An AI system governance review reveals that an AI model produces accurate predictions but uses features that encode protected characteristics through non-obvious proxies. What type of control should be implemented?
Select an answer to reveal the explanation.
Short Explanation
You need feature importance analysis followed by controlled ablation testing because it shows which features drive the model and measures how proxy features affect protected class outcomes. This gives you evidence before choosing a fix.
Full Explanation
The correct control is feature importance analysis followed by controlled ablation testing. Feature importance analysis identifies which features drive predictions, and ablation testing removes or neutralizes specific features to measure how much proxy features affect outcomes for protected groups. This is an evidence-based way to diagnose and quantify discriminatory impact before deciding on the appropriate remedy. The model-discontinuation approach, replacing the model with only rule-based systems that avoid all statistical correlations, is disproportionate without evidence that alternatives are better. The post-processing correction-factor approach, adjusting predictions for protected groups after the fact, may mask root-cause bias. The blanket-correlated-feature-removal approach, removing all features with any statistical correlation with protected characteristics, could eliminate legitimate predictive features.