Generative AI Leader Use AI to make hiring decisions Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Pre-train the model on a dataset that is balanced across all demographics

  Why it's wrong here

  Balancing pre-training data does not satisfy Google's AI Principles, which require testing for unfair bias and ensuring human oversight rather than assuming a balanced corpus removes disparate impact. It is tempting because dataset balance is a recognised mitigation, and it would be the correct choice when curating training corpora for representation.

• ✗

  Blind the model to demographic features to ensure fairness

  Why it's wrong here

  Blinding demographic features does not prevent bias, because correlated proxies such as postcodes or university names still encode group membership. It is tempting because removing protected attributes appears fair, and it would be the correct choice only where no proxy variables exist and outcomes are independently audited.

• ✗

  Only use the model for initial resume screening, with final decisions by humans

  Why it's wrong here

  Human review of every final decision does not by itself satisfy Google's AI Principles, which require proactive bias testing, diverse data and accountability throughout the pipeline, not merely a human sign-off. It is tempting because human oversight is a recognised safeguard, and it would be the correct choice where the model only ranks candidates and reviewers can override it.

• ✓

  Evaluate the model using diverse test sets and adjust if bias is found

  Why this is correct

  Testing the hiring model on diverse demographic test sets and correcting any measured disparity aligns with Google's fairness principle, which requires avoiding reinforcement of unfair bias. Empirical evaluation before deployment is the most defensible approach for high-stakes hiring decisions.

Go deeper

Related to this question