IAPP AIGP Practice Test 70
Question 690
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. The development team must show evidence before promoting the next model version. What is the most directly applicable governance control for this situation?
- Improve approved data coverage and labeling with a targeted, lawful sampling plan.
- Accept only traceable and lawfully usable data with documented source and processing stages.
- Define and refine labeling policy with expert adjudication and representative agreement testing.
- Add robustness tests, input controls and fallback behavior for unsafe predictions.
Explanation
Correct Answer
C:Define and refine labeling policy with expert adjudication and representative agreement testing. Correct. It directly reduces or governs the specific documented risk. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
A: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 Label quality and inter-rater agreement.
B: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 Label quality and inter-rater agreement.
D: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 Label quality and inter-rater agreement.
Question 691
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. The development team must show evidence before promoting the next model version. What should an independent reviewer test before signing off on this specific risk?
- Recreate selected features from source data and confirm sensitive transformations are visible.
- Evaluate both error differences and selection effects on representative groups after mitigation.
- 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.
Explanation
Correct Answer
D:Measure agreement and error on a reviewed gold-standard sample before training. Correct. It yields measurable evidence that the case-specific safeguard works. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
A:Recreate selected features from source data and confirm sensitive transformations are visible. Not selected. That test would not demonstrate this particular risk has been addressed. 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 Label quality and inter-rater agreement.
B:Evaluate both error differences and selection effects on representative groups after mitigation. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Bias measurement and mitigation: Historic training labels can reproduce inequity even when protected fields are omitted. The present case instead concerns Label quality and inter-rater agreement.
C: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 Label quality and inter-rater agreement.
Question 692
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. The development team must show evidence before promoting the next model version. Select TWO steps that together establish the issue and directly address it.
- Audit labeling guidance, disagreement patterns, annotator context and ambiguous categories.
- Add robustness tests, input controls and fallback behavior for unsafe predictions.
- Identify how sample selection, labels and historical practices may propagate protected-group disadvantages.
- Define and refine labeling policy with expert adjudication and representative agreement testing.
Explanation
Correct Answers
A:Audit labeling guidance, disagreement patterns, annotator context and ambiguous categories. Correct. As the first, evidence-gathering half of the requested pair, this establishes the specific risk boundary. Ground-truth quality shapes model behavior and can hide systematic harm.
D:Define and refine labeling policy with expert adjudication and representative agreement testing. Correct. As the second, implementation half of the requested pair, this directly applies the needed safeguard. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
B:Add robustness tests, input controls and fallback behavior for unsafe predictions. Not selected. It addresses a different governance mechanism, not the required assessment-and-control pair. It belongs with Adversarial robustness testing: Security and robustness tests must cover plausible operating and adversarial conditions. The present case instead concerns Label quality and inter-rater agreement.
C: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 Label quality and inter-rater agreement.
Question 693
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. Model assurance reviewers need a documented control before approving continued experimentation. What review step identifies the specific control gap rather than an unrelated AI risk?
- Check entity, time and duplicate overlap that may leak future information or memorized records.
- Audit labeling guidance, disagreement patterns, annotator context and ambiguous categories.
- Inspect collection origin, licenses, consent where applicable, lineage and transformation history.
- Measure sample coverage, label consistency, subgroup representation and production-population mismatch.
Explanation
Correct Answer
B:Audit labeling guidance, disagreement patterns, annotator context and ambiguous categories. Correct. It establishes the relevant factual risk boundary before implementation. Ground-truth quality shapes model behavior and can hide systematic harm.
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 Label quality and inter-rater agreement.
C:Inspect collection origin, licenses, consent where applicable, lineage and transformation history. Not selected. That check primarily diagnoses a different governance issue. It belongs with Training data provenance: Data provenance establishes what entered training and under what conditions. The present case instead concerns Label quality and inter-rater agreement.
D: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 Label quality and inter-rater agreement.
Question 694
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. Model assurance reviewers need a documented control before approving continued experimentation. Which safeguard most directly treats the described failure mode?
- Use temporal or group-aware splits and isolate evaluation records from training and tuning.
- Define and refine labeling policy with expert adjudication and representative agreement testing.
- Maintain reproducible experiments and protected evaluation artifacts with traceable approvals.
- Accept only traceable and lawfully usable data with documented source and processing stages.
Explanation
Correct Answer
B:Define and refine labeling policy with expert adjudication and representative agreement testing. Correct. It directly reduces or governs the specific documented risk. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
A:Use temporal or group-aware splits and isolate evaluation records from training and tuning. Not selected. That action treats another control objective, not the specific gap described. 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 Label quality and inter-rater agreement.
C:Maintain reproducible experiments and protected evaluation artifacts with traceable approvals. Not selected. That action treats another control objective, not the specific gap described. It belongs with Testing evidence and reproducibility: Reproducibility makes AI testing evidence reviewable rather than anecdotal. The present case instead concerns Label quality and inter-rater agreement.
D: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 Label quality and inter-rater agreement.
Question 695
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. Model assurance reviewers need a documented control before approving continued experimentation. Which verification task should support the next approval decision?
- Run a held-out dialect evaluation and compare false-negative rates and uncertainty thresholds.
- Recreate selected features from source data and confirm sensitive transformations are visible.
- Measure agreement and error on a reviewed gold-standard sample before training.
- Test similar inputs for explanation consistency and verify reviewers do not treat correlates as proven causes.
