IAPP AIGP Practice Test 2
Question 10
A benefits agency calls a rules engine and a self-learning applicant-ranking model the same thing, so reviewers cannot tell which decisions are probabilistic. A material change to the operating model has made the earlier risk acceptance incomplete. What action should the accountable team prioritize to govern this use responsibly?
- Restrict tools by least privilege, define human approval boundaries and cap autonomous execution.
- Document the model, its inputs, predictions, decision points and accountable humans in the system inventory.
- Set confidence thresholds and an escalation pathway for uncertainty, with clinician review for consequential cases.
- Adopt suitable local explanations, decision records and a human challenge process without asserting that every feature is causal.
Explanation
Correct Answer
B:Document the model, its inputs, predictions, decision points and accountable humans in the system inventory. Correct. It directly reduces or governs the specific documented risk. AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions.
Incorrect Answers
A:Restrict tools by least privilege, define human approval boundaries and cap autonomous execution. Not selected. That action treats another control objective, not the specific gap described. It belongs with Autonomy and agentic scope: Agentic AI requires controls over actions and delegated authority as well as generated text. The present case instead concerns AI system boundary.
C:Set confidence thresholds and an escalation pathway for uncertainty, with clinician review for consequential cases. Not selected. That action treats another control objective, not the specific gap described. It belongs with Probabilistic uncertainty: A probabilistic output is not a deterministic guarantee, especially in high-impact applications. The present case instead concerns AI system boundary.
D:Adopt suitable local explanations, decision records and a human challenge process without asserting that every feature is causal. Not selected. That action treats another control objective, not the specific gap described. It belongs with AI opacity and explanation: Transparency and useful explainability must match the affected audience and impact, not simply display a score. The present case instead concerns AI system boundary.
Question 11
A benefits agency calls a rules engine and a self-learning applicant-ranking model the same thing, so reviewers cannot tell which decisions are probabilistic. A material change to the operating model has made the earlier risk acceptance incomplete. What result would provide the strongest case-specific assurance?
- Measure calibration and false-negative performance by subgroup on representative evaluation data.
- Trace test cases from model output to the human decision, including cases in which the output is ignored.
- Attempt unauthorized and irreversible test actions and confirm the agent stops or requests approval.
- Test whether affected users and reviewers can understand and contest representative decisions.
Explanation
Correct Answer
B:Trace test cases from model output to the human decision, including cases in which the output is ignored. Correct. It yields measurable evidence that the case-specific safeguard works. AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions.
Incorrect Answers
A:Measure calibration and false-negative performance by subgroup on representative evaluation data. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Probabilistic uncertainty: A probabilistic output is not a deterministic guarantee, especially in high-impact applications. The present case instead concerns AI system boundary.
C:Attempt unauthorized and irreversible test actions and confirm the agent stops or requests approval. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Autonomy and agentic scope: Agentic AI requires controls over actions and delegated authority as well as generated text. The present case instead concerns AI system boundary.
D:Test whether affected users and reviewers can understand and contest representative decisions. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with AI opacity and explanation: Transparency and useful explainability must match the affected audience and impact, not simply display a score. The present case instead concerns AI system boundary.
Question 12
A benefits agency calls a rules engine and a self-learning applicant-ranking model the same thing, so reviewers cannot tell which decisions are probabilistic. A material change to the operating model has made the earlier risk acceptance incomplete. Which TWO actions form the appropriate investigation-and-control pair for this case?
- Identify which component infers patterns from data and where outputs affect applicants, rather than treating every rule as machine learning.
- Document the model, its inputs, predictions, decision points and accountable humans in the system inventory.
- Estimate aggregate impact using exposure volume, severity and unequal subgroup effects.
- Restrict tools by least privilege, define human approval boundaries and cap autonomous execution.
Explanation
Correct Answers
A:Identify which component infers patterns from data and where outputs affect applicants, rather than treating every rule as machine learning. Correct. As the first, evidence-gathering half of the requested pair, this establishes the specific risk boundary. AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions.
B:Document the model, its inputs, predictions, decision points and accountable humans in the system inventory. Correct. As the second, implementation half of the requested pair, this directly applies the needed safeguard. AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions.
