📄 Teaching a Computer to Read
Why reading a bill is much harder than it looks.
🔥 Warm Up — 2 minutes
Write the same word twice — once neatly, once in your worst handwriting. Ask: “Which one would a computer find harder?”
“Today we find out what a computer actually sees when it looks at writing.”
🤖 The Activity — 15 minutes
Write a pretend shop receipt on paper: shop name, three items, three prices, a total. Make one price deliberately smudged.
Now play the computer. Your child reads it out loud and, after every single item, says how sure they are out of ten.
The smudged one gets a low number. That number is the whole point. A real system does the same thing, and anything below its line goes to a human instead of being guessed at.
🧩 The Twist — 5 minutes
Hand them the receipt upside down. Then at arm’s length. Then folded so one item is hidden.
Ask what the computer should do about the hidden one. The right answer is not to guess — it is to say “I could not find it.” Saying nothing and saying a made-up number are very different.
Reader Badge
Learned that a good computer says how sure it is
Parent Notes — tap to expand
▼That document AI produces a confidence score alongside every value, and that the honest behaviour when confidence is low is to escalate rather than guess.
- What should the computer do if it is only 3 out of 10 sure?
- Is a wrong answer worse than no answer? When?
- How could you make the receipt easier for a computer to read?
The insight to listen for is your child distinguishing “I do not know” from “here is a guess”. That distinction is the entire safety argument for AI, in one sentence.
If you use a chat AI for this session, photograph the pretend receipt you made together. Never photograph a real bill, statement or card — real documents have real numbers on them and those do not belong in a chat window.