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

Bias in AI Systems

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

  1. A face-matching tool has few useful training photos of darker-skinned faces. Which risk should its team check?

    • Better results for darker-skinned faces because examples are scarce.
    • Less need to test darker-skinned faces after the model is trained.
    • More errors on darker-skinned faces.

  2. Which description best matches bias in an AI system?

    • A repeated pattern treats some groups unfairly.
    • One mistaken label on a photo.
    • A tool sorts training photos into groups by their dates.

  3. Unfair patterns in training data can contribute to bias in an AI system.

    Circle one:   True   False

  4. Each group had 100 photos in this made-up test. The chart shows correct face matches. What can you conclude?

    Group A95Group B60
    • Group B had fewer useful training photos.
    • Group B had more matching errors in this test.
    • Group B's matching errors were caused by poor lighting in its test photos.

  5. A translation tool repeatedly chooses a male word for 'doctor', although the sentences give no gender clues. What could explain this?

    • The English word 'doctor' could tell it the person is a man.
    • A gender pattern in its training text could shape its word choice.
    • Missing gender details could count as evidence of a male doctor.

  6. A team's face-matching training set lacks useful photos of several skin tones. Which change addresses this missing coverage?

    • Add varied training photos that include the missing skin tones.
    • Use a larger test set without changing the training set.
    • Add more copies of the photos already in the set.

  7. A voice assistant often misheard children. Its team adds child-voice recordings and retrains it. How should they check whether the change helped?

    • Check how the retrained version handles the child recordings used in training.
    • Compare adult results from the old version with child results from the new one.
    • Compare both versions on new recordings from each group.

  8. A team reports 90 percent accuracy across several groups combined. Which claim still needs separate group results?

    • The reported score combines results from several groups.
    • The reported overall accuracy is 90 percent.
    • The tool works equally well for the groups tested.
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. More errors on darker-skinned faces. · Few useful examples can leave gaps that lead to more errors for that group. Testing is needed to measure the result.
  2. A repeated pattern treats some groups unfairly. · AI bias concerns a pattern that can treat groups unfairly. A single error or grouping photos by date does not establish that pattern.
  3. True · A system can learn unfair patterns from its examples. Data is one possible source of bias.
  4. Group B had more matching errors in this test. · Group B had fewer correct matches out of the same total, so it had more errors. Training coverage and lighting are not shown.
  5. A gender pattern in its training text could shape its word choice. · A learned gender pattern could influence the translation. The job title and missing details do not establish the person's gender.
  6. Add varied training photos that include the missing skin tones. · Adding varied photos of the missing skin tones addresses training coverage. A larger test set can measure performance but does not fill that training gap.
  7. Compare both versions on new recordings from each group. · New recordings test beyond training examples. Comparing both versions for each group helps separate a version change from a difference between groups.
  8. The tool works equally well for the groups tested. · The prompt gives the combined score, but not each group's accuracy. A high combined score can hide unequal performance.
Worksheet · LightMySky