- Workshop
Hour of AI
A single hour that goes from “can machines think?” to a model the group has trained themselves — and to using a chatbot as something better than an answer machine.
Overview
The hour is built so that the group trains something before anyone tells them what training is. It opens with the question Turing asked in 1950 and separates the science fiction from what is actually on the table. Then everyone trains an image classifier in the browser and breaks it on purpose — which is where the lesson about training data lands. From there the same idea scales up to language models, and the last third goes on prompting: the same question asked four ways, four different answers, and the point that the person asking is the one steering. It closes on the limits — invented facts, weak reasoning about physical space, and bias inherited from the data.
Who it is for
Written for 10 to 12 year olds and their parents, and works with a whole class as well as with a small group. No prior knowledge is needed and no programming is involved.
Learning outcomes
- Say what machine learning is: a model shaped by examples, not a set of rules someone wrote by hand.
- Train a working image classifier in the browser and explain why it fails on something it has never seen.
- Describe in their own words what a large language model does and where its answers come from.
- Write a prompt that sets a role and a situation, and notice how much the answer changes.
- Name concrete failure modes — invented facts, confident nonsense about physical space, bias carried in from the training data.
- Judge when an answer needs checking rather than accepting.
Lesson plan
| Stage | Minutes |
|---|---|
| Where the question comes from Turing's 1950 question and the imitation game. Separate two things that get confused: solving hard intellectual tasks is not the same as having a mind, and narrow AI is not general AI. A short comparison between what the films promised and what the room actually has. | 5 |
| Learning from data What training actually means — examples in, a model out, no rules written by hand. The Code.org AI activity fits well here and gives everyone something to do with their hands. | 10 |
| Train your own model Everyone trains an image classifier in Teachable Machine: a few classes, a few dozen camera shots each, and it works within minutes. Then break it deliberately — show it something outside the training set, or change the lighting — and let the group work out why it is confidently wrong. This is where the point about training data lands. | 15 |
| From pictures to words The same idea at a different scale: instead of photographs of cats, nearly every sentence people have written. The model learns which words tend to follow which. What that gets you — answers, explanations, translation, drafts — and what it does not: no memory between chats, no knowledge of yesterday, arithmetic it gets wrong, facts it invents. | 10 |
| Prompting A prompt is an instruction, and the same question asked four ways gives four different answers: ask as a grandmother, as a mother, as an eleven-year-old friend, as a teacher. Then let them try something real — planning a birthday for friends who want different things, asking an astrophysicist to explain black holes and then saying which part was unclear, or describing a game and reading the code that comes back. The contrast worth drawing: asking for an answer to copy teaches nothing, asking the next question is the whole point. | 12 |
| Where it breaks, and who decides Bias learned from the data, generated images and voices used as fakes, how much gets collected about people, and models trained on work somebody else made. Show a case where the model reasons confidently and wrongly about physical space. Close on the position worth holding: a powerful tool that you steer with your questions, whose answers you check, and whose decisions are still yours. | 8 |
| Total | 60 |
Equipment
- A laptop or tablet per participant or pair, each with a working camera — Teachable Machine needs one.
- Internet access, with teachablemachine.withgoogle.com, code.org and experiments.withgoogle.com all reachable from the school network.
- A projector for the demonstrations and for the prompting stage.
- One chatbot account for the prompting stage. Driving it from the projector works well and is the simplest answer to the age limits on those accounts.
Before the lesson
- Run the Teachable Machine flow yourself first. A camera blocked by browser permissions and a school network that filters the site are the two things that stop this lesson.
- Have three or four prompts of your own ready. The prompting stage stalls badly if the room has to invent one from nothing.
- Decide in advance how the chatbot gets used — driven by you from the projector, or on participants' own accounts — and check your school's position on it.
- Emoji Scavenger Hunt is a good five minutes if you finish early; it makes the point about image recognition a second time.
Bonus task
Compare two or three models on the same question and talk about why the answers differ — and about what, beyond capability, should decide which one you use: what it costs, how open it is, and where it comes from.