What's really going on inside a chatbot?
Have you ever watched your child type a question into a chatbot and get a smooth, confident answer back in seconds? It can feel like magic. It can even feel like there's someone in there.
There isn't. In this lesson we'll lift the lid together, so that you can explain it to your child in words they'll actually remember.
The world's biggest game of "finish the sentence"
Try this with your child. Say, "Twinkle, twinkle, little..." and stop. They'll shout "star!" straight away.
Why? Because they've heard that pattern loads of times. They didn't look it up. They just knew what usually comes next.
That's the heart of a chatbot. It's a type of AI called a large language model. Before anyone ever used it, it was trained on a huge amount of text: websites, books, articles and much more. From all that reading it picked up patterns in how words tend to follow one another.
So when you ask it something, it doesn't go and fetch a stored answer. It predicts a likely next bit of a word, adds it, then predicts the next bit, and the next. It builds the whole reply piece by piece, a bit like laying stepping stones one at a time.
Here's a child-friendly way to put it: "A chatbot is like a super-powered version of the predictive text on a phone. It's read so much that it's brilliant at guessing which words should come next."
So why does it sound so clever?
Good question. If it's "just" guessing words, how can it write a poem or explain photosynthesis?
The answer is scale. Because it has learned from so much writing, its patterns carry a lot of grammar, facts, styles and ways of explaining. So its guesses are often really good.
But notice the word often. Good guesses are still guesses.
Why chatbots make things up
Here's the part every family needs to know. A chatbot is built to produce answers that sound right. Sounding right and being right aren't the same thing.
When it doesn't really know something, it may fill the gap anyway, producing a fact, a date, a quote or a book title that doesn't exist. People call this "hallucinating".
And it'll say it in exactly the same confident voice it uses for true things. There's no wobble in its voice and no "um". That's what makes it tricky.
Why does it guess rather than say "I don't know"? Researchers at OpenAI, the company behind ChatGPT, published a paper about this in September 2025. They compared it to a pupil doing a multiple-choice test. If a blank answer scores zero, a lucky guess might score a point, so guessing gets rewarded. Chatbots have been trained in a similar way, so they've learned to have a go.
This isn't a small, rare glitch. In a well-known 2023 case in New York, lawyers sent a court a document that ChatGPT had helped write. It listed earlier court cases that turned out not to exist. The judge fined the lawyers $5,000. If grown-up professionals can be caught out, so can a 12-year-old doing a history project.
It's also why chatbots like ChatGPT show a small warning on screen saying they can make mistakes and that you should check important information. Point that line out to your child next time you see it.
It doesn't think or feel like a person
Now for the gentler, but just as important, bit.
A chatbot might say "I'd love to help!" or "That sounds hard, I'm sorry." Those are words that fit the pattern of a kind conversation. They aren't feelings.
It doesn't know your child. It doesn't care about them, get tired, or have a favourite colour. It has no life outside the chat. It's a very clever word-predicting tool, not a friend.
Younger children especially can blur this line. So it's worth saying plainly: "It can sound friendly, but there's nobody in there."
Conversation starters for different ages
Ages 5-8: "Can you finish this? 'Once upon a...' How did you know? A chatbot does that too, but with every word!"
Ages 9-12: "If a chatbot is guessing the next word, what might happen when it doesn't know the answer?"
Teens: "Why do you think a chatbot would rather make something up than say 'I don't know'? How would you catch it out?"
You don't need perfect answers yourself. Wondering about it together is the whole point.
Family action: the chatbot challenge
This week, sit down together with a chatbot and test it like detectives. Here's a simple order to follow:
- Pick questions you can check. Mix easy facts ("What's the capital of France?") with things only your family knows ("What's our dog called?") and something niche, like a local landmark or a lesser-known children's book.
- Predict first. Before you hit enter, ask everyone: "Will it get this right?"
- Ask, then check. Compare the answers with a trusted book, an official website, or your own family knowledge.
- Talk it through. What did it get right? What did it get wrong, and did it sound any less sure when it was wrong?
One more tip: try asking it the same question twice. Sometimes you'll get a different answer, which is a great talking point all by itself.
You might be surprised how quickly your child becomes a sharp-eyed checker. That habit is worth more than any single rule.
Key takeaways
- Chatbots are trained on huge amounts of text and build answers by predicting the next bit of a word, over and over.
- They don't look up stored answers. They produce what sounds likely.
- They can make things up ("hallucinate") and sound just as confident as when they're right.
- They can sound warm and friendly, but they don't think or feel like a person.
- Checking against trusted sources is the habit that matters most, and a family experiment is a fun way to build it.