Bias and Fairness in AI: The Honest Picture

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ai-foundations-bias-fairness-final1-creators

1. Through which channels does AI bias enter a model?

   1. Only via the user prompt
   2. Training data, labeling, and optimization choices
   3. Only at deployment
   4. Only through the model name

2. Is bias 'solvable by good intentions' alone?

   1. Yes, intent is enough
   2. Only on weekends
   3. No — measurable mitigation requires systematic work, not intent
   4. Only with prayer

3. Which is a legitimate use of AI to study bias?

   1. Predicting an individual's race from their name
   2. Picking a single 'best' demographic
   3. Removing all demographic data and assuming fairness
   4. Measuring outcome differences across demographic groups

4. What does 'auditing training data for representation gaps' mean?

   1. Checking which groups are under- or over-represented in the data
   2. Counting total rows only
   3. Measuring file size
   4. Sorting alphabetically

5. What does a counterfactual prompt do?

   1. Asks the model to lie
   2. Varies one demographic feature at a time to surface systematic differences
   3. Doubles the cost
   4. Disables the model

6. Where should known model limits live for users to find?

   1. Internal Slack only
   2. Engineer's head only
   3. Model cards and product docs
   4. Nowhere

7. Can prompt-level 'debiasing' alone eliminate bias?

   1. Yes, prompt is enough
   2. Only with multilingual prompts
   3. Only on Tuesday
   4. No — it cannot fix bias baked into data and weights

8. Can data-set bias be cleanly fixed after the fact?

   1. No — post-hoc cleanup is partial at best
   2. Yes, perfectly
   3. Only by retraining from zero on perfect data
   4. Only by editing the model file

9. Which review function is irreplaceable by AI?

   1. An automated red-team script
   2. Diverse evaluation teams and stakeholder review
   3. A model card
   4. A metric dashboard

10. What concrete bias-audit experiment is recommended?

    1. Run one prompt and call it good
    2. Skip prompts entirely
    3. Run 50 prompts varying only the implied demographic of the subject and look for systematic differences
    4. Always trust the default output

11. In which domains is bias a legal issue, not optional polish?

    1. Only games
    2. Only weather
    3. Only cooking
    4. Hiring, lending, healthcare, and housing

12. What does 'disparate impact' require to apply legally?

    1. No requirement of intent — measurable disparity is enough
    2. Documented intent to discriminate
    3. A jury trial
    4. A presidential order

13. What treatment does recommended for bias in regulated domains?

    1. Treat it as a marketing exercise
    2. Treat it as compliance work, not optional polish
    3. Skip it
    4. Wait until lawsuits arrive

14. Which of these is a misleading 'fix'?

    1. Running an outcome-difference audit
    2. Building stakeholder review into release
    3. Adding a 'be fair' instruction to the prompt and calling it solved
    4. Documenting limits

15. Which mindset best fits responsible AI fairness work?

    1. Bias is mythical
    2. Bias was solved in 2020
    3. Bias only affects others
    4. Bias is real, measurable, and requires ongoing auditing