The part that matters needs no electricity
Almost everything worth teaching about AI is about prediction, pattern and confident error. All three can be taught with chalk, paper and thirty voices. The button-pressing is the easy half and can be picked up in an afternoon whenever a device appears.
So do not treat the paper version as the fallback. Build the lesson on paper and treat a working projector as a bonus. Reverse the default and you will never lose a lesson to a power cut again.
Activity one: the class becomes the model
No materials at all.
Write a sentence stem on the board. Every student writes one word to complete it. Silently, no discussion, everyone commits before anyone hears anyone else.
The old woman walked slowly into the ______
Tally by show of hands. Room, house, shop, kitchen, one "sea". You now have a distribution on the board. Tell them: the machine has the same kind of guess, built from billions of sentences instead of thirty.
Second stem: The capital of Japan is ______. Nearly unanimous.
Ask them why one was a scatter and one was not. From there, two conclusions arrive by themselves. Why the machine gives different answers to the same question if you ask twice. And why it is dependable about Tokyo and unreliable about the population of your district.
Twenty minutes, no equipment, and they now hold the actual mechanism rather than a picture of a robot.
Activity two: where its habits come from
Give one student a short story to read, then have it retold down a line of five. What survives is what each person found worth remembering.
Then the teaching question: if nearly every story ever written about doctors used "he", what word will a machine guess after "the doctor said that"? Let them answer. Then ask what that means for a machine that learned from everything people have written.
You have taught bias in training data in ten minutes, with no equipment, and they will not forget it because they were the training data.
Activity three: the confident wrong answer
This needs a phone with data once, not a phone in class.
Whenever you have connection — at home, early morning, at an internet cafe — ask an AI five questions from your syllabus. Write down its answers word for word, and write down the truth. Choose questions where it slips: local history, a district detail, a multi-step calculation, a plant that grows only in your region.
In class, read the five answers aloud as though they were authoritative. Students vote true or false on each. Then reveal.
The lesson is not that it got two wrong. The lesson is that the two wrong answers sounded exactly like the three right ones. That is the thing you want them to carry, and it lands harder on paper than on a screen, because nothing distracts from the words.
Practical notes from teachers who work this way
- A printed transcript is a real teaching resource. One conversation, printed once, photocopied, teaches as well as a live demonstration and never fails halfway through.
- Batch the online work. Do all your generating in one session when power and data are available. Bring paper. Never build a lesson whose spine depends on connectivity arriving at 10am.
- One phone at the front beats forty half-charged ones. If devices are scarce, a demonstration you control is better than a scramble, and it keeps the class in one conversation.
- Write the answers down before class. Live demonstrations fail, models change their answers, and a lesson built on "watch what it says" has no plan B. A lesson built on "here is what it said, and here is what was true" always works.
The honest worry
There is a real risk that this becomes another line of inequality — schools with connectivity teaching it hands-on, schools without teaching it in the abstract.
But look at what actually transfers. That it predicts. That it is confident when it is wrong. That it learned from us and carries what we wrote. A student who holds those three things and has never touched a keyboard will use the tool better, the first day they get one, than a student who has typed into it for a year without understanding any of it.
Before you move on