Home Learn Class 9
Classes 4–12 · Digital literacy and AI

How does an AI model actually work?

There is one sentence that explains almost everything an AI does, well and badly: <b>it predicts what comes next.</b> Not what is true. Not what you asked. What is most likely to follow the words in front of it.

It continues, it does not answer

Give a model the words "The best thing about a rainy day is" and it does not think about rain. It works out which word most often follows those words, picks one, adds it, and then does the whole thing again with the new sentence.

One word at a time, thousands of times a second. That is the entire machine.

Most of the time, continuing a question well is the same as answering it. Most of the time.

Where the patterns came from

Before you ever used it, the model read an enormous quantity of text and adjusted itself so its guesses got better. That is training, and it finished months before you opened the app.

Nothing was stored. You cannot ask it for page 41 of something it read, because there are no pages in there — only the shapes those pages had in common.

Why this explains the failures

It sounds confident when it is guessing, because guessing is all it ever does — a 55% guess and a 99% guess come out in exactly the same voice.

It fills blanks, because a blank is a thing to continue.

It cannot do arithmetic reliably, because it is predicting what a number looks like rather than working it out.

Watch one sentence get built

The best thing about a rainy day isthe (44%) · that (31%) · not (25%)
…is thesmell (48%) · sound (34%) · way (18%)
…is the smellof (81%) · that (11%) · coming (8%)
Result"…is the smell of wet earth."

Nobody decided that sentence. It was assembled, one most-likely word at a time, by something that has never smelt rain.

Practice

Does the model know what rain is?
Show the answer
No.
It knows which words tend to follow which other words. That is not the same as knowing anything, though it can look identical from the outside.
Why does it give a different answer if you ask twice?
Show the answer
It picks one likely word out of several.
There are usually many good continuations. It samples one. Ask again and it may sample a different one — both perfectly reasonable.
Can it tell you where an answer came from?
Show the answer
Not from its own memory.
There is no index inside a model to point at. If you need a source, you have to give it the document to read.

Be the model for a minute

In our free course you build a sentence by clicking the most likely next word, with the real percentages showing. It takes ten seconds and you never forget it.

▶ Play Lemonade Empire free

Common questions

Is it just copying text it has seen?
No. It is generating new text using patterns learnt from what it read — which is why it can write a sentence nobody has ever written, and also why it can confidently write one that is not true.
Does it get better while I use it?
No. Training finished before you met it. Your conversation does not change the model.

Next concepts