// AI for children and teenagers
What a 12 year old can actually build with AI (and what we don't let them do)
What a 10 to 14 year old can build with AI under supervision, what stays out of reach at each age, and the lines we do not cross on accounts and personal data.
The title question is direct: a 12 year old can build an AI chatbot that remembers a conversation, answers questions from their own revision notes, and runs on a Linux server we host. They write the Python that connects to it. That is the Intermediate programme in one sentence.
The 'what we don't let them do' half matters as much. No child under 13 creates an account with any AI provider. No child at any age pastes personal details into a prompt. The AI runs through a service we operate, with a guardrail system prompt and moderation on every message. The advanced techniques you may have read about, such as training a large model from scratch or chaining several AI programs together, are not on the curriculum at any age.
This guide covers both halves: what a child actually produces at 10 to 12 and at 12 to 14, and where we draw the line.
What a 10 to 12 year old builds with AI
At Foundations level, a child builds two AI projects: a text classifier they train themselves, and a chatbot character with a personality they designed. Both run in the browser with no accounts in the child's name.
The classifier starts with collecting training examples. The child picks two labels (say, 'happy messages' and 'worried messages'), writes 40 to 60 varied phrases, and trains the model using a training tool we run. After training, they test it with new phrases and retrain where it fails. The model then becomes the brain of a browser game: type a message, the model classifies it, a character reacts. That is a real machine learning pipeline, at a pace a 10 year old can follow.
The chatbot is built in the club's chatbot builder, routed through the same AI proxy every child uses. The child writes a character card: a name, a subject area, a tone, and rules about what the character will never say. The result is a chatbot embedded on their club website. Your child trains a real AI model from examples and learns how ChatGPT actually works, at a level a 10 year old understands.
How ChatGPT actually works, at this age
ChatGPT does not look things up. It predicts the most likely next word, given everything it read during training. The 'guess the next word' game in session 5 makes this concrete: 'The cat sat on the ___.' Children fill it in immediately because they have seen that pattern before. A language model does the same, but has read billions of sentences.
Two things follow from this. First, the model can be confidently wrong (a hallucination), because the most likely next word is not always the true next word. Second, the model carries whatever biases existed in its training data, because there is no other source for its opinions. Children at Foundations level find at least one hallucination on purpose, then add a 'not sure' rule to their chatbot character card.
The Experience AI materials at experience-ai.org form the teaching spine for this session. A token visualiser on our platform shows children how a sentence splits into the fragments the model actually processes. Seeing 'ChatGPT' rendered as two tokens, not one word, is the kind of small concrete fact that lands with a 10 year old.
What a 12 to 14 year old builds
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 module at Intermediate level is the fourth of six. By the time a child reaches it, they can write Python functions and make HTTP requests from a script.
The module runs over six weeks (twice-weekly plan) or twelve weeks (once-weekly plan). A child starts by studying how language models work and probing the club proxy for hallucinations, documenting what they find. They then write Python that calls the AI through the club proxy and reads the response. Memory comes next: the messages list that every call sends in full, accumulating the conversation history. A system prompt gives the chatbot a name, a job, a personality, and rules about what it will never do.
In the ninth session, the child loads their own revision notes into the context and asks the bot questions about them. The final deliverable is a personal tutor bot: a system prompt they wrote, notes they loaded, a conversation history that grows with every turn, and code they reviewed line by line. The code lives in their GitHub repository.
The account model, and why it matters
No child under 13 creates an account with any AI provider. Anthropic's own terms set 18 as the minimum age for direct access to claude.ai. OpenAI's documentation advises parental consent for under-18 users. We do not work around these limits.
At Foundations, the browser tools require no account at all. At Intermediate, the club AI proxy at ai.buildersclub.org.uk handles every vendor API call. Children use a class token the tutor issues at the start of the module. They never see a vendor key, never sign up for anything, and the proxy adds a guardrail system prompt and moderation to every message. No child data is retained by the vendor after the session ends.
The one account a child does hold is GitHub, created at 13 or older, in session with a parent present. That is stated in the GitHub terms of service and it is what we follow.
What we don't let them do
Several things are off the curriculum at every age. These are design decisions, not gaps.
Children never paste personal details into a prompt. We teach this explicitly in the hallucinations session and the AI safety rules session; it is not buried in terms and conditions. The three rules children memorise are: check important answers against a real source, the model can be biased, and never paste personal information.
Fine-tuning a large language model is never taught. What happens at Foundations level, training a classifier from examples in a browser tool we run, produces a lightweight text classifier. It trains on examples the child collected. That is classifier training, not fine-tuning a language model, and the distinction matters: the child is working with a simple model they can fully understand, not modifying a system with billions of parameters.
Chains of AI programs talking to each other, and pipelines that run without a person watching, are not on the curriculum at any age, and nor are the specialist search databases used in large AI products. The most advanced the programme goes is a single agent loop with one tool call and an iteration cap, taught to 14 to 16 year olds in one session of the Advanced module. AI-generated code is reviewed, not submitted unchecked: the Intermediate module teaches children to read generated functions line by line, find edge cases the code misses, and document every change they make.
What Builders Club does about all of this
Live online sessions run by engineers who work at AWS. Your child builds a website, a game, an app and their own AI chatbot, all live on the internet, all under our supervision. Your child trains a real AI model from examples and learns how ChatGPT actually works, at a level a 10 year old understands. 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.
No accounts in your child's name under 13. Everything runs in the browser on platforms we manage. A project every six weeks (twice-weekly plan) or every twelve weeks (once-weekly plan), demoed to you at a Showcase. The first session is free, and you can cancel anytime.