Generative AI Leader Practice Question: A company develops a generative AI model for…

Answer analysis

Option-by-option breakdown

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

• ✓

  Audit the training data for demographic representativeness

  Why this is correct

  Auditing training data for demographic representativeness identifies whether skewed historical hiring data caused the disparate rejection rates. Fixing the data source precedes mitigation techniques such as reweighting, threshold adjustment or post-processing, which treat symptoms rather than the underlying cause.

• ✗

  Reduce the model's complexity to avoid overfitting to biased patterns

  Why it's wrong here

  Reducing complexity does not remove the historical bias encoded in the training data or labels; it can simply reduce overall accuracy while leaving disparate impact intact. It tempts as a general remedy for overfitting, and would be reasonable when a model memorises noise rather than generalisable patterns.

• ✗

  Apply adversarial debiasing to the model

  Why it's wrong here

  Adversarial debiasing alters model training to suppress learned bias, but it cannot diagnose whether the disparity originates in the data, labels or proxy features. It tempts because it is a recognised mitigation technique, and would be appropriate after the root cause has been identified through measurement and auditing.

• ✗

  Collect more data from all demographic groups equally

  Why it's wrong here

  Collecting more data equally across groups does not correct bias already present in labels, historical decisions or proxy features; it can entrench the same patterns at greater scale. It tempts because data volume often improves performance, and would help where underrepresentation, not label bias, causes the disparity.

Go deeper

Related to this question