Week 16 · Theme 3 Digital Citizenship, Lesson 1 Explorer in Action · 3 lessons
2
Fair data and unfair data
ICT P6 Term 2, Theme 3: pp. 8–13
GoalTell neutral data from biased data and explain why it matters.
New words
neutral data · data that covers all the relevant cases, not only one group or placebiased data · data that comes from only one group or place, so it gives a one-sided picturefair · treating everyone in the same honest waytrain · to teach a computer by showing it examples
Explanation
An AI learns what people teach it, so it is only as good as its examples. Neutral data covers all the relevant cases, not only data from one group or one place. An AI taught with neutral data can learn about everything. Biased data comes from only one group or place. An AI taught with biased data may make mistakes or act unfairly toward everything that was left out. Nobody has to want to be unfair for this to happen; the problem is simply that some examples are missing. Here is an invented case. A fruit sorter is taught only with fruit grown in one region. When it meets fruit from another region it may not know it. A voice helper taught only with adult voices may not understand children. The cure is to ask a question before training: who or what is missing from my examples? Then add more varied examples, and test again.
Kind of data
What the AI learns
neutral: many kinds, places and people
learns about all of them
biased: only one kind, place or group
may fail or be unfair to the rest
Worked examples
only red apples shown→a green apple arrives→AI says ‘not an apple’→add green and yellow apples
An invented apple sorter is taught with pictures of red apples only. A green apple arrives, and it says ‘not an apple’. The data was biased. Adding green and yellow apples fixes the problem.
Trained with
Works well for
Fails for
adult voices only
adults
children
adult and child voices
both
few people
A voice helper is trained only with adult voices. Then a child asks a question, and the helper often misunderstands. The child is not the problem. The training data left out children’s voices.
Who is missing? An invented AI should find out which sport pupils like best. It is trained with answers from one class only. Youssef’s class loves football, but other classes like swimming and running. The AI’s answer would be one-sided. The fix is to include answers from all classes.
Class activity
Fair data hunt. Imagine an AI that learns to recognise animals. List eight animals you would show it. Look at your list: are they all from one place, one size or one colour? Change or add two animals to make your data more neutral. Write one sentence on why this helps the AI.
Practice
1
What is the difference between neutral data and biased data?
Neutral data covers all the relevant cases; biased data comes from only one group or place.
2
Describe a case of biased data and say what goes wrong for an AI.
Answers differ. For example: an apple sorter taught with red apples only does not know green apples.
3
What is the benefit of using neutral data?
So the AI can learn about everything and treat everyone fairly.
4
An AI keeps making mistakes about one group of people. What should you check first?
Its data: is that group missing from the examples?
