Question

You deployed a loan-approval model on Vertex AI. Regulators require you to demonstrate that approval rates are equitable across racial groups. However, for legal and privacy reasons, your company never collects race as a feature and does not store it anywhere. Your compliance team still needs you to monitor the model for potential disparate impact across racial groups. What is the most appropriate approach?

Accepted Answer

Use a proxy-attribute inference technique such as Bayesian Improved Surname Geocoding (BISG) to statistically estimate group membership, then evaluate approval-rate disparities on that estimated attribute in a controlled fairness audit.

Answer

Add race as an input feature to the model so it can be logged and monitored directly by Vertex AI Model Monitoring.

Answer

Skip fairness monitoring entirely, since the model cannot be biased on an attribute it never receives as input.

Answer

Enable Vertex AI Model Monitoring feature-attribution drift on the existing features and treat any attribution drift as evidence of racial bias.