Bias in AI Systems · seed 1 · A4, ink-friendly. The answer key prints on its own page for grown-ups.

Bias in AI Systems

Computing · Artificial Intelligence · ages 9-11
Name ______________________   Date ____________

  1. A face-recognition app was trained mostly on light-skinned faces. What is most likely to happen?

    • It works equally well for all skin tones
    • It works less well for darker skin tones
    • It stops working completely

  2. What does 'bias in AI' mean?

    • The AI works the same for everyone
    • The AI works better for some groups than others because of its training data
    • The AI never makes mistakes

  3. AI can be biased because of the data it was trained on.

    Circle one:   True   False

  4. Look at the chart. The AI scores well for Group A but poorly for Group B. What most likely caused this gap?

    Group A95Group B60
    • Group B's faces are harder to see
    • The training data had far fewer examples from Group B
    • The AI dislikes Group B

  5. A translation tool always changes 'the doctor' into a male word. Why does this happen?

    • The AI understood the sentence's meaning
    • The AI copied a pattern from text where doctors were mostly written as men
    • The AI made a random guess

  6. Which change would most likely reduce bias in a face-recognition AI?

    • Training it on more photos of many different skin tones
    • Making the app run faster
    • Adding more storage space to the computer

  7. A voice assistant understands adults well but often misunderstands children's voices. What is the most likely cause?

    • The assistant's microphone is broken
    • Children speak a different language
    • The training data had far fewer recordings of children's voices

  8. A team says their AI is fair because it scored 90% accuracy overall. What might they be missing?

    • The AI could still score much worse for specific groups, even with a high overall average
    • A 90% score always means every group is treated fairly
    • Overall accuracy is the only number that matters
LightMySky · lightmysky.comW1-mt__BbOjiY5A5-s1

Answer key

For grown-ups. Fold this page away before handing over the rest.

Bias in AI Systems W1-mt__BbOjiY5A5-s1

  1. It works less well for darker skin tones · With fewer examples of darker skin tones to learn from, the app makes more mistakes on those faces.
  2. The AI works better for some groups than others because of its training data · Bias means the training examples did not fairly represent everyone, so the results come out uneven.
  3. True · Correct. If training data leaves out certain groups, the AI learns those groups less well.
  4. The training data had far fewer examples from Group B · When a group is underrepresented in training data, the AI has less to learn from and performs worse for that group.
  5. The AI copied a pattern from text where doctors were mostly written as men · The AI is not reasoning about gender. It is repeating a pattern that was common in the text it trained on.
  6. Training it on more photos of many different skin tones · More diverse training examples help the AI learn patterns for every group, not just some.
  7. The training data had far fewer recordings of children's voices · Just like with skin tones or gendered words, a group with fewer training examples gets less accurate results.
  8. The AI could still score much worse for specific groups, even with a high overall average · An overall average can hide big gaps between groups, so testing each group separately matters.
Worksheet · LightMySky