Teaching "Good Data vs Bad Data" to Explorer mode (ages 5–7)

Hook & Warm-Up

Gather students close — on the rug, or with devices face-down for a moment so all eyes are on you. This lesson's mascot is Pixel, so read the opening line with warmth and a little wonder, as if Pixel is telling the class a secret:

"What if you only showed A.I. pictures of BIG dogs? 🐕 It would think ALL dogs are big! And it would be WRONG about small dogs! That's why GOOD data matters!"

Pause after "wrong about small dogs" and ask, with genuine curiosity in your voice: "Has anyone here ever seen a REALLY tiny dog? Like, tiny enough to fit in a teacup?" Let a few students describe small dogs they know (chihuahuas, teacup poodles, a neighbor's puppy). Then say: "So if a smart helper only ever saw pictures of big dogs like Great Danes, and never saw a tiny dog like that... what do you think it would say if it saw one?" Take guesses — most students will land on "it wouldn't know it's a dog!" without much prompting. That's the whole lesson in one sentence: a smart helper only knows what it's been shown.

Tell the class: "Today we're going to be data detectives. We're going to figure out what makes examples GOOD for teaching a smart helper, and what makes examples BAD — and why it matters so, SO much."

Before opening the app, do a thirty-second warm-up with an object already in the room. Hold up a shoe and say: "Imagine I only ever taught you the word 'shoe' by showing you MY shoe, this exact one, a hundred times. Then I showed you your friend's sneaker. Would you know that's a shoe too?" Most students will say yes, because people generalize easily — but push a little further: "What if I only showed you high heels, never sneakers or sandals or boots — would 'shoe' in your head only mean high heels?" This gets at the same idea as the big-dog hook from a second angle before the vocabulary gets any harder, so slower processors in the room have two chances to grab the idea rather than one.

Main Activity

Open the lesson together. It has three parts — walk through each one like a game, always asking the class to guess before you reveal the answer.

Part 1 — Data Quality ✨. Show or describe each idea, one at a time, and connect it back to the big-dog example from the hook:

• ✨ Clean Data — "Good examples are neat and correct — like a clear, correct picture of a dog labeled 'dog.' Nothing mixed up." Ask: "What if a picture of a CAT got labeled 'dog' by accident? Would that confuse a smart helper?" (Yes!)
• 🗑️ Dirty Data — "Messy examples — wrong labels, blurry pictures, mistakes — confuse a smart helper, just like it would confuse YOU if your teacher mixed up all your worksheets." Act this out: hold up an imaginary "dog" picture and say "cat" out loud, then look confused.
• 📏 Enough Data — "If you only studied 3 problems before a big math test, would you be ready?" (No!) "A smart helper is the same — it needs to see MANY, many examples, not just a few, before it really learns."
• 🌈 Diverse Data — Return to the hook: "If ALL the dog pictures show big dogs, in sunny weather, standing still... the smart helper only really knows THAT one kind of dog-picture. Different sizes, different weather, different poses — that's what 'diverse' means here: lots of DIFFERENT kinds of examples."

Part 2 — Bias in Data ⚖️. Introduce the word "bias" gently: "Bias is a big word for a simple idea: it means the examples weren't fair to everyone." Walk through the four items, translating each into something concrete:

• ⚖️ What is Bias? — "If a smart helper only ever saw big dogs, it's 'biased' toward big dogs — it's not being fair to small dogs, because it never got to learn about them."
• 👤 Face Bias — "Some smart helpers that try to recognize faces were only shown LOTS of pictures of people with light skin, and only a few pictures of people with dark skin. So guess what happened? They got much better at recognizing light-skinned faces, and made more mistakes with dark-skinned faces. That's not fair — and it happened because of the examples, not because anyone wanted it to happen."
• 💼 Hiring Bias — "A company once built a smart helper to help pick out good job resumes. But it had only ever seen resumes from mostly men, from years and years before. So it started liking 'men' resumes better — even resumes that just mentioned things like a women's sports club got pushed down! The company found out and stopped using it."
• 🏦 Loan Bias — "Some smart helpers that decide who gets to borrow money learned from old records where certain neighborhoods were unfairly turned down a lot in the past. So the smart helper kept being unfair to those same neighborhoods, copying the old unfairness."

