100% Latest & Updated Google Generative AI Leader Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 530
A legal assistant has no access to the clause asked about. What should it do instead of guessing?
- Answer abstention
- Model confidence calibration
- Hallucination
- Source citation verification
Explanation
Correct Answer
A:An evidence-aware system should state its limitation. Abstention allows a system to say evidence is insufficient instead of inventing facts.
Incorrect Answers
B:Calibration examines whether confidence signals correspond to actual correctness. This does not directly meet the case requirement to avoid pretending to know inaccessible source text.
C:A hallucination is a model output that asserts unsupported or factually incorrect information. This does not directly meet the case requirement to avoid pretending to know inaccessible source text.
D:Citation verification checks whether a specific source genuinely supports the statement attributed to it. This does not directly meet the case requirement to avoid pretending to know inaccessible source text.
Question 531
A hiring model learns patterns from past discriminatory recruitment decisions. What underlying problem is present?
- Training-data bias
- Outcome fairness evaluation
- Representative sampling
- Continuous bias monitoring
Explanation
Correct Answer
A:Historical labels can transmit past unfairness into the model. Training-data bias occurs when source examples encode systematic skew or historical discrimination.
Incorrect Answers
B:Fairness evaluation compares model behavior across relevant population groups. This does not directly meet the case requirement to identify biased historical training examples.
C:Representative sampling ensures training and test examples cover the intended deployment population. This does not directly meet the case requirement to identify biased historical training examples.
D:Monitoring checks whether disparate outcomes emerge or worsen after deployment. This does not directly meet the case requirement to identify biased historical training examples.
Question 532
A credit-screening system rejects one demographic group at a much higher rate. What assessment is needed?
- Continuous bias monitoring
- Human escalation
- Representative sampling
- Outcome fairness evaluation
Explanation
Correct Answer
D:Outcomes should be compared using appropriate fairness measures. Fairness evaluation compares model behavior across relevant population groups.
Incorrect Answers
A:Monitoring checks whether disparate outcomes emerge or worsen after deployment. This does not directly meet the case requirement to evaluate disparate decisions across groups.
B:Human escalation routes sensitive or contested model decisions to qualified people. This does not directly meet the case requirement to evaluate disparate decisions across groups.
C:Representative sampling ensures training and test examples cover the intended deployment population. This does not directly meet the case requirement to evaluate disparate decisions across groups.
Question 533
A clinical model was trained on one age group but is used for patients of all ages. What data issue should be addressed?
- Continuous bias monitoring
- Outcome fairness evaluation
- Representative sampling
- Human escalation
Explanation
Correct Answer
C:Population coverage must match deployment needs. Representative sampling ensures training and test examples cover the intended deployment population.
Incorrect Answers
A:Monitoring checks whether disparate outcomes emerge or worsen after deployment. This does not directly meet the case requirement to include representative age groups in evaluation.
B:Fairness evaluation compares model behavior across relevant population groups. This does not directly meet the case requirement to include representative age groups in evaluation.
D:Human escalation routes sensitive or contested model decisions to qualified people. This does not directly meet the case requirement to include representative age groups in evaluation.
Question 534
Two annotators consistently apply different definitions of 'high risk' to similar people. What data process should be corrected?
- Label quality audit
- Outcome fairness evaluation
- Training-data bias
- Representative sampling
Explanation
Correct Answer
A:Labels need agreed definitions and consistent application. A label audit checks annotation consistency and whether categories embed unfair judgments.
Incorrect Answers
B:Fairness evaluation compares model behavior across relevant population groups. This does not directly meet the case requirement to audit category annotation consistency.
C:Training-data bias occurs when source examples encode systematic skew or historical discrimination. This does not directly meet the case requirement to audit category annotation consistency.
D:Representative sampling ensures training and test examples cover the intended deployment population. This does not directly meet the case requirement to audit category annotation consistency.
Question 535
An automated benefits decision is contested by a vulnerable applicant. What control supports accountability?
- Representative sampling
- Continuous bias monitoring
- Human escalation
- Training-data bias
Explanation
Correct Answer
C:A human escalation route can provide review and correction. Human escalation routes sensitive or contested model decisions to qualified people.
Incorrect Answers
A:Representative sampling ensures training and test examples cover the intended deployment population. This does not directly meet the case requirement to send disputed sensitive cases to a person.
