The correct option is Skewed training data that encodes past hiring bias and teaches unfair patterns. The model was trained on many years of historical hiring decisions from a field with persistent underrepresentation, which makes it very likely that it learned to reproduce those biased patterns even when candidates have equivalent qualifications.

Historical decisions often encode human and systemic biases in the labels. Training on those labels teaches the model to associate proxies for protected characteristics with lower potential. The fairness finding that candidates from certain communities are scored lower despite equivalent education and experience strongly indicates label and representation bias in the training set rather than an issue that emerged after deployment. This type of bias would be present from day one and will persist unless data, labels, or training objectives are corrected.

Production data drift that changed resume patterns since deployment is unlikely because the disparity appears despite equivalent qualifications and aligns with long standing historical underrepresentation. That points to biased training signals rather than a sudden change in incoming resumes after launch.

Lack of model interpretability which makes it hard to see why predictions were made is not a root cause. Limited explainability can hinder diagnosis and remediation, yet the unfair scoring stems from what the model learned from biased data, not from whether we can explain the predictions.

Vertex AI Model Monitoring is a monitoring capability and not a cause of biased outcomes. It can help detect drift or skew and surface quality issues, but it does not create the unfair pattern described.

Full AWS Practitioner Certification Question