Prepare and test your skills

Prepare and test your skills

Free sample question with answers and detailed explanation below. The Cloud Digital Leader practice bank has an additional 3,070 practice questions.Try 10 free realistic questions: instant score, no credit card, and no account required.
A university is developing a machine learning model to assist with evaluating student scholarship applications. To promote fairness and prevent discriminatory outcomes against traditionally underrepresented groups, the project team wants to audit their raw dataset before building and training the model to check for the following potential issues:
Which strategy should the team implement to identify these sources of bias?
Cloud Logging is an observability and operations service that records system events, user access, and runtime decisions across cloud workloads. Within an AI governance framework, it is used to establish accountability by maintaining an immutable audit trail of how an AI system operates in production, who interacted with it, and what decisions the model generated over time.
The university had already deployed the scholarship AI system into production and needed to establish accountability by recording an audit trail of every automated decision and system event for compliance review.
In AI governance, Cloud Logging and Error Reporting help teams monitor deployed models and investigate operational anomalies, while pre-training dataset audits rely on data bias metrics and data lineage tracking to uncover bias at the source.
Differential privacy is a data protection and governance technique that introduces controlled mathematical noise into a dataset or query result. Its primary purpose is to protect individual privacy by making it statistically difficult to reverse-engineer or identify specific individual data points, while still preserving the overarching statistical patterns required for a machine learning model to learn effectively.
The scenario asked how the university could protect the privacy of individual student applicants during model training so that specific student records could not be reconstructed or identified from the trained model.
While both fairness and privacy are core pillars of Google's responsible AI governance framework, they solve fundamentally different problems. Differential privacy ensures confidentiality, whereas data bias metrics ensure equitable representation across demographic groups.
Data bias metrics are specialized pre-training evaluation tools used in a machine learning workflow to detect whether raw training data contains systemic imbalances or historical prejudices before a model is built. By dividing a dataset into distinct demographic subsets—known as slices—teams can mathematically quantify disparities in how different groups are represented and labeled in the ground-truth data.
Evaluating data bias metrics across dataset slices is the only approach that directly inspects raw sample counts and ground-truth label distributions prior to model training. Other governance mechanisms—such as post-training prediction checks, privacy obfuscation, or runtime auditing—either occur too late in the lifecycle or address security and privacy rather than dataset fairness.
Model bias metrics are post-training evaluation measurements used to determine whether a trained machine learning model produces biased or inequitable predictions across different subsets of data. By comparing confusion matrices (which track true positives, false positives, true negatives, and false negatives) between two data slices, teams can measure performance gaps using indicators like Accuracy Difference, Recall Difference, and Specificity Difference.
The team had already trained their scholarship evaluation model and wanted to perform a post-training evaluation to verify whether the model's predictions were less accurate or had a higher false-negative rate for applicants from a specific geographic state or demographic group.
A comprehensive responsible AI strategy uses both types of fairness indicators sequentially: data bias metrics are used prior to training to validate the raw dataset, and model bias metrics are used after training to validate the model's predictive behavior.
Answer them yourself and get your score the moment you finish. No account needed.
Full access includes
Keep the momentum going with these hand-picked practice scenarios
Last updated