Teaching "When AI Isn't Fair" to Builder mode (ages 8–10)

Hook & Warm-Up

Open with the lesson's real hook line, read with some weight — this is meant to land as a surprise, not a throwaway fact:

"An AI hiring tool rejected women. A face scanner couldn't see dark skin. AI can be unfair — understanding WHY is the first step to fixing it."

Ask the class: "Before today, did any of you think a computer could be unfair? Most people assume computers are super neutral, like a calculator — 2+2 is always 4, no matter who's asking. Today we're going to find out that some AI is NOT like a calculator, and figure out why."

Draw out the difference a little more: "A calculator follows exact rules a person wrote down, step by step. A lot of AI doesn't work that way — instead of following rules a person wrote, it looks at thousands of past examples and figures out its OWN rules by finding patterns. That's powerful, but it also means it can pick up patterns nobody wanted it to learn, if the examples it saw weren't fair to begin with."

Run a quick warm-up demonstration. Say: "Imagine I want to build a robot that can tell you if a drawing is a 'good' drawing of a dog. I train it by showing it 100 drawings — but 90 of them are drawings of golden retrievers, and only 10 are other breeds." Ask: "If someone draws a really good picture of a chihuahua, what do you think the robot will say?" Let students guess (most will correctly guess it'll say it's "not a good dog drawing," or "not a dog at all"). Land the point: "The robot isn't being mean to chihuahuas on purpose. It just didn't get enough practice with them. That's the whole idea behind today's lesson, called AI bias."

Push the warm-up one step further before moving into the main activity: ask, "what if instead of dog breeds, the 90-to-10 split was about something that mattered a lot more — like which resumes get hired, or whose face a security camera can recognize?" Let that sit for a second. "That's not a hypothetical. Both of those things have actually happened, and we're about to look at exactly how."

Main Activity

Walk the class through the lesson's three sections on the app (or on a shared screen), pausing after each item to explain the real-world story behind it in Builder-level language — specific and factual, but not yet loaded with technical vocabulary like "algorithm" or "fairness metric."

Set expectations up front: "Some of these stories are about grown-up things, like getting a job or buying a house. You don't need to know exactly how a mortgage works to get the point — just notice that in every story, a computer program made a decision that affected real people's lives, and it got that decision wrong for a specific, findable reason. Our job today is to find that reason."

Section 1 — What Is Bias?

• Training Bias: Explain the resume example concretely: "A company built a tool to sort job resumes automatically. It learned from ten years of resumes the company had hired in the past — and because the company had mostly hired men, the tool started marking down resumes that even mentioned the word 'women's,' like 'women's chess club captain.' Nobody told it to do that. It just noticed that word wasn't common in the resumes that got hired before."
• Face Recognition: "A famous study tested face-recognition programs and found they were about 99% accurate for light-skinned men, but only about 65% accurate for dark-skinned women. Same idea as the chihuahua drawings — the programs just hadn't been shown enough examples of every kind of face."
• Dice Roll (not AI): Contrast it — a fair die gives every number an equal chance. That's the standard we want AI to meet, and mostly doesn't by default.
• Lending Bias: "Some AI systems that decide who gets approved for a home loan ended up saying 'no' more often to people in certain neighborhoods — even when those people could actually afford the loan. This has a name: digital redlining, after an old, very real practice of banks literally drawing red lines on maps around neighborhoods they wouldn't lend to."

Pause here and ask the class to spot the pattern across all four examples: "In every single one of these, did anyone sit down and decide 'let's be unfair to this group'?" The answer is no every time — that repetition is intentional, and it's the single most important thing to land in Section 1 before moving on.

Section 2 — Why It Happens

• Historical Data: AI trained on old records learns whatever patterns — fair or unfair — were already in those records.
• Label Bias: Someone has to label training examples ("this resume = hire," "this face = match"). If the people doing the labeling have their own blind spots, the AI absorbs them.
• Measuring Tape (not AI): A tape measure doesn't care who's holding it. That's the goal — AI decisions should hold up the same way, regardless of who's being measured.
• Proxy Variables: This is the one worth slowing down for. "A bank isn't allowed to use race to decide loans. But a ZIP code can carry almost the same information, because of how neighborhoods have historically been divided. So an AI using ZIP code can end up biased by race without ever 'seeing' race." Ask: "Can you think of another piece of information — besides race or gender directly — that might secretly tell a computer a lot about someone's background?" (School name, first name, and neighborhood are all good answers.)

If the class is engaged, try a quick "spot the proxy" round: name a Philippine example, like "the barangay someone lives in," and ask whether that could quietly carry information about income or background the same way a ZIP code does in the lending example. Students usually get this quickly once they see the ZIP code pattern once.

Section 3 — How to Fix It

Introduce the four fixes as a matched set: diverse training data, bias audits (testing separately for different groups before launch, not just one overall score), and diverse teams building the AI. If time allows, run a short sorting activity using index cards: write each of these eight examples on its own card — biased training data, fair dice, biased face recognition, measuring tape, automated bias audit, paper checklist, AI lending discrimination, inclusive design — and have small groups sort them into "This is AI" and "This is NOT AI." (Answer: training data, face recognition, bias audit, and lending discrimination are AI; the dice, tape, checklist, and inclusive design are not.) This mirrors the sorting game inside the app and reinforces that fairness tools themselves don't have to be AI — sometimes a paper checklist and a mixed team of people do more good than another algorithm.

After the sort, ask groups to notice something specific: every "not AI" card in the pile is something a regular person could pick up and use immediately — a die, a tape measure, a checklist, a mixed team. Point out: "None of our fixes require inventing a fancier AI. Sometimes the best tool for catching an AI's mistake isn't another piece of technology at all — it's a group of people with different eyes looking carefully at the same problem."

Discussion

• Why didn't the people who built the biased hiring tool notice the problem before it launched?
• If a company only has photos of certain kinds of faces to train with, whose responsibility is it to go find more varied photos — the company, the government, or someone else?
• Can you think of a "proxy" in real life — something that seems unrelated but actually reveals a lot about a person, like their address or their school?
• If a bias audit found that an AI was less accurate for one group of people, what should the company do — stop using the AI, fix the data and try again, or something else?
• Do you think a team made up of very similar people (same age, same background) would catch the same problems as a more varied team? Why or why not?
• If you found out an app you use every day (a game, a social app, a homework helper) had been trained mostly on data from one country or one age group, would that change how much you trusted its suggestions? Why?

Quiz Walkthrough

Wrap-Up & Extension

Close with: "AI isn't automatically fair just because it's a computer. It learns whatever it's shown, which means it's up to the people building it — and someday, that could be some of you — to make sure it's shown a fair picture of the whole world."

Extension activity — "Audit the Class Robot" (15–20 minutes): As a class, invent a simple fictional AI, like "SnackBot," which recommends a snack based on a written description of a student. Have students in pairs write five short fictional "training examples" (a one-line description plus the snack SnackBot recommended). Collect all the examples on the board and look for patterns — did certain kinds of descriptions always get the same recommendation? Then discuss what a real bias audit for SnackBot would check, and how the class could make its training examples more varied and fair. This turns the audit concept from something abstract into something the class builds and inspects themselves.

To wrap the activity, have the class vote on one concrete fix for SnackBot's training data — for example, "add five more examples from students we haven't heard from yet" — and actually collect those examples on the spot from students who haven't contributed one yet. This small step makes the "diverse data" fix something the class did, not just something they talked about, and it's a natural bridge into the next lesson's themes around designing systems responsibly from the start.