Bias Audits That Catch Problems Before Deployment: A Production Audit Pipeline

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ethics-safety-bias-audit-pipeline-adults

1. Which fairness metric is satisfied when the probability of receiving a positive outcome is equal across all demographic groups?

   1. Demographic parity
   2. Equal opportunity
   3. Calibration
   4. Predictive parity

2. What is 'fairness drift' in a production AI system?

   1. A gradual increase in model accuracy over time
   2. A deliberate adjustment of fairness thresholds by operators
   3. The process of retraining models on new data
   4. Changes in fairness metric values due to shifts in real-world data distributions

3. A company chooses demographic parity over equal opportunity for their hiring algorithm. What is the most accurate characterization of this decision?

   1. A purely technical optimization choice
   2. A values-based decision with significant consequences
   3. A decision that can be automated by the AI system
   4. A choice that has no impact on affected individuals

4. Why can AI systems not fully resolve trade-offs between competing fairness metrics?

   1. Because the metrics were poorly designed
   2. Because the algorithms are not sophisticated enough
   3. Because different fairness metrics encode different ethical values that require human judgment
   4. Because production data is insufficient

5. In a production bias audit pipeline, what is the primary function of setting thresholds that trigger human review?

   1. To completely automate fairness decisions
   2. To reduce the need for any human involvement
   3. To flag situations requiring contextual judgment beyond automated metrics
   4. To permanently disable the system when exceeded

6. What is a 'proxy' attribute in fairness monitoring?

   1. The primary output of the AI model
   2. A fairness metric that is no longer in use
   3. A feature that inadvertently correlates with protected attributes
   4. An alternative name for protected attributes

7. What does 'calibration' measure in fairness contexts?

   1. Whether different groups receive similar numbers of positive predictions
   2. Whether predicted probabilities match actual outcomes within each group
   3. Whether the model performs equally well across all groups
   4. Whether the model treats similar individuals similarly

8. Why is stakeholder input necessary when defining what 'fair' means for an AI system?

   1. To increase the model's prediction accuracy
   2. To comply with technical requirements
   3. Because stakeholders can help clean the training data
   4. Because different stakeholders may have legitimately different conceptions of fairness relevant to their communities

9. What is 'disparate impact' in employment-related AI auditing?

   1. A policy or practice that disproportionately harms a protected group even without discriminatory intent
   2. Intentional discrimination against a protected group
   3. The difference in pay between executives and workers
   4. A measure of model accuracy across demographic groups

10. What is the fundamental limitation of running bias audits only at the time of deployment?

    1. They require too much computational resources
    2. They cannot detect issues that emerge from real-world usage patterns, distribution shift, and edge-case interactions
    3. Deployment audits are illegal
    4. Deployment audits are too expensive

11. What should trigger a fairness incident response process?

    1. A single instance of model prediction error
    2. A complaint from a competitor
    3. Automated audit flags exceeding defined thresholds
    4. A change in the company's stock price

12. What role do automated audits play in a production bias audit pipeline?

    1. They continuously monitor production traffic and alert when fairness metrics drift
    2. They determine the final fairness decisions for the organization
    3. They are only run during system development
    4. They replace the need for any human oversight

13. Why is publishing fairness documentation externally important?

    1. It is required by all AI companies
    2. It enables external accountability, builds trust with affected communities, and allows independent verification
    3. It reduces the company's legal liability
    4. It improves the model's accuracy

14. When fairness metrics in production exceed threshold values, what is the appropriate immediate action?

    1. Ignore the alert if the system is still functioning
    2. Initiate human review and follow the incident response process
    3. Retrain the model with the same data
    4. Immediately shut down the system permanently

15. Why might 'equal opportunity' be preferred over 'demographic parity' in some contexts?

    1. Demographic parity has been proven to be illegal
    2. Equal opportunity is mathematically easier to achieve
    3. Equal opportunity focuses on treating similar individuals similarly regardless of group membership, which may align better with merit-based contexts
    4. Equal opportunity requires less computational resources