AI Bias Quiz

Answers and explanations appear as you take the quiz.

1. Where can bias enter an AI system according to the guide?
   • Data, measurement, modeling choices, human judgments, and the surrounding workflow
   • Only through a sensitive column
   • Only through the user interface
   • Only after a model is deployed
2. Why can removing a sensitive attribute fail to eliminate bias?
   • Sensitive attributes are always the only source
   • Other features or historical labels may encode related patterns
   • Removing columns guarantees fairness
   • Model evaluation no longer needs groups
3. A dataset’s label reflects a past human decision. What should the team ask?
   • Whether the sensitive column was deleted
   • Whether the label is easy to predict
   • Whether the label captures the intended outcome or an old process
   • Whether the model has enough layers
4. Why report sample sizes and uncertainty for subgroup metrics?
   • An average score is always sufficient
   • Every subgroup has the same sample size
   • Uncertainty makes group testing impossible
   • Small groups can have unreliable estimates and averages can hide disparities
5. Why can choosing one fairness criterion leave other concerns unresolved?
   • One statistical criterion settles every legal concern
   • Different criteria answer different questions and may conflict
   • Any single metric guarantees fairness in every context
   • Choosing a metric removes the need for affected-community input
6. Who should help select fairness criteria for a consequential system?
   • Only the engineering team
   • Only the model vendor
   • Domain experts and people affected by the system
   • A language model without review
7. Why review the workflow around a model?
   • How predictions are used and challenged can change the distribution of harm
   • The model output fully determines fairness
   • Appeals cannot affect outcomes
   • Workflow has no effect after training
8. What should a team do after applying a bias mitigation?
   • Hide subgroup results
   • Assume the mitigation certifies fairness
   • Stop monitoring
   • Evaluate its intended effect and possible new problems