How to build an AI that can tell rock from paper from scissors

Here is what you can have by the end of a Saturday morning: a rock-paper-scissors game that does not know what rock, paper or scissors look like until your child teaches it. Not a game you download. A game your child builds, trains and can watch get things wrong.

The tool is free, it is called Machine Learning for Kids, and a volunteer developer at IBM named Dale Lane built it for schools near his home in the UK before it was opened up to everyone. IBM now maintains it as an open-source project on GitHub, and AI4K12, the US National Science Foundation-backed initiative that sets AI curriculum guidance for American schools, lists it as a resource for “elementary and up”. Around two dozen projects sit behind it, each with its own worksheet. Rock-paper-scissors is the one worth starting with, because within about twenty minutes your child has something that visibly works, or visibly doesn’t, which turns out to be the more interesting outcome.

What actually happens

A child creates a free project on the site and tells it they want to teach it to recognise “images”. They open the Train tab, make a label called “rock”, and hold up a fist to the webcam ten times from slightly different angles and distances. Then a label called “paper”, ten more photos, flat hand. Then “scissors”, ten photos, two fingers. Thirty photos in total, all taken by the child, none of them downloaded from anywhere.

The site’s own worksheet is blunt about why this matters more than it sounds: “if you have a lot of examples for one type, and not the other, the computer might learn that type is more likely, so you’ll affect the way that it learns to recognise photos.” That line is doing more teaching than the whole rest of the activity. It is the honest, hands-on version of a sentence every parent has now half-heard about AI: that a model only knows what it was shown, and uneven examples produce an uneven, overconfident machine. Here, a child can watch it happen to something they built themselves, in real time, with their own hand.

Training it, then testing it

Once the photos are in, the child clicks “Learn & Test” and then “Train new machine learning model”. Training takes a few minutes, during which the worksheet suggests a multiple-choice quiz to fill the gap, a small kindness to anyone who has sat a ten-year-old in front of a progress bar. What comes out the other end is a genuine, if small, machine learning model: a piece of software that has learned, from thirty photographs, to sort a new photograph into one of three buckets.

It will not be very good yet. That is the point of the next step, not a flaw in it. If the model keeps calling scissors “rock”, the fix is not a settings menu, it is going back and adding more scissors photos from more angles. The worksheet says so directly, and says it as a normal, expected step rather than a failure: more examples, more variety in position and size, try again.

Wiring it into a game

The last phase is called “Make”. The child picks Scratch, opens a ready-made rock-paper-scissors project template that already has a scoreboard and a computer opponent built in, and drops a single new block into it: “recognise image”, wired to a photo from the webcam. That block is the trained model, sitting inside an ordinary Scratch script. Click the green flag, hold up a hand shape, snap a photo, and the game tells you whether you won, using a model your child trained an hour earlier on their own fist.

It is a small project. It is also, as far as we can tell, a genuinely rare thing to hand a child: a complete loop, in one sitting, from “I showed it examples” to “now it plays a game with me”, with every step visible and nothing happening inside a black box.

The Singapore version of the same idea

Singapore’s schools already run a version of this. The Ministry of Digital Development and Information said in 2024 that new “AI for Fun” modules, five to ten hours long, would be added from 2025 to the existing Code for Fun programme run jointly by IMDA and the Ministry of Education, reaching more than 50,000 students a year. The release describes the activities as “hands-on exploration and ‘tinkering’ with technology, such as through the design of prototypes incorporating artificial intelligence”, and gives one example directly: students “understand what a smart robot is and train such robots to respond to external signals”. That is the same idea as the rock-paper-scissors project, run by a teacher, during school hours, for five to ten hours a year.

A Saturday morning at the kitchen table is not a replacement for that. It is seconds. But it is a chance to watch your own child do the training themselves, ask why the model keeps confusing their sister’s paper for a rock, and answer it together instead of waiting for term to come round again.