Class 7 · Unit 5: Bias and citizenship · Lesson 5.1 · 40 min

What is bias?

The app is "95% correct". It calls a scabbed Ambri leaf healthy. It wasn't lying — it had just never seen a Kashmir sky.

Today you will: Train a leaf detector that's secretly biased · Catch it by testing group by group · Make it fairer — and prove it

Story

Bilal's uncle sells apples in Sopore. His new phone app promises to spot scab, the dark spots that ruin apple leaves.

He points the phone at a leaf from his own orchard. "Healthy," says the app.

Hiba looks closer. "Mamu, that leaf has spots. It's scab."

The uncle scratches his head. "The company said this app is 95% correct."

Later they find out why. The app was trained almost entirely on photos of one apple variety, taken in bright sunshine on another continent. Bilal's uncle grows Ambri, under a grey Kashmir sky.

The app wasn't lying about its 95%. It was 95% correct on the leaves it had seen.

Watch

When AI is unfair without meaning to be — — scan code 7.5.1 in the printed book.

Warm up your fingers

Skill: typing a short paragraph, then fixing it (Term 3 focus, 7 min)

On typing.com, open an Intermediate paragraph lesson.
Then type this in a new Writer document, without looking at the keys: "Bias means a system is unfair to some people or things. It often comes from the data."
Go back and fix every mistake with the arrow keys and Backspace, not by retyping the whole line.

On the laptop

You need: Teachable Machine in the browser (no account) · LibreOffice Calc · unit-5/leaves-healthy-sunny/ and leaves-scab-sunny/ (20 each) · leaves-test-mixed/ (20: shade, cloud, close-ups, other varieties) · later, leaves-extra-shade/

Mission 1: The biased model (groups of 3–4, 11 min)

☐1
Teachable Machine → Image Project → Standard image model. Don't sign in. Name the classes Healthy and Scab.
☐2
Upload leaves-healthy-sunny into Healthy and leaves-scab-sunny into Scab. Train Model — don't switch tabs.
☐3
Predict: will it do better on sunny test leaves or shady ones? Note it.
☐4
Preview → File, test each photo in leaves-test-mixed. In bias-test.ods, record: Photo · Real answer · Model said · Correct? (Y/N) · Light (sunny / shade / cloudy).
☐5
Count right answers in sun and in shade or cloud: =COUNTIFS(E2:E21;"sunny";D2:D21;"Y").

Mission 2: Make it fairer (groups, 9 min)

☐6
Add the leaves-extra-shade photos to the right classes and train again.
☐7
Test the same photos. Fill a second column: Model said (after).
☐8
Compare. Which group improved most?

No laptop today?

The teacher shows 10 printed leaf cards. The class "trains" a volunteer Model on the 4 sunny cards only, then the Model sorts 6 shade cards. Record the mistakes, add 4 shade cards to training, try again.

Now you know

An app trained on one variety in bright sunshine is tested on Ambri under grey skies. Illustrative results: on sunny leaves it is right 95 percent of the time before and after; on shade and cloudy leaves 40 percent before and 85 percent after adding shade photos to training. Break results down by group; varied data is the first fix.

Bias means results are unfairly tilted — better for some people, places or things than others.
It usually comes from the data (who or what was left out) and from human choices in building it.
Overall accuracy can hide it. "95% correct" can still mean failing one group most of the time.
Always ask: accurate for whom? Break results down by group.
Fair, varied data is often the first fix.

Debate it

The leaf app works well for big orchards that bought expensive cameras, and badly for small farmers with old phones.

Is that just bad luck, or is it unfair? Who should fix it: the farmer, the company, or the government?

Check yourself — practice, not a test

1. What does bias in an AI system mean?
- ○ The AI is broken and gives random answers
- ○ The AI's results are unfairly tilted, working better for some than for others
- ○ The AI is slow
- ○ The AI has an opinion like a person
2. If a model is 95% accurate overall, it must be fair to every group.
- ○ True
- ○ False
3. Where does bias in AI usually come from? (choose all that apply)
- ○ Training data that leaves some groups out
- ○ Choices made by the people who build it
- ○ The colour of the laptop
- ○ The AI deciding it dislikes someone
4. To check a model for bias, we ask: is it accurate for ______?

Remember

Don't just ask "Is it accurate?" Ask "Is it accurate for everyone?"

Story

Bilal's uncle sells apples in Sopore. His new phone app promises to spot scab, the dark spots that ruin apple leaves.

He points the phone at a leaf from his own orchard. "Healthy," says the app.

Hiba looks closer. "Mamu, that leaf has spots. It's scab."

The uncle scratches his head. "The company said this app is 95% correct."

Later they find out why. The app was trained almost entirely on photos of one apple variety, taken in bright sunshine on another continent. Bilal's uncle grows Ambri, under a grey Kashmir sky.

The app wasn't lying about its 95%. It was 95% correct on the leaves it had seen.

Warm up your fingers

Skill: typing a short paragraph, then fixing it (Term 3 focus, 7 min)

On typing.com, open an Intermediate paragraph lesson.
Then type this in a new Writer document, without looking at the keys: "Bias means a system is unfair to some people or things. It often comes from the data."
Go back and fix every mistake with the arrow keys and Backspace, not by retyping the whole line.

On the laptop

Mission 1: The biased model (groups of 3–4, 11 min)

☐1
Teachable Machine → Image Project → Standard image model. Don't sign in. Name the classes Healthy and Scab.
☐2
Upload leaves-healthy-sunny into Healthy and leaves-scab-sunny into Scab. Train Model — don't switch tabs.
☐3
Predict: will it do better on sunny test leaves or shady ones? Note it.
☐4
Preview → File, test each photo in leaves-test-mixed. In bias-test.ods, record: Photo · Real answer · Model said · Correct? (Y/N) · Light (sunny / shade / cloudy).
☐5
Count right answers in sun and in shade or cloud: =COUNTIFS(E2:E21;"sunny";D2:D21;"Y").

Mission 2: Make it fairer (groups, 9 min)

☐6
Add the leaves-extra-shade photos to the right classes and train again.
☐7
Test the same photos. Fill a second column: Model said (after).
☐8
Compare. Which group improved most?

No laptop today?

Now you know

Bias means results are unfairly tilted — better for some people, places or things than others.

It usually comes from the data (who or what was left out) and from human choices in building it.

Overall accuracy can hide it. "95% correct" can still mean failing one group most of the time.

Always ask: accurate for whom? Break results down by group.

Fair, varied data is often the first fix.

Check yourself — practice, not a test

1. What does bias in an AI system mean?

○ The AI is broken and gives random answers
○ The AI's results are unfairly tilted, working better for some than for others
○ The AI is slow
○ The AI has an opinion like a person

2. If a model is 95% accurate overall, it must be fair to every group.

○ True
○ False

3. Where does bias in AI usually come from? (choose all that apply)

○ Training data that leaves some groups out
○ Choices made by the people who build it
○ The colour of the laptop
○ The AI deciding it dislikes someone

4. To check a model for bias, we ask: is it accurate for ______?