🧠 How AI Actually Works
Go under the hood — understand what is really happening.
🔥 Warm Up — 3 minutes
"You have used AI for 24 sessions now. But do you actually know what is happening when you type a question and get an answer? Today we find out."
"Not using AI to get answers today — using AI to understand itself. That is a different kind of session."
🤖 The Activity — 20 minutes
Today is conceptual. We go deeper than any previous session.
Read carefully. Ask: Does this make sense? What is still confusing? Keep asking follow-up questions until the mechanism is clear.
These are the three building blocks of understanding AI. Your child should be able to explain them after this session.
This thought experiment makes training data concrete and intuitive. The answer reveals why AI has biases, gaps, and unexpected strengths.
🧩 The Twist — 5 minutes
AI's answer to this is genuinely interesting. Then ask your child: Is AI being honest here, or is it just telling us what it thinks we want to hear about its limitations? There is no clean answer — but the question is worth sitting with.
Understanding how AI works makes this rule richer. AI does not know it is wrong — it generates the most statistically likely continuation of text. That is why it sounds confident even when it is wrong. The mechanism explains the mistake pattern.
Under the Hood Badge 🧠
Understood how AI actually works — not just what it does
Parent Notes — tap to expand
▼Mechanistic AI literacy — moving from user to understander. The training data thought experiment is particularly powerful for building intuition about AI biases and gaps.
- Can you explain what a token is in your own words?
- Why does AI sound confident even when it is wrong?
- If you were going to build an AI — what would you train it on, and why?
Some children find this session abstract. Keep it concrete with the thought experiments. The cookbook analogy is particularly accessible.
Understanding the mechanism helps children maintain appropriate scepticism. When they know AI is doing sophisticated pattern matching rather than thinking, they approach its outputs more carefully.