Learning from Examples

Learning Objectives

This is the first lesson of World 3, "How AI Learns," and it introduces the single most common way AI systems are actually trained: supervised learning. Everything students meet later in this world — clustering, reinforcement, neural networks — builds on the loop this lesson teaches first: label an example, show it to the AI, let it guess, tell it whether the guess was right, and repeat thousands of times. The lesson runs 5–15 minutes in-app across all four age modes, using the same three fact scenes (Supervised Learning, Teaching AI, Real Examples) with a different hook, summary, and quiz for each age band.

By the end of the lesson, a student should be able to:

• Explain, in their own words, that this kind of AI learns by being shown many labeled examples — like flashcards with the answer already written on the back — rather than by following one fixed rule a programmer wrote for every possible case.
• Describe what happens when the AI's guess doesn't match the label: it doesn't break, get angry, or give up — it adjusts its internal numbers by a small amount and tries again, across many repeated passes through the same examples.
• Name at least one real supervised-learning system from daily life — the lesson itself uses a spam filter, a medical X-ray screener, Shazam, and weather forecasting — and identify what the labeled training data for it would have been (millions of emails marked spam/not-spam, X-rays a doctor has already diagnosed, songs already tagged with title and artist).
• (Ages 11+) Use this stage's vocabulary precisely: training set, prediction, accuracy, and epoch — one complete pass through all of the training data.
• (Ages 15–18) Explain, at a level suitable for a technically minded audience, how a model's error is measured (a loss function) and reduced (gradient descent), and why performance is always checked on data the model was never trained on.

Teacher Background

The lesson's hook asks a genuinely good question: how did you learn what a cat looks like? Not by being handed a dictionary definition — by seeing lots of cats, usually with someone pointing and saying "cat." Supervised learning is AI doing something structurally similar. A person (or a team of people) takes a large pile of data — photos, emails, X-rays, songs — and attaches the correct answer to each one: "this photo is a cat," "this email is spam," "this X-ray shows pneumonia." That answer is called a label, and the whole labeled pile is the training set. The AI is shown the training set over and over, and each time it makes a guess, it is told the real answer.

Here is the mechanism in full, in the same order the app's three scenes present it: the AI looks at a labeled example and makes a guess; the guess is compared against the real label; if the guess matches, nothing dramatic happens — the AI's current settings worked, so it keeps them; if the guess doesn't match, the AI adjusts a large number of internal numeric settings by a tiny amount, in whatever direction seems to make future guesses on that example more accurate. Do this across every example in the training set once, and that's called an epoch. Real training runs the whole set through many epochs — sometimes thousands — which is why the app's "Real Examples" scene mentions training taking days or weeks on many computers running continuously.

"Accuracy" is simply the percentage of correct guesses — but the guesses that count are on examples the AI has never seen during training, not the ones it practiced on. This distinction matters more than it looks. An AI shown the same 1,000 cat photos over and over can eventually "solve" that specific set of 1,000 photos without ever learning what makes something a cat in general — it may have quietly memorized incidental details of those exact files instead. This is called overfitting, and it's the reason real training always holds back some labeled examples the AI never trains on, purely to test it honestly afterward. The app's own line that "more examples usually means better accuracy" is true and is exactly what younger students should take away; the nuance that a model can still fail by training on the same limited set for too long is the layer the Hacker and Architect scripts below add, because their quizzes test it directly.

The app's quiz gets more demanding as the age band goes up, and it introduces vocabulary that the three fact scenes above don't spell out on screen. Ages 11–14 are quizzed on overfitting, a loss function (the running score of how wrong the guesses are), the train/validation/test split (holding back data specifically to test honestly), and gradient descent (the actual procedure behind "adjusts its internal numbers" — nudging each number a small step in whichever direction reduces the loss). Ages 15–18 go further, into the bias-variance tradeoff, L1/L2 regularization, the Adam optimizer, and cross-validation. None of that is a separate topic — it's the same predict-compare-adjust- repeat loop, named with the precision a technical audience expects. The age-specific scripts linked below build that bridge explicitly before each quiz question, so you don't have to invent the explanation live in class.

Materials & Prep

No prep and nothing to print for the core lesson. Each student (or pair, if devices are shared) needs a phone, tablet, or computer with a browser and the AIQ app loaded — the hook, the three fact scenes, the practice matching round, and the quiz all run entirely on-device, with no login required. Read the Teacher Background above once before class, especially if you're teaching Hacker or Architect mode, since their quizzes assume vocabulary the in-app scenes don't spell out.

The extension activities in each age script add a little more: Explorer and Builder work best with a small stack of picture cards or two piles of real classroom objects (nothing that needs buying — drawings, magazine cutouts, or printed photos all work); Hacker and Architect extensions ask students to sketch a design on paper or, optionally, try a free browser-based tool. Nothing in this lesson or its extensions costs money, consistent with how AIQ itself is built and run.

Common Misconceptions

Pick your grade's script