Is it fair to everyone?
An AI that learned from one-sided data works well for some people and badly for others. People have to check it and fix it.
Story
The school got a machine that opens the door when it hears a voice. When Jun said, “Open up!” it opened right away. But when Mina, whose voice is quiet and high, said it, the door hardly ever opened. Mina was upset. “Why doesn’t it work for me?” The machine had learned only from voices like Jun’s.
Let’s learn
An that saw only red apples does not know green apples. The same thing happens with people.
If you teach an AI with lots of from only some people, it works well only for those people.
- that learned only from grown-up voices does not understand children well.
- A translator that learned from just one language sounds strange when it translates other languages.
- that learned only from photos of one group often gets other people’s faces wrong.
AI does not do this on purpose. It happens because the data was one-sided. But the people it gets wrong feel upset and lose out. That is not fair.
So the people who make AI have to put in a good mix of data from many kinds of people, and test that it works well for everyone. And we who use it can ask, “Is this fair to everyone?”
At a glance
-
Collect a good mix
-
Test with everyone
-
Fix the one-sided parts
In simple words
If a shoe store had only one size, only people with feet that size could wear the shoes. For everyone to wear shoes, the store needs many sizes.
Try it
Think about ‘a door for everyone.’
- Imagine you are making a ‘voice door’ for your school.
- Write down everyone who will use this door. There are children, grandmothers, friends with a cold and friends who came from other countries, too.
- Decide whose voices you should collect to teach it.
- How could a friend who has trouble speaking open the door?
Quiz
If the sentence is right, choose [True]. If it is wrong, choose [False].
-
1AI is a machine, so it is always fair to everyone.
That’s right!
Not quite. Read the reason.
If it learns from one-sided data, AI becomes one-sided, too. Being a machine does not make it fair by itself.
-
2Putting in a good mix of data from many kinds of people makes AI fairer.
That’s right!
Not quite. Read the reason.
You have to put in a good mix and test that it works well for everyone.
-
3Checking whether AI is fair is a job for people.
That’s right!
Not quite. Read the reason.
The people who make it and the people who use it have to check together.
Correct answers / 3
Remember
- One-sided data makes unfair AI.
- AI needs a good mix of data from many kinds of people.
- Checking that it is fair is a job for people.
Today’s words
Tap a word to see what it means.
Go deeper
These pages are for older students and grown-ups. Read them together with a grown-up.