IAPP AIGP Practice Test 75
Question 740
An insurance model provides a feature attribution chart but explanations change wildly across near-identical cases. The development team must show evidence before promoting the next model version. A material exception remains unresolved. Which containment decision best protects affected parties?
- Withdraw unverifiable performance claims and rerun controlled tests before launch.
- Block unsupported causal claims and route consequential decisions for human review.
- Switch to a validated safe mode when input quality falls outside the assessed operating domain.
- Suspend adverse automated decisions when discriminatory outcomes cannot be justified or corrected.
Explanation
Correct Answer
B:Block unsupported causal claims and route consequential decisions for human review. Correct. It limits further reliance or exposure while the risk remains unresolved. An explanation tool may be misleading unless its reliability and limitations are evaluated.
Incorrect Answers
A:Withdraw unverifiable performance claims and rerun controlled tests before launch. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Testing evidence and reproducibility: Reproducibility makes AI testing evidence reviewable rather than anecdotal. The present case instead concerns Interpretability testing.
C:Switch to a validated safe mode when input quality falls outside the assessed operating domain. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Adversarial robustness testing: Security and robustness tests must cover plausible operating and adversarial conditions. The present case instead concerns Interpretability testing.
D:Suspend adverse automated decisions when discriminatory outcomes cannot be justified or corrected. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Bias measurement and mitigation: Historic training labels can reproduce inequity even when protected fields are omitted. The present case instead concerns Interpretability testing.
Question 741
An insurance model provides a feature attribution chart but explanations change wildly across near-identical cases. Model assurance reviewers need a documented control before approving continued experimentation. Before selecting a remedy, what should the accountable reviewer investigate first?
- Check entity, time and duplicate overlap that may leak future information or memorized records.
- Assess explanation stability, fidelity and whether the method matches model and use-case limitations.
- Measure sample coverage, label consistency, subgroup representation and production-population mismatch.
- Audit labeling guidance, disagreement patterns, annotator context and ambiguous categories.
Explanation
Correct Answer
B:Assess explanation stability, fidelity and whether the method matches model and use-case limitations. Correct. It establishes the relevant factual risk boundary before implementation. An explanation tool may be misleading unless its reliability and limitations are evaluated.
Incorrect Answers
A:Check entity, time and duplicate overlap that may leak future information or memorized records. Not selected. That check primarily diagnoses a different governance issue. It belongs with Dataset leakage and temporal splits: A contaminated test set can give convincing but invalid evidence of generalization. The present case instead concerns Interpretability testing.
C:Measure sample coverage, label consistency, subgroup representation and production-population mismatch. Not selected. That check primarily diagnoses a different governance issue. It belongs with Data quality and representativeness: Training data quality includes representativeness for the intended use population. The present case instead concerns Interpretability testing.
D:Audit labeling guidance, disagreement patterns, annotator context and ambiguous categories. Not selected. That check primarily diagnoses a different governance issue. It belongs with Label quality and inter-rater agreement: Ground-truth quality shapes model behavior and can hide systematic harm. The present case instead concerns Interpretability testing.
Question 742
An insurance model provides a feature attribution chart but explanations change wildly across near-identical cases. Model assurance reviewers need a documented control before approving continued experimentation. What action should the accountable team prioritize to govern this use responsibly?
- Define and refine labeling policy with expert adjudication and representative agreement testing.
- Validate explanation methods and document what attributions do and do not establish.
- Add robustness tests, input controls and fallback behavior for unsafe predictions.
- Inspect collection origin, licenses, consent where applicable, lineage and transformation history.
Explanation
Correct Answer
B:Validate explanation methods and document what attributions do and do not establish. Correct. It directly reduces or governs the specific documented risk. An explanation tool may be misleading unless its reliability and limitations are evaluated.
Incorrect Answers
A:Define and refine labeling policy with expert adjudication and representative agreement testing. Not selected. That action treats another control objective, not the specific gap described. It belongs with Label quality and inter-rater agreement: Ground-truth quality shapes model behavior and can hide systematic harm. The present case instead concerns Interpretability testing.
C:Add robustness tests, input controls and fallback behavior for unsafe predictions. Not selected. That action treats another control objective, not the specific gap described. It belongs with Adversarial robustness testing: Security and robustness tests must cover plausible operating and adversarial conditions. The present case instead concerns Interpretability testing.
