Week 11 — Bias: When the Data Is Lopsided¶

At a glance


Date	Friday, October 30, 2026
Unit	2 — AI Unmasked
Big idea	Bias can appear when training data is missing important kinds of examples.
Goal	Students deliberately build a lopsided classifier, discover its failure, then repair it with better data.
Student accounts needed	None. Teachable Machine runs without login.
Prep time	5 minutes, plus the same upload fallback as Week 10

Why this lesson looks the way it does¶

Bias is easiest to moralize and hardest to teach well. Discovery is better. If students first create a lopsided dataset with their own hands, the failure is not abstract and not accusatory. It is visible: the model performs well on the examples it saw and poorly on examples it did not see.

Keep the categories non-personal. Do not use faces, gender, race, or bodies as classroom classifier categories. You can teach the technical mechanism of bias with pencils, hands, paper shapes, lighting, backgrounds, and angles. Once students understand the mechanism, the broader social implications are easier to discuss honestly and calmly.

Your professional framing matters: bias is not only about bad intentions. In real systems, it often appears because data was convenient, incomplete, old, cheap, or collected for a different purpose.

CSTA 2026 standards¶

Concept area	Connection
Algorithms & Design	Students intentionally design an unfair training set and then redesign it.
Programming	Students observe behavior changes caused by data changes rather than code changes.
Data & Analysis	Students compare lopsided and more representative datasets.
Systems & Security	Students consider the conditions under which a system is safe or unsafe to use.
Computing & Society	Students reason about fairness, representation, and consequences without fear-based framing.

"I can…" objectives¶

By the end of today, students can say:

I can explain bias as a pattern of unfair or unreliable errors.
I can show how missing examples can make a model fail.
I can improve a biased classifier by adding more representative data.
I can ask who or what might be left out of a dataset.

Teacher prep — the 5-minute version¶

Open https://teachablemachine.withgoogle.com.
Write on the board:

Last Friday: We trained a model.
Today: We train it unfairly on purpose.
Under that, write:

Bias question: What examples are missing?
Choose a safe lopsided setup to model if needed:
Train pencil only on yellow pencils, then test a mechanical pencil.
Train hand only from one person's hand and one lighting angle, then test another person's hand or a shadowed hand.
Train paper shape only on white paper, then test colored paper.

Minute-by-minute run sheet¶

0:00–0:04 — Anchor¶

"Last Friday you trained a model and saw that it learned from your examples. Seven days later, here is the important follow-up: what if the examples are incomplete? Today we are going to make that problem happen on purpose, safely, so we can understand it."

Point back two weeks:

"Our unit sentence is still: examples → patterns → predictions. Bias often starts at the examples step."

0:04–0:12 — Mini-lesson: bias without blame¶

Say:

"Bias does not always mean someone tried to be unfair. In AI, bias can mean the system makes errors in a pattern because the data was lopsided. Maybe the examples came from one kind of camera, one kind of room, one kind of person, one season, or one language."

Use a simple demonstration:

"If I train 'pencil' using only yellow wooden pencils on this desk, and then it fails on a blue mechanical pencil, the model is not being stubborn. It learned the world I showed it."

Add the key sentence:

"A model can be accurate on average and still fail badly for the examples that were missing."

0:12–0:36 — Paired work: build, break, repair¶

Pairs use Teachable Machine again. The task has three phases.

Phase 1 — Build it lopsided (6 minutes)

Students train two classes, but Class A must be narrow. Examples:

Pencil class: only yellow pencils on the same desk
Book class: only one book cover from one angle
Hand class: only one person's hand, same pose, same lighting
Shape class: only white triangles, same size

Phase 2 — Break it fairly (8 minutes)

They test with examples that should count but were missing:

different pencil type
different background
different angle
different person holding the object
different color or size

Phase 3 — Repair it (10 minutes)

They add the missing examples and retrain.

Say:

"Do not trick your model with nonsense. Give it fair examples it should handle. Bias shows up when a reasonable example fails because the training data was too narrow."

Circulate with:

"What kind of example did you leave out?"
"Did the mistake affect both classes equally?"
"What did you add to repair it?"
"Would you trust this model outside this room? Why or why not?"

Use a three-question pattern.

"What did you leave out?"

"How did the model fail?"

"What data helped repair it?"

Then widen carefully:

"Now imagine the same issue in a bigger system: speech recognition trained mostly on one accent, medical images mostly from one kind of machine, or a recommendation system trained mostly on one group of users. The technical question is the same: who or what is missing from the data?"

0:42–0:45 — Teaser + prompt¶

"Next Friday we move from pictures to words. We will build a sentence one word at a time and see why language models are prediction machines."

Tiered challenges¶

🟢 Starter (grades 5–6): Name one kind of example your model did not see and what happened when you tested it.
🟡 Builder (grades 7–8): Explain the difference between a random mistake and a biased pattern of mistakes.
🔴 Extender (grades 9–10): Design a fairness test: what subgroups of examples should this classifier handle before you would call it reliable?

"What was missing from your training data, and how did that change the model's behavior?"

🤔 Think About This¶

Notice a system that works better for some people, places, or situations than others.

It could be voice captions, autocorrect, a game controller, a search result, or a recommendation. Ask: what examples might be missing?

🔌 Unplugged Plan B — if the internet is down¶

Use paper cards to make bias visible.

Time	Activity
0:00–0:04	Same anchor: last week we trained; today we train unfairly.
0:04–0:12	Mini-lesson with the pencil example and the sentence: "It learned the world I showed it."
0:12–0:24	Groups create a lopsided training set for a category, such as "school supplies": only rectangular supplies, only one color, or only one size.
0:24–0:34	Another group tests with fair-but-missing examples and records which ones fail.
0:34–0:40	Groups add 4 better examples and retest.
0:40–0:45	Share: "What was missing?" Same teaser and prompt.

Common problems & fixes¶

Problem	Fix
Students try to trick the model with absurd examples	Restate: "Fair but missing. The example should belong, but the model has not seen that kind."
Discussion turns personal too fast	Bring it back to neutral objects first. Mechanism before social examples.
Students say bias means "wrong on purpose"	Use the phrase: "pattern of errors from lopsided data."
Model does not fail	Make the lopsided data narrower: same background, same angle, same object, then test a real variation.
Cameras are blocked	Use upload mode or run one teacher-led model with students suggesting missing examples.

Vocabulary¶

Bias — a pattern of unfair or unreliable errors
Representative data — examples that include the important variety a system will face
Edge case — an unusual but important example a system still needs to handle
Accuracy — how often a system is correct on a test
Fair test — a test using examples the system should reasonably handle

Reflection — fill in after class¶

Which lopsided setup produced the clearest failure?
Did students understand bias as data-related, not just attitude-related?
Did any pair create a strong fairness test worth reusing?
Should future Teachable Machine lessons assume cameras or upload/teacher-led mode?

Next week: Week 12 — How Language Models Actually Work