trickiest problems in AI right now.
The Secret Garden. There's a boy in the story who's in a wheelchair — not because his legs are broken, but because he never uses them. When he finally starts walking, his legs get stronger. Your brain works the same way. If you let ChatGPT do all your homework, do all your thinking — your brain won't get stronger. It'll get weaker. Use AI as a tool, but do your own thinking.
I showed the images and asked the kids what they thought each one meant. They figured it out themselves. Then I gave them homework: "Show these pictures to your parents tonight. See if THEY can figure out what each one means."
Making kids the teachers of AI safety to their own parents — that sticks.
This was the highlight. I handed out pieces of paper and pencils: "Write what you want AI to draw."
Some of the prompts:
I photographed each kid's handwritten note, sent the photo to my agent on Telegram, and within seconds the AI-generated image appeared on the website. The handwriting transcription step alone amazed them — "It can read my writing?!"
By the end we had 19 images in the gallery. The last prompt was collective: "Nasza klasa — złoty pył i kolorowe motyle" (Our class — golden dust and colorful butterflies).
Before the kids left, each one got a handout: the tic-tac-toe paper algorithm — that same "intelligent piece of paper" from the beginning — plus a QR code linking to our website. I told them to show their parents the website when they got home.
Twenty kids walked out with a physical piece of AI in their backpacks and a URL to a website they'd built together.
The tech stack is deliberately simple:
I've packaged everything into a reusable, open-source skill:
ai-lesson-for-kids in the ai-hotpot repo includes: - Full setup instructions - The HTML website template - Image generation script - Detailed lesson plan with timing and tips
You need: a Netlify account (free), a Gemini API key, a laptop with a webcam, paper, markers, and 45 minutes.
If you run this lesson in your kid's class, I'd love to hear about it.