Ask AI for Three Clues, Then Check If They Work

Most lessons about AI being unreliable are a grown-up saying so. This one hands your child a red pen and lets them prove it in twenty minutes.

How it goes

1. Pick a secret object in the room. A kettle, a shoe.
2. Ask for clues: Give me three clues that would let someone guess a kettle, without saying the word. Number them.
3. Audit each clue before anyone plays. Your child asks three questions of every one: is it true, does it only fit a kettle, and could a person actually use it?
4. Mark them in red. True but useless. False. Fits eleven other things. Most sets of three have one genuine dud.
5. Play the game with the surviving clues and see whether anyone guesses.
6. Write a replacement clue by hand for the one that failed, and compare it with the machine’s.

Who does what

AI produces clues, some of which are wrong or lazy, which is the raw material for the lesson.

Your child audits, marks in red, decides what survives, and writes the replacement.

You guess honestly, including guessing wrong when a clue is bad, so the flaw is visible rather than argued about.

Make it easier or harder

For a six-year-old: do the audit out loud together and let them rule on each clue with a thumbs up or down. The three-question framework is too abstract at this age, but the judging is not, and the judging is the habit.

For a twelve-year-old: ask for clues about something abstract rather than an object, such as jealousy or a rule. Abstract clues fail in more interesting ways, and auditing them requires deciding what the thing even is, which is a good argument to have at a kitchen table.

Take it off the screen

Keep a running tally on the fridge: clues generated, clues binned. Over a month it becomes a number your child quotes at people, which is a healthier relationship with these tools than any warning produces.

Play the same game with clues written by family members and audit those the same way. Humans write duds too, and noticing that keeps the scepticism fair rather than cynical.

When it goes sideways

Sometimes all three clues are fine. That is a real result and worth marking as one, otherwise the game teaches that machines are always wrong, which is its own kind of miscalibration. Write three ticks on the fridge tally and move on.

If your child starts rejecting clues that are actually good, push back. Ask them to say which of the three questions it fails and why. Scepticism that cannot explain itself is just contrarianism, and the difference is the whole point of the exercise.

One thing to watch

The aim is calibration, not distrust. A child who concludes AI is always wrong is as poorly served as one who thinks it is always right. Point out the clues that were good, and count those too. Teaching kids healthy scepticism covers keeping that balance.

Related: family trivia night applies the same audit to facts, and the family escape room is what to build once your child is good at spotting a broken clue.