Responsible AI Practitioner practice test

Sample questions

Two questions from this practice test, in the same style as the certification exam. Answer before opening the explanation.

1. Question 1

   In a typical AI governance process, when should a new AI use case first go through intake and an initial risk classification?

   1. After a pilot has run, so that the classification can draw on real performance results
   2. Just before go-live, when the final configuration and the vendor contract are known
   3. When the idea or purchase request is raised, before major build or buying commitments
   4. At the first scheduled inventory review after the system has been deployed
   Show answer and explanation

   Correct answer: C. When the idea or purchase request is raised, before major build or buying commitments

   Intake works best at the start: an early classification decides which assessments, controls and contract terms the use case will need, while design and procurement can still change cheaply, and it can be refined as the system takes shape. Waiting for a pilot means the pilot itself, which may already touch real people and data, runs without knowing what it requires. Just before go-live is too late to change the design or the contract without cost and delay. Classifying only after deployment leaves the system running without any assessment.

2. Question 2

   A model card was written when a churn-prediction model launched. Since then the model has been retrained twice, and the latest evaluation shows noticeably weaker performance for customers over 70. What should happen to the model card?

   1. Update it for the current version, adding the new results by age group and the limitation for older customers
   2. Keep it as written, because it documents the version that was originally reviewed and approved at launch
   3. Leave it unchanged and record the weaker results in the team's internal ticket for the next retraining
   4. Replace the group results with the latest overall accuracy, which improved after the second retraining
   Show answer and explanation

   Correct answer: A. Update it for the current version, adding the new results by age group and the limitation for older customers

   A model card is useful only if it describes the model actually in use: each version should carry its own evaluation results, broken down by relevant groups, and its known limitations, so that users and reviewers can decide how far to rely on it. Keeping the launch version leaves readers with outdated claims about the current model. An internal ticket hides a limitation from the people who depend on the card. Replacing group results with an overall figure removes exactly the information that shows who the model serves less well.