How does AI work? Explained for kids (and everyone else)

The one idea underneath all of it: learning from examples

Ordinary software follows rules a person wrote. If a bank program needs to spot an overdraft, someone typed the rule: if the balance is below zero, flag it. That approach falls apart the moment the job is something nobody can write rules for, like telling a cat from a dog in a photograph. Try it - every rule you come up with has exceptions.

Machine learning does something different. Instead of rules, you supply examples that are already labelled with the right answer, and the system works out for itself what separates them. Nobody tells it to look at ear shape or nose length. It finds whatever happens to be reliably different, which is both the reason it works and the reason it sometimes learns something absurd.

This is worth being precise about with a child, because the popular explanation - the computer has a brain and learns like you do - makes everything that follows harder to understand. It does not understand anything. It has found a pattern that worked on the examples it was given.

Why AI is sometimes biased or unfair

If the examples are one-sided, the pattern will be too, and the system has no way to know. A famous real case: a recruiting tool trained on a company's past hiring decisions learned to downgrade CVs that mentioned women's activities, because the historical data it learned from reflected who had been hired before.

There is nothing mysterious in that. The system did exactly what it was asked to do - find the pattern in the examples - and the examples carried the bias. This is why the phrase to teach is not 'the AI was unfair' but 'the training data was one-sided'.

The fastest way to make this land is to let someone cause it themselves. In Train the Bot you personally choose which creatures the bot learns from. Pick five examples that all happen to share a feature and the bot will confidently decide that feature is the answer, then fail on everything new. The bias is visibly your doing, and fixing it is a matter of choosing examples that vary.

• The model finds whatever separates your examples, not what you meant
• One-sided examples produce a one-sided rule, every time
• It cannot tell you this has happened - it looks like success on the training data
• The fix is better examples, not a cleverer model

How ChatGPT and other chatbots actually work

A language model is doing one thing over and over: given the text so far, how likely is each possible next word? It picks one, adds it to the text, and asks again. Everything that looks like conversation, reasoning or personality is built out of that single repeated step.

It follows that a chatbot has no separate store of facts to check against. Nothing in it knows whether a sentence is true - it knows whether a sentence is the kind of thing that usually gets written. A plausible-sounding wrong answer scores exactly as well by that measure as a right one, which is the whole explanation for why models invent citations, dates and names with total confidence.

Next Word puts a person in that seat. You see a half-finished line, pick the word you think comes next, and then see roughly how a model would rate each option. Five minutes of it and the difference between predicting language and knowing things stops being an abstraction.

What a neural network actually is

Under the hood, a neural network is a large pile of extremely small pieces of arithmetic. One artificial neuron takes some numbers in, multiplies each by a weight, adds them up with a bias, and fires if the total clears zero. That is the entire neuron.

Training the network means adjusting those weights until the outputs come out right. There is no other magic step. A large model has billions of weights, but each one is doing the job you can do by hand with three sliders in Neuron Lab.

That game also demonstrates the limit that forced the field forward. One neuron can only draw one straight line, so there are patterns it can never separate no matter how the weights are set. Stacking neurons into layers is what lets the boundary bend, and that is the whole reason modern systems are called deep learning.

What to actually teach a child about AI

Three things, in this order. First, it learns from examples, so the examples decide what it learns. Second, it predicts rather than knows, so it can be fluent and wrong at the same time. Third, it is worth checking anything that matters, for the same reason you would check a very confident stranger.

That is enough to use these tools sensibly, and it is a great deal more useful than a list of rules about when not to use them. Children who understand why a chatbot invents things tend to catch it doing so.