Beyond Accuracy: Evaluating AI Classifiers for Fairness Across Subgroups

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ethics-safety-AI-classifier-fairness-evaluation-adults

1. A machine learning model reports 95% overall accuracy, but separate analysis shows 70% accuracy for one demographic group and 99% for another. What does this scenario illustrate?

   1. Aggregate accuracy metrics can hide demographic-specific performance failures
   2. The model is fundamentally broken and cannot be used
   3. Overall accuracy is the most reliable metric for deployment decisions
   4. Demographic information should be removed from datasets before evaluation

2. When conducting a subgroup fairness evaluation for an AI classifier, what is the most important factor in determining which subgroups to analyze?

   1. Groups that have the largest sample sizes in the dataset
   2. Groups that are most convenient to measure and report on
   3. Any group explicitly listed in current regulatory frameworks
   4. Which groups are relevant to the use case and could experience harm from model errors

3. Beyond overall accuracy, which additional metrics should be calculated for each subgroup during a fairness evaluation?

   1. Training loss, validation loss, and overfitting metrics
   2. Model confidence scores and inference time
   3. False positive rate, false negative rate, and calibration metrics
   4. Precision, recall, and F1 score only

4. When subgroup accuracy metrics diverge significantly during evaluation, what should investigators examine FIRST?

   1. The programming language the model was written in
   2. The model's hyperparameters and learning rate settings
   3. The hardware used for model inference
   4. Whether the training data has unequal representation across subgroups

5. Why are sample size requirements important when calculating subgroup fairness metrics?

   1. Sample size affects model training speed but not fairness
   2. Regulatory frameworks mandate minimum sample sizes for all demographic groups
   3. Large sample sizes always indicate better data quality
   4. Small subgroups produce statistically unreliable estimates with wide confidence intervals

6. A fairness researcher claims their model satisfies demographic parity, equal opportunity, and calibration simultaneously. What is the fundamental problem with this claim?

   1. Calibration only applies to regression models, not classifiers
   2. These fairness metrics often conflict with each other and cannot all be satisfied in practice
   3. The researcher has not used enough training data
   4. Demographic parity is not a real fairness metric

7. Why is achieving statistical fairness metrics insufficient for ensuring substantive equity in AI systems?

   1. Statistical fairness can be satisfied while groups still experience different types and magnitudes of harm
   2. Statistical fairness metrics require too much computational resources
   3. Substantive equity cannot be measured quantitatively
   4. AI systems are inherently neutral and do not create equity issues

8. Why is it insufficient to only examine overall accuracy when conducting subgroup fairness evaluation?

   1. Accuracy is not a standard machine learning metric
   2. Accuracy measurements vary too much between runs to be useful
   3. Overall accuracy is too difficult to calculate precisely
   4. Overall accuracy can be high while error types differ dramatically across groups

9. A hiring AI achieves demographic parity with equal selection rates across gender groups, but investigation reveals it selects fewer women for technical roles. What does this demonstrate about demographic parity?

   1. Sample size issues caused the apparent discrepancy
   2. The model is actually fair despite the investigation findings
   3. Equal selection rates can mask substantive differences in the types of positions or outcomes groups receive
   4. Demographic parity is the strongest form of fairness protection

10. What is a key reason to examine false positive rate and false negative rate separately for each subgroup?

    1. The harm from each error type often differs by group—false positives may harm one group while false negatives harm another
    2. Separating these rates violates privacy regulations
    3. False positive rate and false negative rate are always equal in well-trained models
    4. This examination is only required for government-regulated AI systems

11. When a fairness evaluation reveals significant subgroup performance gaps, which remediation approach directly addresses the training data?

    1. Hiring more machine learning engineers
    2. Adjusting the model's output threshold for each subgroup at inference time
    3. Resampling or reweighting the training data to improve subgroup representation
    4. Increasing the model's neural network depth

12. What does calibration measure about an AI classifier?

    1. How fast the model makes predictions
    2. Whether the model produces the same outputs regardless of input order
    3. The degree to which the model avoids false positives
    4. Whether the model's confidence scores match the actual frequency of positive outcomes

13. A classifier satisfies demographic parity but fails equal opportunity. What does this indicate about the model's predictions?

    1. Different demographic groups are selected at equal rates, but qualified individuals from some groups are less likely to be selected
    2. The model only makes errors on one demographic group
    3. The model has perfect accuracy
    4. The model is randomly assigning outcomes regardless of qualifications

14. For a medical screening AI that could miss serious conditions, which fairness consideration is most critical?

    1. Ensuring false negative rates are equal across demographic groups, since missing a diagnosis is particularly harmful
    2. Prioritizing model speed over accuracy
    3. Ensuring demographic parity in screening referrals
    4. Using the simplest possible model for interpretability

15. A classifier shows excellent accuracy across all subgroups but poor calibration for one group. Why is this still a fairness concern?

    1. Poor calibration only affects model training, not predictions
    2. The group receives unreliable confidence scores, leading to inconsistent or unjustified decisions for its members
    3. The group will receive faster predictions due to calibration issues
    4. This indicates the model has too many parameters