AI for Kids
Robots seem very clever. But robots are not magic! They do a few tiny tricks, over and over, super fast.
The Guessing Game
A robot that talks does one tiny trick, over and over. It guesses the next word! Then it guesses the word after that. And after that. It does not know what it is saying. It is just very, very good at guessing what comes next.
Tap any word the robot offers you. Watch it build a silly sentence, one guess at a time.
Pick a word to start your story! 👇
👀 For grown-ups: what is really happening here?👀 For grown-ups
This is a bigram Markov chain built from nursery rhymes. It counts which word followed which, then offers the most common continuations. A large language model is the same idea scaled up enormously: far more context than one word, learned weights instead of raw counts — but still next-token prediction.
What children learn about AI here
These 21 free missions explain AI to kids aged 4–8 by letting them play with small, real versions of it in the browser. Each one earns a single big word, reads aloud for children who cannot read yet, and needs no login. Under every mission is a note for grown-ups saying exactly what the demo really is.
1. The Guessing Game
Big word: Prediction — Guessing what comes next.
A robot that talks does one tiny trick, over and over. It guesses the next word! Then it guesses the word after that. And after that. It does not know what it is saying. It is just very, very good at guessing what comes next.
👀 For grown-ups: what is this demo really doing?
This is a bigram Markov chain built from nursery rhymes. It counts which word followed which, then offers the most common continuations. A large language model is the same idea scaled up enormously: far more context than one word, learned weights instead of raw counts — but still next-token prediction.
2. The Word Chopper
Big word: Tokens — The little pieces a sentence is chopped into.
Robots cannot read letters like you do. First they chop your sentence into little pieces. Then they give every piece a number. Numbers are the only thing a robot really understands!
👀 For grown-ups: what is this demo really doing?
A deliberately naive tokenizer: it splits on whitespace and punctuation, then breaks long words into sub-word chunks. Real tokenizers learn their vocabulary from data (BPE, WordPiece), but the lesson holds — text becomes integer ids before a model ever sees it, and one word is often several tokens.
3. The Giant Story Library
Big word: LLM — A giant robot reader that remixes millions of books.
What is an LLM? It is a robot that read millions of books! It remixes words from all those books to answer your questions. But watch out: when it does not know the answer, it makes up silly daydream stories with a straight face!
👀 For grown-ups: what is this demo really doing?
A Large Language Model (LLM) is pre-trained on billions of words from text corpora. Its "creativity" is controlled by temperature sampling; when hallucinating, the model generates statistically plausible word continuations without grounded factual verification.
4. Agent Miles’s Mission
Big word: Agent — A robot helper with a goal and a backpack of tools.
An AI Agent does not just talk. It has a real job to do! It packs a backpack full of tools, looks around, picks a tool, and then checks what happened. If the tool did not work, it thinks again and tries a different one. Look, think, try, check — round and round until the job is done!
👀 For grown-ups: what is this demo really doing?
An autonomous agent loop (ReAct pattern). Rather than a passive text responder, the agent maintains a goal, inspects environment state, dispatches a tool (keys, bridges, API actions), then observes the result and feeds it back into the next decision. Turn on the storm to watch the loop actually close: the log is tried, it fails, and only that observation triggers the replan onto the balloon.
5. Inside a Robot Brain
Big word: Neuron — One tiny helper in a robot brain.
A robot brain is full of tiny helpers. Each helper looks at one clue and shouts how sure it is. Then all the shouts get added up. If the total is big enough — the robot decides YES!
👀 For grown-ups: what is this demo really doing?
A single perceptron. Each slider is an input, each has a weight (how much that clue matters), the products are summed, and the result is compared to a threshold. Stack millions of these in layers and you have a neural network; training is just adjusting the weights.
6. Teach the Robot
Big word: Training — Showing a robot lots of examples so it can learn.
Nobody tells a robot the rules. You show it examples instead! Show it lots of Zibs and lots of Zogs. Then show it a new creature. It will look for the ones it already knows that look most like the new one.
👀 For grown-ups: what is this demo really doing?
A k-nearest-neighbour classifier over three binary features. The child supplies the labels, so the robot learns their rule — including a wrong or inconsistent one. That is the whole point: the model has no notion of truth, only of the examples it was given.
7. The Puppy Robot: Treats & Stars
Big word: Reward — A star or treat the robot earns when it does a good job.
How does a robot learn to walk or play games? Practice! Every time Puppy Robot avoids mud and finds the golden bone, we give it a star treat. After practicing over and over, the robot figures out the best path all by itself!
👀 For grown-ups: what is this demo really doing?
