A bank's AI governance policy requires that all high-impact AI models undergo an ethical impact assessment before deployment. The assessment for a mortgage pricing model reveals that the model relies heavily on neighborhood characteristics that correlate with race. What is the MOST appropriate next step?
Select an answer to reveal the explanation.
Short Explanation
Here's the deal — when features correlate with protected attributes, you need a disparate impact analysis, then remove or replace proxy features and revalidate before deployment. Deploying with a disclaimer or just capping the weight doesn't fix the legal risk.
Full Explanation
The correct step is a disparate impact analysis, then remove or replace proxy features and revalidate, because features correlated with protected attributes (race via neighborhood) require quantifying disparate impact, feature remediation such as removing proxies or applying fairness constraints, and revalidation—not disclaimers or arbitrary weight caps. Deploying as designed ignores anti-discrimination law on disparate impact. Adding a disclaimer and limiting neighborhood weight to 5% are insufficient responses to a potential legal violation.