Book 2 · Chapter 2 · Mission
Fool Your Parent
Build an AI classifier, then hand it to your parent and let them try to break it
⏱ 30–40 minutes
Build an AI classifier, then hand it to your parent and let them try to break it
⏱ 30–40 minutes
Ready to jump in? Tool: teachablemachine.withgoogle.com. Free, no account, runs in your browser, data stays on your computer. You already used this in Book 1 Ch 2 — this mission goes deeper.
Your kid is about to hand you an AI and challenge you to break it. This is how real AI safety researchers work — they build a model, then try to find the cases where it fails. Don’t go easy. The harder you try to fool it, the more your kid learns.
Train a classifier that matters to you, deploy it live with your webcam, then invite your parent (or older sibling) to try to break it on purpose. Every place they fool your AI is a place your training data came up short. That’s real AI testing work.
5 minutes
Teachable Machine tells two classes apart. Pick two things that feel specific and personal to you, so your AI feels yours and not like an example from a tutorial.
Pick one pair. Write the two class names on paper.
2 minutes
Go to teachablemachine.withgoogle.com. Click Get Started → Image Project → Standard. Rename the two class boxes to match your pair from Step 1.
15 minutes
For Class 1: click Webcam, take 25–30 photos from different angles. Move your hand between shots. Vary lighting a little.
For Class 2: same. 25–30 photos, different angles, some variety.
Click Train Model. Wait about thirty seconds.
5 minutes
Hold the camera up to something the AI has never seen. Watch the probability bars:
Get a feel for how your AI behaves before you hand it over.
10 minutes
This is the part nobody forgets.
Explain to your parent what the AI is trying to tell apart. Then challenge them: "Try to get it to pick the wrong class on purpose."
Watch what your parent does. Maybe they tilt the object. Maybe they use weird lighting. Maybe they bring in a third object the AI has never seen. Maybe they show a mix of both classes at once.
Every time they fool your AI, write down:
You’re now doing real adversarial AI testing.
“Real AI testers build a model, then try to break it on purpose. That’s the whole job. Every failure tells them what was missing. You just did the exact same thing a researcher at a tech company does every day.”
“And the conversation with your parent while they’re fooling your AI — that’s the part nobody warns you about. You’ll both be laughing. That’s the real skill transferring.”
“My AI is always right. My parent can’t fool it.”
Good problem, and also unlikely. If it seems unbeatable, try harder cases. Mix both classes in one shot. Introduce something the AI has never seen. Use bad lighting. Sooner or later you’ll find the edge. The edge is where the interesting learning lives.
“My AI is confused by everything.”
Your training data needs more variety. Add 10–15 more photos to each class — different lighting, different angles, different backgrounds. Retrain. Most classifier problems are really training-data problems.
“My parent is too good at fooling it.”
That’s a gift. Write down every way they fool your AI. You’re learning what an adversarial tester does, and you’re learning it from somebody who knows you and won’t go easy.