Explanation
Correct Answer
C:Measure agreement and error on a reviewed gold-standard sample before training. Correct. It yields measurable evidence that the case-specific safeguard works. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
A:Run a held-out dialect evaluation and compare false-negative rates and uncertainty thresholds. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Data quality and representativeness: Training data quality includes representativeness for the intended use population. The present case instead concerns Label quality and inter-rater agreement.
B:Recreate selected features from source data and confirm sensitive transformations are visible. Not selected. That test would not demonstrate this particular risk has been addressed. 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 Label quality and inter-rater agreement.
D: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 Label quality and inter-rater agreement.
Question 696
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. Model assurance reviewers need a documented control before approving continued experimentation. If the control fails, which immediate response most appropriately limits the documented risk?
- Remove problematic derived features and invalidate affected model releases when lineage cannot be verified.
- Switch to a validated safe mode when input quality falls outside the assessed operating domain.
- Provide accessible non-AI service paths while underrepresented speech performance is corrected.
- Pause use of unreliable labels and reannotate ambiguous categories under an approved process.
Explanation
Correct Answer
D:Pause use of unreliable labels and reannotate ambiguous categories under an approved process. Correct. It limits further reliance or exposure while the risk remains unresolved. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
A: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 Label quality and inter-rater agreement.
B: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 Label quality and inter-rater agreement.
C:Provide accessible non-AI service paths while underrepresented speech performance is corrected. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Data quality and representativeness: Training data quality includes representativeness for the intended use population. The present case instead concerns Label quality and inter-rater agreement.
Question 697
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. A material training or architecture change makes the previous technical approval insufficient. Which initial evidence check most directly addresses the governance concern?
- Inspect collection origin, licenses, consent where applicable, lineage and transformation history.
- Select foreseeable corruption, adversarial perturbation and distribution-shift scenarios to test.
- Audit labeling guidance, disagreement patterns, annotator context and ambiguous categories.
- Measure sample coverage, label consistency, subgroup representation and production-population mismatch.
Explanation
Correct Answer
C:Audit labeling guidance, disagreement patterns, annotator context and ambiguous categories. Correct. It establishes the relevant factual risk boundary before implementation. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
A:Inspect collection origin, licenses, consent where applicable, lineage and transformation history. Not selected. That check primarily diagnoses a different governance issue. It belongs with Training data provenance: Data provenance establishes what entered training and under what conditions. The present case instead concerns Label quality and inter-rater agreement.
B:Select foreseeable corruption, adversarial perturbation and distribution-shift scenarios to test. Not selected. That check primarily diagnoses a different governance issue. It belongs with Adversarial robustness testing: Security and robustness tests must cover plausible operating and adversarial conditions. The present case instead concerns Label quality and inter-rater agreement.
D: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 Label quality and inter-rater agreement.
Question 698
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. A material training or architecture change makes the previous technical approval insufficient. Which specific implementation best addresses the identified gap?
- Define and refine labeling policy with expert adjudication and representative agreement testing.
- Record end-to-end feature lineage and review proxy risks in engineered attributes.
- Improve approved data coverage and labeling with a targeted, lawful sampling plan.
- Maintain reproducible experiments and protected evaluation artifacts with traceable approvals.
Explanation
Correct Answer
A:Define and refine labeling policy with expert adjudication and representative agreement testing. Correct. It directly reduces or governs the specific documented risk. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
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 Label quality and inter-rater agreement.
C: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 Label quality and inter-rater agreement.
D:Maintain reproducible experiments and protected evaluation artifacts with traceable approvals. Not selected. That action treats another control objective, not the specific gap described. It belongs with Testing evidence and reproducibility: Reproducibility makes AI testing evidence reviewable rather than anecdotal. The present case instead concerns Label quality and inter-rater agreement.
Question 699
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. A material training or architecture change makes the previous technical approval insufficient. Which acceptance test most directly verifies the relevant control?
- Measure agreement and error on a reviewed gold-standard sample before training.
- Test for leakage indicators and compare results on a genuinely unseen time period.
- Recreate selected features from source data and confirm sensitive transformations are visible.
- Test similar inputs for explanation consistency and verify reviewers do not treat correlates as proven causes.
Explanation
Correct Answer
A:Measure agreement and error on a reviewed gold-standard sample before training. Correct. It yields measurable evidence that the case-specific safeguard works. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
B: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 Label quality and inter-rater agreement.
C:Recreate selected features from source data and confirm sensitive transformations are visible. Not selected. That test would not demonstrate this particular risk has been addressed. 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 Label quality and inter-rater agreement.
D: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 Label quality and inter-rater agreement.
Question 700
An image-safety classifier uses inconsistent human labels for borderline images with no adjudication instructions. A material training or architecture change makes the previous technical approval insufficient. Which controlled response should follow an unsuccessful assurance test?
- Provide accessible non-AI service paths while underrepresented speech performance is corrected.
- Pause use of unreliable labels and reannotate ambiguous categories under an approved process.
- Withdraw unverifiable performance claims and rerun controlled tests before launch.
- Remove problematic derived features and invalidate affected model releases when lineage cannot be verified.
Explanation
Correct Answer
B:Pause use of unreliable labels and reannotate ambiguous categories under an approved process. Correct. It limits further reliance or exposure while the risk remains unresolved. Ground-truth quality shapes model behavior and can hide systematic harm.
Incorrect Answers
A:Provide accessible non-AI service paths while underrepresented speech performance is corrected. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Data quality and representativeness: Training data quality includes representativeness for the intended use population. The present case instead concerns Label quality and inter-rater agreement.
C: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 Label quality and inter-rater agreement.
D: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 Label quality and inter-rater agreement.