D:Inspect collection origin, licenses, consent where applicable, lineage and transformation history. Not selected. This action relates to Training data provenance. Data provenance establishes what entered training and under what conditions. For the control decision, it does not resolve the specifically described issue of Interpretability testing in An insurance model provides a feature attribution chart but explanations change wildly across near-identical cases. Model assurance reviewers need a documented control before approving continued experimentation.
Question 743
An insurance model provides a feature attribution chart but explanations change wildly across near-identical cases. Model assurance reviewers need a documented control before approving continued experimentation. What result would provide the strongest case-specific assurance?
- Test for leakage indicators and compare results on a genuinely unseen time period.
- Re-run an approved benchmark and compare results within documented tolerances.
- Inspect collection origin, licenses, consent where applicable, lineage and transformation history.
- Test similar inputs for explanation consistency and verify reviewers do not treat correlates as proven causes.
Explanation
Correct Answer
D:Test similar inputs for explanation consistency and verify reviewers do not treat correlates as proven causes. Correct. It yields measurable evidence that the case-specific safeguard works. An explanation tool may be misleading unless its reliability and limitations are evaluated.
Incorrect Answers
A:Test for leakage indicators and compare results on a genuinely unseen time period. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Dataset leakage and temporal splits: A contaminated test set can give convincing but invalid evidence of generalization. The present case instead concerns Interpretability testing.
B:Re-run an approved benchmark and compare results within documented tolerances. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Testing evidence and reproducibility: Reproducibility makes AI testing evidence reviewable rather than anecdotal. The present case instead concerns Interpretability testing.
C:Inspect collection origin, licenses, consent where applicable, lineage and transformation history. Not selected. This action relates to Training data provenance. Data provenance establishes what entered training and under what conditions. For the validate decision, it does not resolve the specifically described issue of Interpretability testing in An insurance model provides a feature attribution chart but explanations change wildly across near-identical cases. Model assurance reviewers need a documented control before approving continued experimentation.
Question 744
An insurance model provides a feature attribution chart but explanations change wildly across near-identical cases. Model assurance reviewers need a documented control before approving continued experimentation. Which TWO actions form the appropriate investigation-and-control pair for this case?
- Validate explanation methods and document what attributions do and do not establish.
- Identify how sample selection, labels and historical practices may propagate protected-group disadvantages.
- Assess explanation stability, fidelity and whether the method matches model and use-case limitations.
- Use temporal or group-aware splits and isolate evaluation records from training and tuning.
Explanation
Correct Answers
A:Validate explanation methods and document what attributions do and do not establish. Correct. As the second, implementation half of the requested pair, this directly applies the needed safeguard. An explanation tool may be misleading unless its reliability and limitations are evaluated.
C:Assess explanation stability, fidelity and whether the method matches model and use-case limitations. Correct. As the first, evidence-gathering half of the requested pair, this establishes the specific risk boundary. An explanation tool may be misleading unless its reliability and limitations are evaluated.
Incorrect Answers
B:Identify how sample selection, labels and historical practices may propagate protected-group disadvantages. Not selected. It addresses a different governance mechanism, not the required assessment-and-control pair. It belongs with Bias measurement and mitigation: Historic training labels can reproduce inequity even when protected fields are omitted. The present case instead concerns Interpretability testing.
D:Use temporal or group-aware splits and isolate evaluation records from training and tuning. Not selected. It addresses a different governance mechanism, not the required assessment-and-control pair. It belongs with Dataset leakage and temporal splits: A contaminated test set can give convincing but invalid evidence of generalization. The present case instead concerns Interpretability testing.
Question 745
A research group reports performance but cannot recreate the model because seeds, data versions and tuning settings were discarded. The development team must show evidence before promoting the next model version. What review step identifies the specific control gap rather than an unrelated AI risk?
- Assess explanation stability, fidelity and whether the method matches model and use-case limitations.
- Identify how sample selection, labels and historical practices may propagate protected-group disadvantages.
- Measure sample coverage, label consistency, subgroup representation and production-population mismatch.
- Identify missing code, data, parameter, environment and test-set versions.
Explanation
Correct Answer
D:Identify missing code, data, parameter, environment and test-set versions. Correct. It establishes the relevant factual risk boundary before implementation. Reproducibility makes AI testing evidence reviewable rather than anecdotal.
