Home › Blog › How does ChatGPT work?
How does ChatGPT actually work?
No maths, no jargon, and nothing you need a technical background to follow. If you have used a phone keyboard, you already know most of it.
Plenty of people are being told to worry about AI before anyone has told them what it is. This is that missing explanation, and it takes about six minutes.
Start with your phone keyboard
Type “I’ll be home in” into a text message and look at what your phone offers you above the keys. Probably a, five, ten. It has no idea when you will be home. It has noticed that after those words, those words usually come next.
That is the entire idea behind ChatGPT. Not similar to it. The same idea.
Try it now. Open a message on your phone and type “Thanks so much for”. Whatever your keyboard suggests, it is guessing from patterns — not from knowing anything about your life.
Now make it enormous
Your keyboard has looked at a little of your typing. ChatGPT has been shown a staggering quantity of written text — books, websites, articles, conversations — and has worked out, in extraordinary detail, which words tend to follow which.
Give it “Instead of turning left, she turned…” and it knows right is likely. Give it “explain photosynthesis to a ten-year-old” and it produces a paragraph, one word at a time, each chosen because it fits what came before.
It is doing that all the way to the full stop. There is no moment where it looks something up, and no moment where it decides whether the answer is true.
So where did it get all that from?
From text that already existed, gathered up to a particular date. That date is called the cut-off, and it matters more than it sounds.
Nothing after the cut-off is in there. A shop that changed its hours last month, a rule that changed last week, the result of yesterday’s match — unless the tool is also searching the web, it has no way of knowing. It will still answer, in the same steady voice as everything else.
Why does it sound so sure of itself?
Because it was built from writing, and confident writing is what most writing looks like. Nobody publishes “the Battle of Hastings was in 1066, though I am honestly not certain”.
So the tone you are reading is a property of the writing, not of the answer. A sentence it has good reason for and a sentence it produced because the shape felt right arrive looking identical. There is no hesitation in the voice, because there is no voice — there is a prediction of what a confident sentence looks like.
This is the sentence worth remembering. How certain an answer sounds tells you nothing at all about how certain it is.
Four things it is not doing
It is not looking things up. Search finds pages that already exist and hands you the links. This writes a new sentence. That is why it can produce a book quotation, a page number or a study author that sounds perfectly real and simply is not.
It is not checking. Nothing between your question and its answer verifies anything.
It is not understanding you in the way another person would. It is very good at producing the words that fit — which is not the same thing, and is easy to forget because the result reads so much like someone who does.
It is not remembering you, unless you have switched that on. Most conversations start from nothing.
None of which makes it useless
Predicting words that fit turns out to be enormously useful. Explain this simply. Give me an example. Rewrite this so it is shorter. Help me draft an awkward email. What might I be forgetting? It is genuinely good at all of that, and a person who refuses to touch it is giving up something real.
The useful habit is not suspicion. It is one question, asked only when the answer matters:
Don’t ask only “What does it say?”
Ask “How would it know?”
If the answer is “it read something and can show me”, that is checkable. If the answer is “it wrote a sentence that fits”, that is a starting point — not a finish.
Ask it to explain a diagnosis and you are using it well. Ask it whether to take two medicines together and you are asking a pattern-matcher a question that needs a pharmacist.
Read next
- ChatGPT vs Google: what parents and teachers should know
- Why AI makes things up — and how to catch it
Was this useful?
Would you want your child to know this?
You have just learned what the thing is actually doing. FutureMinds teaches children 9–16 the same idea — by playing, not by being warned — so they meet it before they need it. Launching 15 September.
Take the 2-minute Challenge →
Join the conversation
Comments are read by a person before they appear — usually within a day or two. Nothing is published automatically.