How would you programmatically monitor a deployed model for demographic bias?

WHAT THIS TESTS: This question probes whether you can move beyond academic fairness definitions and build a production system that continuously surfaces demographic disparities. The interviewer cares about telemetry design, metric selection, and operational response, not just knowing the names of bias metrics.

A GOOD ANSWER COVERS: First, define the protected attributes and the granularity of analysis, such as race, gender, or age bracket, and acknowledge when direct labels are unavailable so you must use legitimate proxies or self-reported data with privacy safeguards. Second, select group-aware metrics that match the business context; for binary classification this often includes statistical parity difference, equalized odds gap, and calibration error across groups, while for regression you might compare mean absolute error by subgroup. Third, implement stratified logging so every prediction is tagged with the protected attribute and ground truth when available, then compute these metrics in a streaming or batch pipeline with a sliding window. Fourth, set dynamic thresholds based on historical variance rather than static limits, and wire alerts into an incident management system so breaches trigger a human review workflow. Fifth, version your fairness policies and model cards so stakeholders can audit what threshold was in force at any point in time.

COMMON WRONG ANSWERS: A major red flag is proposing a one-time pre-deployment audit with no ongoing monitoring, because data drift and population shifts make fairness a moving target. Another mistake is tracking only aggregate accuracy while ignoring subgroup breakdowns, which masks disparate performance. Candidates also err by assuming demographic labels are always available; in practice you often need proxy inference or user-consented data, and failing to discuss privacy or consent signals naivety. Finally, suggesting manual spreadsheet reviews instead of automated pipelines shows you have not operationalized the concept.

LIKELY FOLLOW-UPS: The interviewer may ask how you would handle missing demographic data without violating privacy, or how you balance fairness constraints against business metrics like revenue or click-through rate. They might also probe whether you would block predictions automatically when bias thresholds are breached, or merely alert and escalate, and how you would design an A/B test to validate a retrained fairer model.

ONE CONCRETE EXAMPLE: Imagine a lending model where approval rates drop for a specific zip code that correlates with a minority group. You instrument the prediction service to log credit tier, requested amount, and inferred demographic segment via a proxy model trained on surname and geography with documented confidence intervals. Each hour a batch job computes the statistical parity gap and equalized odds difference between this segment and the control group; if the parity gap exceeds three percent for two consecutive windows, a PagerDuty alert fires and the MLOps on-call must decide whether to roll back to the previous model version or throttle traffic while the data science team investigates. All thresholds and proxy definitions are stored in a versioned model card that compliance audits quarterly.