Incorrect Answers
A:Assess explanation stability, fidelity and whether the method matches model and use-case limitations. Not selected. That check primarily diagnoses a different governance issue. It belongs with Interpretability testing: An explanation tool may be misleading unless its reliability and limitations are evaluated. The present case instead concerns Testing evidence and reproducibility.
B:Identify how sample selection, labels and historical practices may propagate protected-group disadvantages. Not selected. That check primarily diagnoses a different governance issue. It belongs with Bias measurement and mitigation: Historic training labels can reproduce inequity even when protected fields are omitted. The present case instead concerns Testing evidence and reproducibility.
C:Measure sample coverage, label consistency, subgroup representation and production-population mismatch. Not selected. That check primarily diagnoses a different governance issue. It belongs with Data quality and representativeness: Training data quality includes representativeness for the intended use population. The present case instead concerns Testing evidence and reproducibility.
Question 746
A research group reports performance but cannot recreate the model because seeds, data versions and tuning settings were discarded. The development team must show evidence before promoting the next model version. Which safeguard most directly treats the described failure mode?
- Maintain reproducible experiments and protected evaluation artifacts with traceable approvals.
- Validate explanation methods and document what attributions do and do not establish.
- Add robustness tests, input controls and fallback behavior for unsafe predictions.
- Improve approved data coverage and labeling with a targeted, lawful sampling plan.
Explanation
Correct Answer
A:Maintain reproducible experiments and protected evaluation artifacts with traceable approvals. Correct. It directly reduces or governs the specific documented risk. Reproducibility makes AI testing evidence reviewable rather than anecdotal.
Incorrect Answers
B:Validate explanation methods and document what attributions do and do not establish. Not selected. That action treats another control objective, not the specific gap described. It belongs with Interpretability testing: An explanation tool may be misleading unless its reliability and limitations are evaluated. The present case instead concerns Testing evidence and reproducibility.
C:Add robustness tests, input controls and fallback behavior for unsafe predictions. Not selected. That action treats another control objective, not the specific gap described. It belongs with Adversarial robustness testing: Security and robustness tests must cover plausible operating and adversarial conditions. The present case instead concerns Testing evidence and reproducibility.
D:Improve approved data coverage and labeling with a targeted, lawful sampling plan. Not selected. That action treats another control objective, not the specific gap described. It belongs with Data quality and representativeness: Training data quality includes representativeness for the intended use population. The present case instead concerns Testing evidence and reproducibility.
Question 747
A research group reports performance but cannot recreate the model because seeds, data versions and tuning settings were discarded. The development team must show evidence before promoting the next model version. Which verification task should support the next approval decision?
- Measure hazard-specific failure rates under glare, obstruction and attack conditions.
- Test similar inputs for explanation consistency and verify reviewers do not treat correlates as proven causes.
- Measure agreement and error on a reviewed gold-standard sample before training.
- Re-run an approved benchmark and compare results within documented tolerances.
Explanation
Correct Answer
D:Re-run an approved benchmark and compare results within documented tolerances. Correct. It yields measurable evidence that the case-specific safeguard works. Reproducibility makes AI testing evidence reviewable rather than anecdotal.
Incorrect Answers
A:Measure hazard-specific failure rates under glare, obstruction and attack conditions. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Adversarial robustness testing: Security and robustness tests must cover plausible operating and adversarial conditions. The present case instead concerns Testing evidence and reproducibility.
B:Test similar inputs for explanation consistency and verify reviewers do not treat correlates as proven causes. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Interpretability testing: An explanation tool may be misleading unless its reliability and limitations are evaluated. The present case instead concerns Testing evidence and reproducibility.
C:Measure agreement and error on a reviewed gold-standard sample before training. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Label quality and inter-rater agreement: Ground-truth quality shapes model behavior and can hide systematic harm. The present case instead concerns Testing evidence and reproducibility.
Question 748
A research group reports performance but cannot recreate the model because seeds, data versions and tuning settings were discarded. The development team must show evidence before promoting the next model version. If the control fails, which immediate response most appropriately limits the documented risk?
- Suspend adverse automated decisions when discriminatory outcomes cannot be justified or corrected.
- Remove problematic derived features and invalidate affected model releases when lineage cannot be verified.
- Discard contaminated benchmark claims and rerun controlled evaluation before release.