B:Monitoring checks whether disparate outcomes emerge or worsen after deployment. This does not directly meet the case requirement to send disputed sensitive cases to a person.
D:Training-data bias occurs when source examples encode systematic skew or historical discrimination. This does not directly meet the case requirement to send disputed sensitive cases to a person.
Question 536
A deployed hiring assistant becomes more discriminatory after a new applicant pipeline is introduced. What should detect it?
- Label quality audit
- Representative sampling
- Human escalation
- Continuous bias monitoring
Explanation
Correct Answer
D:Monitoring tracks changes in subgroup outcomes over time. Monitoring checks whether disparate outcomes emerge or worsen after deployment.
Incorrect Answers
A:A label audit checks annotation consistency and whether categories embed unfair judgments. This does not directly meet the case requirement to observe fairness degradation in production.
B:Representative sampling ensures training and test examples cover the intended deployment population. This does not directly meet the case requirement to observe fairness degradation in production.
C:Human escalation routes sensitive or contested model decisions to qualified people. This does not directly meet the case requirement to observe fairness degradation in production.
Question 537
A support dataset primarily contains complaints from wealthy neighborhoods and omits others. What bias source is possible?
- Continuous bias monitoring
- Representative sampling
- Outcome fairness evaluation
- Training-data bias
Explanation
Correct Answer
D:Skewed training examples may distort responses. Training-data bias occurs when source examples encode systematic skew or historical discrimination.
Incorrect Answers
A:Monitoring checks whether disparate outcomes emerge or worsen after deployment. This does not directly meet the case requirement to recognize selection bias in source records.
B:Representative sampling ensures training and test examples cover the intended deployment population. This does not directly meet the case requirement to recognize selection bias in source records.
C:Fairness evaluation compares model behavior across relevant population groups. This does not directly meet the case requirement to recognize selection bias in source records.
Question 538
A language model provides much lower-quality service for one language community. Which evaluation helps?
- Training-data bias
- Representative sampling
- Outcome fairness evaluation
- Continuous bias monitoring
Explanation
Correct Answer
C:Cross-group quality tests can expose unequal performance. Fairness evaluation compares model behavior across relevant population groups.
Incorrect Answers
A:Training-data bias occurs when source examples encode systematic skew or historical discrimination. This does not directly meet the case requirement to measure performance by relevant language groups.
B:Representative sampling ensures training and test examples cover the intended deployment population. This does not directly meet the case requirement to measure performance by relevant language groups.
D:Monitoring checks whether disparate outcomes emerge or worsen after deployment. This does not directly meet the case requirement to measure performance by relevant language groups.
Question 539
A speech recognizer was evaluated only on studio-quality English audio but will serve noisy multilingual users. What should change?
- Representative sampling
- Outcome fairness evaluation
- Human escalation
- Training-data bias
Explanation
Correct Answer
A:Test data must represent actual deployment conditions. Representative sampling ensures training and test examples cover the intended deployment population.
Incorrect Answers
B:Fairness evaluation compares model behavior across relevant population groups. This does not directly meet the case requirement to include intended languages and acoustic environments.
C:Human escalation routes sensitive or contested model decisions to qualified people. This does not directly meet the case requirement to include intended languages and acoustic environments.
D:Training-data bias occurs when source examples encode systematic skew or historical discrimination. This does not directly meet the case requirement to include intended languages and acoustic environments.
Question 540
After retraining, a customer assistant shows higher harmful error rates for a subgroup. What ongoing practice matters?
- Continuous bias monitoring
- Human escalation
- Training-data bias
- Outcome fairness evaluation
Explanation
Correct Answer
A:Fairness drift monitoring can trigger correction. Monitoring checks whether disparate outcomes emerge or worsen after deployment.
Incorrect Answers
B:Human escalation routes sensitive or contested model decisions to qualified people. This does not directly meet the case requirement to track subgroup safety and quality after release.
C:Training-data bias occurs when source examples encode systematic skew or historical discrimination. This does not directly meet the case requirement to track subgroup safety and quality after release.
D:Fairness evaluation compares model behavior across relevant population groups. This does not directly meet the case requirement to track subgroup safety and quality after release.
Generative AI Leader Practice Tests
