// AI for children and teenagers
ChatGPT is not a coding tutor: how we teach AI tools to teenagers
Why a chatbot alone does not teach a teenager to code, where its confident answers go wrong, and how our sessions use an AI assistant the child checks.
The pitch is easy to understand. A child is stuck, types the error into ChatGPT, gets a confident explanation and working code back in thirty seconds. It looks like tutoring. But what happened is: the chatbot answered, and the child watched. No debugging took place, no mental model formed. The child is not better at reading error messages than they were five minutes ago.
There are four specific ways this fails a learner, and they matter more than the general concern about over-reliance. A chatbot answers instead of teaches, it states wrong things with complete confidence, it has no memory of last week and cannot see what is on the screen right now, and a wrong explanation written in polished prose is harder to unlearn than a right one delivered badly. Each of those deserves a straight look.
We do use AI tools in sessions, from age twelve. The difference is in the role the child plays: reading the AI's output line by line, deciding what to keep, running the code, and checking what actually happens. The chatbot is a pair programmer whose output the child evaluates, not an oracle whose answers the child accepts.
It answers instead of teaches
Teaching means finding the gap in someone's understanding and building a bridge to it. Answering means supplying the correct output. Those two things feel similar when everything goes right, but they produce different learners. A child who gets the answer from a chatbot has not debugged anything. A child who gets a nudge from a tutor ('what does that error message say on the second line?') has practised the skill that will help them next time.
This matters especially at the start, when the whole job is to build habits: reading error messages, forming a guess, trying something, and being wrong before it works. If the habit a child builds instead is 'type it in and accept the answer', that habit survives long after the easy problems are gone. It is roughly equivalent to giving a child a calculator before they understand why multiplication works. The output is correct; the understanding is missing.
A tutor can withhold the answer deliberately. They can sit with the child's confusion for a minute, ask one more question, and let the child find the thing themselves. A chatbot does not have that option. Answering is the only move it has.
It states wrong things with complete confidence
Large language models generate text by predicting which word (more precisely, which token) is most likely to come next, given everything that came before. They do not retrieve facts from a database and they do not verify what they produce. OpenAI's GPT-4 technical report describes hallucinations as an expected behaviour of this architecture: the model sometimes produces statements that are plausible-sounding but factually wrong, with no reliable signal to the reader that this has happened. Anthropic makes the same observation about Claude. The model is not lying; it has no mechanism for distinguishing 'I know this' from 'this is what a confident explanation on this topic would sound like'.
For an experienced developer, a wrong suggestion from a coding assistant is usually detectable. They have enough context to notice that the function the chatbot invented does not exist, or that the logic is subtly off. A ten or twelve year old building their first mental model of how Python works has none of that. They have no independent source to check the answer against, so the confident prose goes straight in.
Stephen Wolfram's 2023 explainer on how ChatGPT works describes the process plainly: the model is 'just adding one word at a time, always asking what word should come next given what's come before'. Plausible continuations of text and true statements about the world overlap a lot of the time, but not always. A beginner cannot tell the difference.
It cannot see the screen or remember last week
A tutor in a live session watches what the child does. They can see that the indentation is two spaces instead of four, or that the child is stuck and confused but not saying so. They remember that this child struggled with the same concept last week and can go back to it without being asked. A chatbot has none of that. Each conversation starts from zero, and the only input it has is the text you type.
When a child is stuck, the most useful information is often not in what they say but in what they do. The pause before a question and the specific mistake they make repeatedly: all of that is invisible to a chatbot. A tutor working with the same small group every week builds a picture of each child's particular gaps that no single conversation can replicate.
This also means the chatbot cannot adjust. A good tutor notices when a child has grasped something and moves on, or when they have not and slows down. A chatbot answers the question as asked, every time, in roughly the same register, whether the child is ten sessions in or one.
A wrong answer in fluent prose is harder to unlearn
There is a specific problem with the chatbot's output style that makes it worse than a vague or partial answer for a beginner. The language is polished and authoritative. It reads like a textbook. When a child reads something that sounds like it was written by someone who knows what they are talking about, they update their model of how things work accordingly. A wrong answer in that register leaves a stronger impression than a hesitant one.
Wrong mental models in computing are particularly persistent. If a child builds an incorrect picture of what a variable is, or what happens when a function is called, that picture will produce bugs they cannot explain for months. A tutor who can catch the misunderstanding early and correct it clearly, in the moment when the child is confused about it, does a different job from an algorithm that produces fluent text.
This is also why the chatbot's confident tone is part of the problem, not incidental to it. A tutor who is uncertain says so. They might say 'I think it works like this, but let us check'. A language model does not have an uncertain register. It produces the same confident prose whether it is right or wrong, and a child reading it has no way to tell which.
How we use AI tools in sessions
From age twelve, in the Intermediate Build with AI module, children learn to call an AI assistant from Python code and build a chatbot that remembers the conversation. The sessions trace how a language model actually works, from training data to next-token prediction, so that what the model does and what it fails to do are not mysterious. Children run prompts designed to produce hallucinations and record what the model did and why.
The specific skill the module teaches is: use the AI as a pair programmer, read its output line by line, run the code yourself, and check what it produces before accepting it. That is the phrase in the skills checklist: 'using an AI assistant as a pair programmer and checking its work'. The child runs the code and decides whether the AI's suggestion was right. The AI is a faster way to get a first draft, not a way to skip the thinking.
At the Advanced level, in the AI Engineering module, children build a 20-question evaluation set and run it against their own AI system, with the answers scored by an AI judge that measures whether they are correct. They see the before and after improvement. Evaluating AI output is taught as a skill alongside producing it, because the two belong together. A child who has built and scored their own AI system has a much more grounded view of what these tools can and cannot do than one who has only used them.
What we do about it
Live online sessions run by engineers who work at AWS. The people teaching these sessions use AI coding tools in their own work and can give a child a straight answer about when to trust the output and when to check it. That context does not come from a chatbot.
From 12, your child runs their own Linux server that we host, writes Python that does real jobs, and builds an AI assistant that remembers the conversation. The AI is in the room for those sessions as a tool, not as the teacher.
The sessions are small groups of six to eight children, ninety minutes, live. A tutor can see every screen and hear every question. They notice when someone has the answer wrong before that child has finished typing it. That is what a chatbot cannot do.