- Withdraw unverifiable performance claims and rerun controlled tests before launch.
Explanation
Correct Answer
D:Withdraw unverifiable performance claims and rerun controlled tests before launch. Correct. It limits further reliance or exposure while the risk remains unresolved. Reproducibility makes AI testing evidence reviewable rather than anecdotal.
Incorrect Answers
A:Suspend adverse automated decisions when discriminatory outcomes cannot be justified or corrected. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Bias measurement and mitigation: Historic training labels can reproduce inequity even when protected fields are omitted. The present case instead concerns Testing evidence and reproducibility.
B:Remove problematic derived features and invalidate affected model releases when lineage cannot be verified. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Data lineage for transformations: Lineage tracks how raw data becomes model-ready features, including indirect sensitive signals. The present case instead concerns Testing evidence and reproducibility.
C:Discard contaminated benchmark claims and rerun controlled evaluation before release. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Dataset leakage and temporal splits: A contaminated test set can give convincing but invalid evidence of generalization. The present case instead concerns Testing evidence and reproducibility.
Question 749
A research group reports performance but cannot recreate the model because seeds, data versions and tuning settings were discarded. Model assurance reviewers need a documented control before approving continued experimentation. Which initial evidence check most directly addresses the governance concern?
- Measure sample coverage, label consistency, subgroup representation and production-population mismatch.
- Check entity, time and duplicate overlap that may leak future information or memorized records.
- Identify missing code, data, parameter, environment and test-set versions.
- Accept only traceable and lawfully usable data with documented source and processing stages.
Explanation
Correct Answer
C:Identify missing code, data, parameter, environment and test-set versions. Correct. It establishes the relevant factual risk boundary before implementation. Reproducibility makes AI testing evidence reviewable rather than anecdotal.
Incorrect Answers
A:Measure sample coverage, label consistency, subgroup representation and production-population mismatch. Not selected. That check primarily diagnoses a different governance issue. It belongs with Data quality and representativeness: Training data quality includes representativeness for the intended use population. The present case instead concerns Testing evidence and reproducibility.
B:Check entity, time and duplicate overlap that may leak future information or memorized records. Not selected. That check primarily diagnoses a different governance issue. It belongs with Dataset leakage and temporal splits: A contaminated test set can give convincing but invalid evidence of generalization. The present case instead concerns Testing evidence and reproducibility.
D:Accept only traceable and lawfully usable data with documented source and processing stages. Not selected. This action relates to Training data provenance. Data provenance establishes what entered training and under what conditions. For the diagnose decision, it does not resolve the specifically described issue of Testing evidence and reproducibility in A research group reports performance but cannot recreate the model because seeds, data versions and tuning settings were discarded. Model assurance reviewers need a documented control before approving continued experimentation.
Question 750
A research group reports performance but cannot recreate the model because seeds, data versions and tuning settings were discarded. Model assurance reviewers need a documented control before approving continued experimentation. Which specific implementation best addresses the identified gap?
- Accept only traceable and lawfully usable data with documented source and processing stages.
- Record end-to-end feature lineage and review proxy risks in engineered attributes.
- Maintain reproducible experiments and protected evaluation artifacts with traceable approvals.
- Improve approved data coverage and labeling with a targeted, lawful sampling plan.
Explanation
Correct Answer
C:Maintain reproducible experiments and protected evaluation artifacts with traceable approvals. Correct. It directly reduces or governs the specific documented risk. Reproducibility makes AI testing evidence reviewable rather than anecdotal.
Incorrect Answers
A:Accept only traceable and lawfully usable data with documented source and processing stages. Not selected. That action treats another control objective, not the specific gap described. It belongs with Training data provenance: Data provenance establishes what entered training and under what conditions. The present case instead concerns Testing evidence and reproducibility.
B:Record end-to-end feature lineage and review proxy risks in engineered attributes. Not selected. That action treats another control objective, not the specific gap described. It belongs with Data lineage for transformations: Lineage tracks how raw data becomes model-ready features, including indirect sensitive signals. The present case instead concerns Testing evidence and reproducibility.
D:Improve approved data coverage and labeling with a targeted, lawful sampling plan. Not selected. That action treats another control objective, not the specific gap described. It belongs with Data quality and representativeness: Training data quality includes representativeness for the intended use population. The present case instead concerns Testing evidence and reproducibility.