AI Associate Ethical AI and Data Privacy 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.

• ✓

  Exclude gender from the feature set used for model training

  Why this is correct

  Excluding protected attributes is a direct way to prevent the model from using them; it aligns with data minimisation.

• ✗

  Use gender as a feature but ignore the model predictions for certain groups

  Why it's wrong here

  Ignoring predictions is not a systematic solution; the model will still be biased.

• ✗

  Allow gender in training but use a post-processing technique to adjust scores

  Why it's wrong here

  Allowing gender in training and applying post-processing to adjust scores fails to prevent the model from learning biased correlations during training, as the algorithm can still encode gender-based patterns in its internal weights. This approach is tempting because post-processing techniques, such as threshold adjustments, are effective for mitigating bias in a model’s final outputs when the sensitive attribute is already present in the training data. It would be the correct choice if the data scientist had no control over training features but could still access gender labels to recalibrate predictions after inference.

• ✗

  Include gender as a feature and then apply a fairness constraint during training

  Why it's wrong here

  Including gender as a feature directly contradicts the objective of preventing the model from using it at all. The scenario aims to ensure gender is not inadvertently utilised, implying it should be excluded from the feature set. Fairness constraints are applied to mitigate bias *when* a sensitive feature is present, ensuring equitable predictions across groups. This approach would be appropriate if gender was a necessary input, but its potentially discriminatory influence needed to be controlled.

Go deeper

Related to this question