Reinforcement Learning (RL) with reward shaping. The agent navigates a grid-world via trial and error, updating action-value estimates (Q-values) to maximize cumulative rewards while penalizing negative states (mud).
8. The Pixel Detective
Big word: Pixels — The tiny square tiles that make up every picture.
Robots do not have eyes like you. When a camera looks at your drawing, it sees a mosaic of tiny square tiles called pixels. It turns each tile into a number and slides a magnifying glass across to spot corners and shapes!
👀 For grown-ups: what is this demo really doing?
Computer Vision and Convolutional Neural Networks (CNNs). Images are parsed as numeric matrices. Sliding kernels (convolution filters) aggregate local pixel neighborhoods to detect features like edges, curves, and textures.
9. The Orange Cat Problem
Big word: Bias — When a robot only learns about some things, not all things.
Watch what happens when a robot only ever sees ORANGE cats. When a black cat comes along, the robot gets confused! It is not being naughty. Nobody ever showed it a black cat. Robots only know what we show them.
👀 For grown-ups: what is this demo really doing?
Dataset bias, made visible. The classifier is trained on examples that all share an irrelevant feature (colour), so it latches onto that feature instead of the one that matters. Retraining with a fuller set fixes it.
10. Real or Made?
Big word: Made-up Picture — A picture a computer built that never really happened.
Some pictures are photos of real things. Some are drawn by people. And some are made by a computer that mixed up millions of other pictures! Computers sometimes slip up — too many fingers, or writing that is all jumbled. Let us hunt for the slip-ups.
👀 For grown-ups: what is this demo really doing?
Generative image models. A diffusion model learns the statistics of what pictures look like, not the rules of what things are — which is why it renders a plausible hand with six fingers and signage as letter-shaped noise. Worth being straight with children about the ending: these artefacts are disappearing fast, so the durable skill is provenance ("who posted this?"), not artefact-spotting.
11. The Heart & Hug Test
Big word: Human Touch — The feelings, kindness, and love that only real people have.
Robots are super fast math machines, but you have superpowers no computer will ever have! Robots cannot feel love, taste yummy cake, or give warm hugs. You are always the boss of the robot!
👀 For grown-ups: what is this demo really doing?
Human-in-the-loop and alignment ethics. While machines surpass humans in calculation speed and data scale, human beings retain consciousness, subjective experience, empathy, and moral agency.
12. The Robot’s Word Map
Big word: Word Map — A map in the robot’s brain where words with the same meaning live side by side.
A robot that reads keeps a giant map of words in its brain. It is not a map of letters — it is a map of what words MEAN! Words that mean similar things live right next door: dog, puppy, and cat are neighbours. Words that mean very different things live far, far apart.
👀 For grown-ups: what is this demo really doing?
Text embeddings. A real model maps words to vectors in a high-dimensional space learned from co-occurrence, and nearby vectors mean nearby meaning. This demo is a tiny space of hand-chosen features scored with plain cosine similarity — the same math word2vec-style models use, scaled down to a handful of dimensions a child can feel.
13. How the Robot Hears
Big word: Speech — Sounds that become ideas — a voice turning into words, or words into a voice.
When you talk to a robot, your voice is really just wiggly air that a tiny microphone scraps into numbers. The robot chops the sound into the shortest possible pieces, compares each piece with millions of sounds it has practised, and says the word it thinks it heard. Talking and listening are two different robot tricks!
👀 For grown-ups: what is this demo really doing?
Automatic speech recognition (ASR): the waveform is sampled, framed, and mapped to text by models trained on huge audio corpora; text-to-speech (TTS) runs the other way, text to waveform. Both reduce the sound to a sequence of predictions over a vocabulary — the demo uses the browser’s on-device speech synthesis, so the "voice" the child hears is a real speech model running locally, consistently with the site’s nothing-leaves-the-device promise.
14. The Guess-It Machine
Big word: Decision Tree — A tree of yes-or-no questions that leads the robot to one answer.
To find one thing among many, the robot never searches one by one. It asks a question that cuts the pile in half: water or land? Then half again: fur or no fur? Each answer throws away half the possibilities, until only one is left. A chain of questions like this is called a decision tree!
👀 For grown-ups: what is this demo really doing?
A decision tree recursively partitions the candidate set on the single most informative attribute at each node. The demo grows the tree at runtime by always picking the first question whose yes/no truly splits the remaining animals, so the child is watching real supervised classification — pruned 20 Questions, the same idea as ID3/CART, and the seed of random forests.
15. The Show Picker
Big word: Recommendation — When the robot guesses what you will like next, from what you liked before.
