Explaining ChatGPT to My Sons: Very Good Guessing, Not Thinking

Archael asked me yesterday what I do for work. I told him I help computers think. He looked at me like I was pulling his leg. “Dad, computers can’t think.”

Smart kid. That exchange is why I want to explain AI to my kids properly now. They’re going to use these tools whether I like it or not. The thing I can still shape is how far they trust them.

The demo

“So what do you actually do?”

I opened ChatGPT and typed: “write a story about a robot who wants to learn to paint.” His eyes widened as the response appeared, sentence by sentence.

“Is that real? Did a person write that?”

“Sort of. A really smart computer program looked at millions of stories and learned patterns. Then it used those patterns to make something new.”

“But computers just follow rules, right? Like in my game?”

I never expected to have this conversation with a seven-year-old. It’s a fair question, though. Everything else he knows about computers follows rules someone wrote down. This one doesn’t, at least not in a way he can see.

Guessing, not knowing

I tried to explain it without the jargon:

“You know how you learn to draw? You see lots of pictures, you practise, you get better?”

He nodded.

“The computer does something similar. It ‘reads’ millions of books and learns how words usually go together. When you ask it something, it predicts which words should come next, based on what it learned.”

“So it’s guessing?”

“Very educated guessing, yes.”

That’s the most useful sentence I’ve found for explaining a large language model to a child, and honestly for some adults too. It’s a simplification, but an honest one. It also sets the right expectation. Educated guesses are usually good, sometimes wrong, and they sound the same either way.

Archael landed on his own version, which beats mine: “So it’s like autocomplete on my iPad, but really, really good?”

Yes. Exactly that.

Andriel’s questions

The interesting part came when Andriel joined us. He’s 10, on the autism spectrum, and processes information differently.

I showed him the same demo. He asked something else: “Does it understand what it’s writing?”

That question is still argued about by people who build these systems for a living. I gave him my honest answer.

“I don’t think so. It’s really good at predicting what words should come next, but understanding is different.”

Andriel thought about it. “Like when I memorise things but don’t always know what they mean?”

He’d described the gap between sounding fluent and actually understanding better than most explainers I’ve read.

Later he came back with another one: “If it learns from what people write, does it learn bad things too?”

He’s 10 and asking about bias and training data. I love it. The answer is yes. If the stories, articles and forum posts it learned from contain mistakes, unfairness or nastiness, some of that can show up in what it says. The people who build these systems try to filter that out. They don’t catch everything.

What I’m actually teaching them

Not transformers or embeddings. They can learn that later if they want to. What matters at seven and 10 is a handful of ideas:

Confident isn’t the same as correct. A chatbot answers in the same calm, fluent voice whether it’s right or making something up. Checking an answer is part of using the tool.

Use it like a calculator. Pick it up for a specific job, check the result, put it down.

Human creativity matters. The AI can write a story about a robot painter, but it doesn’t care about the story. You do, and that matters.

Good questions get good answers. What you get out depends on what you ask. Asking a clear, specific question is a skill, and it’s worth practising.

A person is responsible. People built these systems, people decide how they’re used, and the person using one is responsible for what they do with its output.

Where the guardrails sit

There’s a practical side too, and it doesn’t come up much when tech people get excited about their kids “getting into AI”.

ChatGPT isn’t designed for young children. OpenAI’s Terms of Use require users to be at least 13 (or older where local law sets a higher age), and anyone under 18 needs a parent’s or guardian’s permission. My kids are seven and 10, so when they use it, it should be on my account with me doing the typing and them beside me. When they turn 13, ChatGPT’s parental controls (launched 29 September 2025) let you link a teen account to yours and set limits such as quiet hours and turning off voice, memory and image generation.

Supervision looks different at each age, in my view. At seven, I’d keep it to 10 to 15 minutes on one task, such as a story or a question about trains, with the adult typing and time left to check the answer together. At 10, the child can suggest the wording of the prompt and decide what to ask next, but personal details, health worries and anything that turns into a chat about feelings stay off the table and go to a parent instead.

What worries me more is the companion-style chatbots built to feel like a friend, rather than a general assistant writing a robot story. In an online safety advisory, Australia’s eSafety Commissioner warned about AI chatbots and companions: some children spend hours a day with them, and conversations can drift into sexual content and self-harm.

Most parents here have heard about the social media minimum age, in force since 10 December 2025, which stops under-16s having accounts on platforms like TikTok, Instagram and Snapchat. It doesn’t cover chatbots like ChatGPT. What does apply are eSafety’s new industry codes, which include rules for AI companion chatbots and require services to protect children from that kind of content from 9 March 2026, plus the provider’s own age terms.

Regulation helps, but it moves slower than kids do. Supervised, purposeful use is fine, even good. An always-on AI “friend” for a child is something I’d keep out of the house.

Judgement, not gatekeeping

My kids will grow up treating chatbots as normal, as ubiquitous as smartphones are now. I’m still deciding how I feel about that.

Part of me wants to protect them from becoming dependent on these tools. The other part knows that’s futile. My job isn’t to gatekeep the technology. It’s to help them develop the judgement to use it well: understand it enough to get value from it, and stay sceptical enough not to trust it blindly.

That’s harder than any technical problem I deal with at work.

If you’re explaining AI to your own kids

• Start with their analogy, not yours. Archael’s “autocomplete, but really good” worked better than anything I prepared.
• Test it on something they already know. Ask the chatbot a question your child knows the answer to, such as their favourite Pokémon’s type or how many carriages a local train has, then check its answer together. Catching a wrong answer teaches “it guesses” better than any explanation.
• Take their questions seriously. Andriel’s questions about understanding and bias are ones experts still debate. Say so.
• Make it hands-on, and sit with them. Check the age requirements of whatever tool you use. Until your child meets them, you do the typing and they watch and suggest.
• Have an answer ready for “Is it alive?” It will come up. Mine is: no, it doesn’t feel anything or want anything, even when it writes as if it does. It’s a program that’s very good at sounding like a person.

For now, Archael wants to use it to help write his Pokémon fanfiction. Andriel wants to know if AI can help him identify different types of trains.

I’ll take that as a win.