Teaching "Learning from Examples" to Builder mode (ages 8–10)

Hook & Warm-Up

Open with the app's own line, read with a bit of drama, as if you're letting the class in on a secret:

"How did you learn what a cat looks like? Your parents showed you cats and said 'cat!'
AI learns the exact same way — from thousands of labeled examples!"

Follow up: "Except AI doesn't just need one cat, or ten cats — it usually needs thousands. Why do you think it needs so many, when you figured out what a cat was after seeing just a handful?" Take a few guesses; there's no wrong answer here, but steer toward: computers don't have the same instincts and life experience a human brain brings to the table, so they need volume to make up for it. Tell the class that today's word is supervised learning — "supervised" because a human supervised the process by labeling every example first. It's the same idea as a teacher grading flashcards before handing them back — someone has to know and mark the right answer before anyone, human or AI, can practice with it.

Quick warm-up before the app: ask, "How many labeled photos do you think it would take to teach a computer what a cat looks like — 10? 100? A million?" Write a few guesses on the board without correcting them yet. Tell the class you'll come back to this number after the lesson's Real Examples section, where they'll see how big these training sets actually get in practice.

Main Activity

Open the lesson and walk through its three scenes, pausing to build the vocabulary as you go — this age group can hold onto real terms if you anchor each one to something concrete.

🏷️ Supervised Learning. The first two cards are Labeled Data and Training Set. Introduce the term this way: "A label is the correct answer attached to a piece of data — 'this photo = cat.' A training set is the whole pile of labeled examples the AI studies. Think of it exactly like a stack of flashcards where every card already has the answer written on the back." Ask the class who actually has to do the labeling in real life — the answer is real people, often thousands of them, going through huge piles of photos or emails and tagging each one by hand before an AI system ever sees them. That's genuine, sometimes tedious human work behind every training set, worth naming so students don't picture the labels appearing by themselves. The next two cards, Make Predictions and Accuracy, show the payoff: "After all that studying, you test it — show it a brand-new photo it's never seen, with the label hidden, and see if it guesses right. If it gets 95 out of 100 right, we say its accuracy is 95%." Ask: "Is 95% good?" and let them debate — it's a good warm-up for the discussion questions later. Push a little further: "95 out of 100 right also means 5 out of 100 wrong. Does that change how good 95% sounds?" This is worth sitting with for a moment rather than rushing past — the app's own next scene builds directly on this same number.

👨‍🏫 Teaching AI. Walk through the four cards as one continuous process: "When the AI's guess matches the label, it keeps doing whatever it was doing — it worked. When the guess doesn't match, it doesn't get angry or delete the example — it adjusts its internal numbers by a small amount and tries a slightly different approach next time." Then introduce the vocabulary word the quiz will ask about directly: "Going through the entire training set once — every single flashcard, start to finish — is called one epoch. If you kept practicing your multiplication flashcards over and over, each full trip through the whole stack would be one epoch." Close with the "Training Time" card: real AI models can take days or weeks to train, running on thousands of computers at once — nothing like one epoch of your times tables. Ask the class: "If one epoch is one full trip through the flashcards, and a real AI model runs through many epochs, why do you think it needs computers instead of a person doing this by hand?" Land on: the sheer number of repetitions (millions of examples, times many epochs, times constant small adjustments) is far beyond what a human could do by hand in a lifetime, but it's mechanically simple enough that a computer can do it very fast, over and over, without getting tired or bored.

🌍 Real Examples. Go through all four and ask the class, for each one, "what do you think the labeled training data was?" before revealing the app's answer:

• 📧 Spam Filter — trained on millions of emails already marked "spam" or "not spam" by real people over years of use.
• 🏥 Medical AI — trained on thousands of X-rays that a real doctor had already diagnosed, so the AI learns from human medical expertise, not from scratch.
• 🎵 Shazam — trained on millions of songs already tagged with the correct title and artist, so it can match a short clip back to the right song.
• 🌤️ Weather Prediction — trained on decades of past weather data that's already known and recorded, connecting today's conditions to what actually happened the next day.

This is the moment to circle back to the warm-up guess: "Remember your guess about how many photos it would take to learn 'cat'? Real systems like these often train on millions of examples, not hundreds. That's the scale we're talking about." Connect the abstract idea back to something tangible: every one of these needed a huge pile of already-labeled real-world data before it could learn anything at all.

Finish with the app's short matching round, which pairs an AI system with a simple non-AI object that does a similar-sounding job without any learning involved: labeled training data next to a paper clip, a trained spam filter next to a plain mailbox, song recognition next to a pair of drumsticks, weather-prediction AI next to a glass thermometer. Ask the class to explain out loud, for each pair, exactly what the AI side had to learn that the non-AI side never needed to — a glass thermometer, for instance, always works the same simple way no matter how many times you use it, while a weather-prediction system's guesses actually improve as more weather data accumulates.

Discussion

• Is 95% accuracy good? Would your answer change if the AI were sorting spam email versus reading X-rays? Why?
• If you were building an AI to sort your class's cafeteria snacks into "the class usually likes this" and "the class usually doesn't," what would your labeled training data need to look like?
• Why do you think AI needs thousands of examples when you figured out what a cat looks like from only a handful?
• What's the difference between an AI "getting it wrong" and an AI "adjusting its internal numbers"? Does calling it the second way change how you think about mistakes?
• Can you think of something at home or school that could be taught this way, if someone collected enough labeled examples?
• Who do you think actually does the work of labeling millions of emails as "spam" or "not spam," or millions of songs with their correct titles? Is that free, or does it take real time and money?
• If a spam filter is trained mostly on English-language emails, what do you predict would happen if it were given a mix of Tagalog and English spam? Why might that matter for an app built for Filipino users?
• The lesson says training a big AI model can take days or weeks using thousands of computers. If you were in charge of a company building one of these, what would you have to plan for besides just writing the computer program?

Quiz Walkthrough

All four questions test the same loop this lesson covered — label, train, predict, adjust, repeat — just from four different angles. If a student misses one, the fastest fix is usually to ask them to retell that loop in their own words before moving on, rather than just stating the right answer.

Wrap-Up & Extension

Close with: "Every AI you've heard about that can sort, recognize, or predict something started the same way today's lesson did — with a giant pile of examples someone had to label first, and a lot of guessing, adjusting, and trying again. That's supervised learning, and now you know its name."

Extension activity — Build a tiny training set: In pairs, have students pick a simple classroom category — for example, sorting canteen snacks into "masarap" (tasty, by class vote) and "hindi gaanong masarap" (not as tasty). Give each pair 8–10 index cards or sticky notes to "label" with a snack name and a vote. Once their training set is built, have them write down (not run — this is a paper exercise) a brand-new snack that wasn't in their pile and predict, based on their labeled examples, which category it would probably land in. Ask pairs to swap training sets and try each other's prediction — do they agree with the original pair's guess? This turns the abstract "label → train → predict" loop into something they've built with their own hands, and stretches the core 10–15 minute lesson into a full 30–40 minute period.

If time allows, add a twist round: have each pair deliberately build a tiny, biased training set (say, only snacks that cost less than ₱20) and predict a snack outside that range. Ask them to notice out loud why their prediction feels shakier this time — this previews, in concrete terms, why real AI systems need large and varied training sets rather than small or narrow ones.