When you love a show, the robot writes it down. Then it looks for shows that are MOST like the ones you loved — same space rockets, same funny jokes. It even slips in one brand-new kind of show to learn more about you. That is how a robot decides what you should watch next!
👀 For grown-ups: what is this demo really doing?
A recommender reduced to content features: each item is a tag vector, love actions update the taste profile, ranking is similarity (mean Jaccard) between each candidate and the loved set, and the deliberate off-profile pick is exploration — the explore/exploit trade-off every recommendation system lives by.
16. Safe or Not Safe?
Big word: Privacy — Keeping your private things to yourself — and deciding who you trust.
A robot has a giant memory. When you type something secret to a robot, it might remember it forever — and a robot cannot promise to keep your secrets. So there are things that are fine to tell a robot (your favourite colour!) and things you should never tell it (your name, your house, your passwords). When in doubt, ask a grown-up first!
👀 For grown-ups: what is this demo really doing?
Data minimization and consent taught as a rehearsed habit: personal information (name, address, school, photos, credentials) should never leave the child’s informed circle. Labels are deliberately strict because the durable skill is the pause before sharing anything identifying. This app transmits nothing by design, making the rule literal here.
17. The Magic Words
Big word: Prompt — The words you give a robot to tell it what to do.
A robot cannot see inside your head. It only knows the words YOU give it. If you say "draw me a robot", the robot has to guess! But if you say "draw me a red round robot with a hat", the robot knows exactly what to do. Words are how you boss a robot around!
👀 For grown-ups: what is this demo really doing?
Prompting as instruction-following. A model conditions on the literal tokens supplied, so an underspecified prompt leaves the output distribution wide — many plausible robots — while naming the distinguishing attributes narrows it to one. The demo is honest about its mechanism: literal feature filtering over a fixed gallery. The durable skill is decomposition, not incantation: say the attributes that actually tell the options apart.
18. Robot Says... Is That True?
Big word: Hallucination — When a robot makes something up and thinks it is true.
Robots sound SO sure when they talk. But sometimes a robot makes things up — and it does not even know it did! A made-up answer from a robot is called a hallucination. The tricky part is that a wrong answer sounds exactly like a right one. So we check with a grown-up or a real book.
👀 For grown-ups: what is this demo really doing?
Hallucination in large language models. Generation is optimised for the statistically plausible next token, not for truth, and the model exposes no reliable internal signal of its own uncertainty — a fabricated claim is emitted as fluently as a fact. The demo pins the confidence meter at 100% for every claim on purpose: the lesson is calibration, not confidence. Children must learn that a sure voice is not evidence.
19. The Attention Spotlight
Big word: Transformer — A super-brain that connects words across a whole sentence using an attention spotlight.
Old robots read sentences like sleepy snails, one word at a time, and forgot the beginning before they reached the end! A Transformer robot is different: it looks at the whole sentence at once. It shines a magic attention spotlight that draws glowing lines between words that belong together!
👀 For grown-ups: what is this demo really doing?
The Transformer architecture and Self-Attention mechanism (Vaswani et al., 2017). Unlike recurrent networks (RNNs) that process tokens sequentially, Transformers process all tokens simultaneously using multi-head self-attention. This allows the model to compute pairwise attention weights across arbitrarily long distances, resolving coreferences and contextual semantics.
20. Code the Robot Brain
Big word: Code Recipe — Step-by-step instructions written by humans that tell a computer what to do.
AI does not magically know what to do. Real humans write recipes called code to tell it how to listen, think, and react! When you snap code blocks together, you give the robot instructions. You are the director, and the robot is the actor!
👀 For grown-ups: what is this demo really doing?
Algorithmic thinking and rule-based conditioning in AI systems. Before or alongside neural weights, software engineers write scripts that define event listeners (triggers), validation thresholds (confidence checks), and action handlers. Coding empowers children to understand that AI is a tool engineered and controlled by humans.
21. The AI Safety Superhero
Big word: Safety Shield — Smart rules that protect your secrets and keep you in charge of the computer.
AI is a super-fast helper, but every hero needs a shield! Remember the four golden rules: keep your private secrets safe, check if robot facts are real, use your own brain for schoolwork, and tell a grown-up if anything feels weird. Wear your superhero cape and stay in charge!
👀 For grown-ups: what is this demo really doing?
AI literacy, digital citizenship, and safety heuristics for early learners. Children learn four protective strategies: data minimization (privacy), hallucination verification (truth checking), cognitive independence (academic integrity), and human escalation (safeguarding against distressing outputs).
Want more? Build a robot of your own in Train Your Robot, hear the story of how AI was invented, or find lesson plans and printables on the parents and teachers page.