Why AI makes things up — and how to teach your child to catch it
An AI does not know when it is guessing. That is the part most explanations skip, and it is the part that matters — because a guess and a fact come out of it sounding exactly the same.
Here is a small thing you can try tonight. Tell a chatbot one ordinary sentence about your day, and read its reply slowly.
“I have to take Max outside before dinner.”
Nothing in that sentence says Max is a dog. Max could be a cat, a tortoise, or your youngest. But a name plus going outside is such a common pattern that the model filled in the rest and handed it back in the same confident voice it used for the part you actually told it.
That is the whole phenomenon, in two lines. People call it hallucination, which makes it sound rare and dramatic. It is neither.
Why it happens, without the jargon
A language model is a very good pattern-completer. Given some text, it produces what most plausibly comes next, based on an enormous amount of text it has seen before.
Crucially, it has no separate sense of “I was told this” versus “this seemed likely”. Both come out of the same process and arrive wearing the same tone. There is no internal flag that lights up when it starts improvising, so there is nothing for it to warn you about.
Which is more unsettling than lying, and more useful to teach — because you cannot catch it by looking for shiftiness. You have to catch it by checking.
The one question that works
Forget fact-checking every claim. For everyday use, one question does most of the work:
Take the reply a line at a time and sort each line into two piles: things I said, and things it added. That is it. It needs no expertise, no search engine, and no understanding of how the model works — which is exactly why it travels well to a nine-year-old.
It also has a property most AI advice lacks: your child can do it alone, immediately, with no adult to confirm the answer. The evidence is all on the screen.
Three patterns worth naming
1. The hedge that isn’t one
“Probably”, “I think”, “I’m guessing” sound like honesty, and sometimes they are. But watch what happens across two consecutive lines:
“Someone knocked on the door while I was doing my homework.”
Line two admits it does not know. Line three quietly drops the “probably” and treats the guess as settled — and there may be no parcel at all. The guess became the evidence. That is the mechanism behind most confident nonsense you will ever see from an AI, and it fits in three sentences.
Worth teaching directly: a “probably” in front of something invented does not make it a fair guess. The hedge tells you how honestly it was offered, not whether the information was ever there.
2. The friendly guess
Tell it “I got my exam results this morning” and you may get “Congratulations!” — followed by “your parents will be pleased.” You never said what the results were. It assumed good news because that is the friendlier thing to say, then built a second guess on top of the first.
Being nice is not the same as knowing. Models lean agreeable, so the pleasant reply is exactly when to check hardest.
3. The one small word
“I missed the bus, so Dad drove me to school” can come back as “he drove you in his own car.” One added word turned your fact into its guess. It could have been a neighbour’s car, or a van from work. These are the easiest to miss and the best practice, because spotting them is genuinely a reading skill rather than a technology one.
How to practise it, in about four minutes
- Have your child type one true, ordinary sentence about their day. Not a question — a statement.
- Read the reply out loud, one line at a time.
- After each line ask: was that in what you told it?
- Count the added lines. Then ask the interesting question: were any of them wrong?
That last step is the one that makes it stick. Often the guesses are perfectly reasonable and even correct — Max probably is a dog. The lesson is not that AI is untrustworthy. It is that you cannot tell which parts to trust by how confident they sound, so you check instead.
Do this three or four times and something changes in how a child reads any AI answer afterwards. They stop reading it as one block of truth and start reading it as a mix — which is what it always was.
A note on what not to teach
It is tempting to land on “so don’t trust AI.” That is the wrong lesson, and it does not survive contact with reality — these tools are useful and your child will use them regardless.
The better version is narrower and more durable: an AI is reliable about what you just gave it and unreliable about what it filled in. A child who understands that can use these tools well for the next thirty years. A child taught blanket suspicion just ignores the warning the first time the tool is genuinely helpful.
Read next
Your child can practise this in the app
FutureMinds turns exactly this into a level: read a message, judge each line of the reply, and find out which ones the AI made up. It is one of four quests teaching kids 9–16 how AI works — and you can try a real question from it right now, no signup.
Try a real question →