After these three examples, pause and say plainly: "Notice something — none of these smart helpers were being MEAN on purpose. They just repeated whatever was in the examples they were shown. That's why picking GOOD, FAIR examples matters so much." If a student asks "so is it the smart helper's fault?" — this is a good moment to say clearly: "No. It's not the smart helper's fault, and it's usually not any one person's fault either. It's about making sure the examples include EVERYONE before you ever start teaching it."

Part 3 — Fixing Data Problems 🔧. End on hope, not just problems — this part is about how people fix it:

• 🔍 Audit Data — "Check your examples! If you never showed any pictures of elderly people, or any tiny dogs, you'd notice — and you'd know what to add."
• ➕ Add Diversity — "Once you notice something's missing, go get more examples of it, until it's fair and balanced."
• 🧪 Test for Bias — "Try the smart helper out on lots of different kinds of people or things, separately, and see if it works just as well for everyone."
• 👥 Diverse Teams — "If everyone building the smart helper is very similar to each other, they might not notice what's missing — because it never occurred to them to look. Different kinds of people on the team notice different missing pieces."

Give one more concrete example to cement Part 3 before moving on: "Imagine a school is building a smart helper that recommends lunch menus. If the team building it only ever ate rice and adobo growing up, they might forget to include vegetarian options, or halal options, or someone with a peanut allergy. It's not that they don't care — it's that it didn't occur to them, because it wasn't part of THEIR examples. That's exactly why 'Diverse Teams' matters: more kinds of people notice more kinds of missing pieces."

Close the activity by reading the lesson's own summary line together: "A.I. needs fair and complete examples to learn right! If it only sees big dogs, it won't know small dogs. If it only sees boys, it might not work for girls. Good data = good A.I.!" Then let students do the in-app practice round, sorting more examples into fair vs. biased data.

Discussion

• "If you were teaching a smart helper what a 'fruit' looks like, and you only ever showed it apples, what would happen when it saw a banana?"
• "Is it the smart helper's fault if it's unfair, or is it about what it was shown?" (Guide toward: it's about what it was shown — the smart helper can't know what it never saw.)
• "Can you think of a time YOU only knew about one kind of something, until someone showed you more kinds?" (Any honest answer works — new foods, animals, games.)
• "If you were building a class photo album of 'what a dog looks like,' what different KINDS of dogs would you want to include so it's fair?"
• "Pixel wants to teach a smart helper what 'breakfast' looks like, but Pixel only eats pandesal every morning. What should Pixel do before teaching the smart helper?" (Guide toward: show it lots of different breakfasts — rice and egg, cereal, oatmeal, tocino — not just one.)

Quiz Walkthrough

Wrap-Up & Extension

Close with: "You're data detectives now! Every time you see a smart helper — a phone, a game, a video app — you can ask yourself: what did it learn from? Did it get to see EVERYONE and EVERYTHING fairly? That question makes you someone who thinks carefully about smart helpers, not just someone who uses them." Give the class a round of applause for finishing the lesson, and remind them that being a "data detective" is a habit they get to keep using forever, not just something for today's lesson.

Extension activity — "Fair Picture Book": Give small groups a stack of magazine cutouts, printed photos, or drawing paper and ask them to build a tiny "picture book" that teaches a pretend smart helper what a "good pet" looks like — but it has to be FAIR, meaning it needs different kinds of pets (big, small, fluffy, short-haired, common, unusual). Groups then swap books and try to spot anything "biased" or missing in another group's book (a book with only dogs, or only fluffy animals). This turns the abstract idea of representative data into something students physically build and can see is missing pieces — a satisfying way to stretch the 5–15 minute core lesson into a full 30-minute activity.

If time is short, a simpler five-minute version works too: hold up two hand-drawn "example books" you prepared ahead of time — one clearly unfair (ten pictures, all the same kind of animal) and one clearly fair (ten pictures, many different kinds) — and have the whole class vote, thumbs up or thumbs down, on which one would teach a smart helper better. Ask a volunteer to explain their vote out loud before revealing which one you intended as the "trick" example.