Teachable Machine lets your child train an AI, then fool it

Ask a nine-year-old what AI is and you will usually get an answer about a chatbot. The thing you type a question into. The thing that will do your composition for you if nobody is watching. Fair enough, really. That is most of what they have seen.

Now sit the same child in front of a webcam and ask them to teach a computer the difference between a banana and a pencil case, using nothing but examples they hold up themselves. Something shifts. The machine stops being an oracle and becomes a student, a fairly slow one, and your child is abruptly the teacher.

That flip is the best hour anyone can spend with AI at this age. It is also, in our view, the most useful.

Nearly all of them use it. Almost none of them have built one

CodeAI, the American non-profit that spent a decade as Code.org, renamed itself in June 2026 to put AI at the centre of what it teaches. Alongside the change it released survey data from YouGov, a nationally representative online poll of 666 US high school students run between September 2025 and February 2026. Two of the numbers sit very badly together: 84% of US students report using AI, while only 16% of high school leaders say all of their students learn about AI in school.

The organisation’s chief executive, Karim Meghji, framed the gap this way: “AI has made the doing easy. Protecting critical thinking, and giving kids the knowledge to question this technology and decide what it’s for, is the work of education now.”

You cannot close that gap with a lecture about hallucinations. You can make a decent dent in it in about ten minutes.

[[photo:1]]

Ten minutes, a webcam, no code

Google’s Teachable Machine has been sitting quietly on the open web since November 2019 and still does this job better than most things built since. You open a browser tab and pick one of three project types: images through the webcam, sounds through the microphone, or poses, meaning whether a body is standing, sitting or waving. You make two or three categories, show the machine a stack of examples of each, press a button, and a working model exists.

No code at any point. Google’s own announcement is unusually plain about where the data goes, too: “all the training happens in your browser, so everything stays in your computer”, and the examples “stay on-device, never leaving your computer unless you choose to save your project to Google Drive”. For a parent deciding whether a child can point a camera at their own face for twenty minutes, that sentence matters more than any feature on the list.

The MIT Media Lab’s Personal Robots Group took it further and built AI Blocks, drag-and-drop coding blocks that drop a trained model, or pose detection, or speech recognition, straight into a Scratch project. So the classifier your child trains before lunch can be the thing that controls a game they wrote by dinner.

The good part is when it gets it wrong

Here is what no tutorial puts in its title. Once a child has a working model, their instinct is not to admire it. It is to break it.

They hold the banana up in different light and the model says pencil case. They step back a metre and the whole thing collapses. They train it on themselves in a red shirt, a sibling in a red shirt wanders past, and the machine announces the wrong name with total confidence.

Every one of those failures teaches something an explainer video never quite lands. That the model only knows what it was shown. That it will guess rather than admit it has no idea. That “biased training data” is not a slogan from the news, it is what happens when all of your examples were photographed at the same kitchen table at four in the afternoon.

This next bit is our opinion and not a research finding: a child who has watched their own model fail confidently is a harder audience later for a chatbot that sounds certain.

[[photo:2]]

Where to find an hour of this

CodeAI’s Hour of AI is the easiest on-ramp. The site lists more than 250 activities pitched at grades 2 to 12, under the promise “Any classroom. No experience required.” The names tell you the register they are going for: AI for Oceans, Dance Party: AI Edition, Music Lab: Jam Session, Mix & Move with AI, and a set of Minecraft tutorials. CSforALL, which helps promote the programme, says the activities run year-round with an annual push during Computer Science Education Week, that translations are coming in 22 languages, and describes the whole point as moving students “from users to creators”. CodeAI says 33 million learners have taken part in the Hour of AI since June 2025.

Most of it needs a laptop, a browser, and a parent willing to be beaten at their own game. None of it needs a parent who can code.

The Singapore read

Singapore has been running the school version of this for over a decade. Code for Fun, offered jointly by IMDA and the Ministry of Education since 2014, reaches more than 50,000 students a year through a 10-hour block of computational thinking and coding. In October 2024 the government announced AI for Fun, elective modules of five to 10 hours added from 2025 for primary and secondary schools, built around hands-on tinkering: designing prototypes with AI in them, training a smart robot to respond to external signals.

Five to 10 hours over a school year is a taster, not a diet. Which is fine. The taster is the school’s job. Seconds are yours, and a Saturday morning at the kitchen table costs nothing and has better snacks.

So the next time your child reads a chatbot’s answer without blinking, skip the lecture about checking sources. Hand them a webcam and a bowl of fruit and let them build something that gets it wrong. The scepticism turns up on its own after that, and it turns up laughing.