Fairness, Accountability and the Limits of a Model · seed 1 · A4, ink-friendly. The answer key prints on its own page for grown-ups.

Accurate overall can still be unfair

Computing · Machine Learning · ages 21-22
Name ______________________   Date ____________
  1. Overall ninety percent, one group fifty. What hides the gap?

    • Using too many decimals
    • Reporting only the overall average
    • Printing the report
  2. What does a model trained on old biased decisions do?

    • Copies the pattern, unfair parts included
    • Automatically repairs the past
    • Refuses to train at all
  3. One system can satisfy every definition of fairness at once.

    Circle one:   True   False

  4. What does accountability require from the team?

    • An answerable owner who can explain each decision
    • A promise that errors never happen
    • A mascot for the project
  5. What is redlining in this lesson?

    • Drawing gym routes in red
    • Refusing areas service by map lines from the past
    • Underlining books in red ink
  6. Which table reveals the unfair gap?

    • Error rates broken out by group
    • A list of employee birthdays
    • Total profit by quarter
  7. A loan tool rejects Maya. What is she owed?

    • The full training dataset to keep
    • A job on the modelling team
    • A clear reason plus a path to challenge it
  8. Your fix meets one fair rule but bends another. What do you report?

    • That maths is now perfect
    • Nothing, averages look fine
    • Which rule you chose and what it gives up
LightMySky · lightmysky.comW1-mt_tTh4WtAyk5-s1

Answer key

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

Accurate overall can still be unfair W1-mt_tTh4WtAyk5-s1

  1. Reporting only the overall average · One average blends the failing group into the passing crowd.
  2. Copies the pattern, unfair parts included · Training copies history as it was, red lines and all.
  3. False · Fair rules clash, so a team must choose openly.
  4. An answerable owner who can explain each decision · Someone must own the choice and explain it to those it touches.
  5. Refusing areas service by map lines from the past · Drawn lines once cut neighbourhoods off, and models can relearn them.
  6. Error rates broken out by group · Per-group errors show where the model fails whom.
  7. A clear reason plus a path to challenge it · Affected people deserve reasons they can act on and appeal.
  8. Which rule you chose and what it gives up · Honest fixes name the chosen rule and its cost.
Worksheet · LightMySky