Incorrect Answers
C:Estimate aggregate impact using exposure volume, severity and unequal subgroup effects. Not selected. It addresses a different governance mechanism, not the required assessment-and-control pair. It belongs with Scale-driven harm: Even a low individual error rate may create material harm when AI is applied at scale. The present case instead concerns AI system boundary.
D:Restrict tools by least privilege, define human approval boundaries and cap autonomous execution. Not selected. It addresses a different governance mechanism, not the required assessment-and-control pair. It belongs with Autonomy and agentic scope: Agentic AI requires controls over actions and delegated authority as well as generated text. The present case instead concerns AI system boundary.
Question 13
A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. The system owner is requesting a documented approval of this existing governance concern. Before selecting a remedy, what should the accountable reviewer investigate first?
- Identify the competing fairness, privacy, utility and lawful-processing objectives in the specific use case.
- Inspect calibration curves and subgroup error rates, not just the displayed confidence percentage.
- Identify whether explanations describe case-specific factors, system limits and meaningful review routes.
- Document the model, its inputs, predictions, decision points and accountable humans in the system inventory.
Explanation
Correct Answer
B:Inspect calibration curves and subgroup error rates, not just the displayed confidence percentage. Correct. It establishes the relevant factual risk boundary before implementation. A probabilistic output is not a deterministic guarantee, especially in high-impact applications.
Incorrect Answers
A:Identify the competing fairness, privacy, utility and lawful-processing objectives in the specific use case. Not selected. That check primarily diagnoses a different governance issue. It belongs with Responsible AI tradeoffs: Responsible AI principles interact; governance should record and justify risk tradeoffs rather than invoke a single slogan. The present case instead concerns Probabilistic uncertainty.
C:Identify whether explanations describe case-specific factors, system limits and meaningful review routes. Not selected. That check primarily diagnoses a different governance issue. It belongs with AI opacity and explanation: Transparency and useful explainability must match the affected audience and impact, not simply display a score. The present case instead concerns Probabilistic uncertainty.
D:Document the model, its inputs, predictions, decision points and accountable humans in the system inventory. Not selected. This action relates to AI system boundary. AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions. For the diagnose decision, it does not resolve the specifically described issue of Probabilistic uncertainty in A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. The system owner is requesting a documented approval of this existing governance concern.
Question 14
A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. The system owner is requesting a documented approval of this existing governance concern. What action should the accountable team prioritize to govern this use responsibly?
- Reassess risk classification, capacity for appeals and monitoring thresholds before scaling.
- Set confidence thresholds and an escalation pathway for uncertainty, with clinician review for consequential cases.
- Restrict tools by least privilege, define human approval boundaries and cap autonomous execution.
- Adopt suitable local explanations, decision records and a human challenge process without asserting that every feature is causal.
Explanation
Correct Answer
B:Set confidence thresholds and an escalation pathway for uncertainty, with clinician review for consequential cases. Correct. It directly reduces or governs the specific documented risk. A probabilistic output is not a deterministic guarantee, especially in high-impact applications.
Incorrect Answers
A:Reassess risk classification, capacity for appeals and monitoring thresholds before scaling. Not selected. That action treats another control objective, not the specific gap described. It belongs with Scale-driven harm: Even a low individual error rate may create material harm when AI is applied at scale. The present case instead concerns Probabilistic uncertainty.
C:Restrict tools by least privilege, define human approval boundaries and cap autonomous execution. Not selected. That action treats another control objective, not the specific gap described. It belongs with Autonomy and agentic scope: Agentic AI requires controls over actions and delegated authority as well as generated text. The present case instead concerns Probabilistic uncertainty.
D:Adopt suitable local explanations, decision records and a human challenge process without asserting that every feature is causal. Not selected. That action treats another control objective, not the specific gap described. It belongs with AI opacity and explanation: Transparency and useful explainability must match the affected audience and impact, not simply display a score. The present case instead concerns Probabilistic uncertainty.
Question 15
A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. The system owner is requesting a documented approval of this existing governance concern. What result would provide the strongest case-specific assurance?
- Trace test cases from model output to the human decision, including cases in which the output is ignored.
- Test whether affected users and reviewers can understand and contest representative decisions.
- Identify which component infers patterns from data and where outputs affect applicants, rather than treating every rule as machine learning.
- Measure calibration and false-negative performance by subgroup on representative evaluation data.
Explanation
Correct Answer
D:Measure calibration and false-negative performance by subgroup on representative evaluation data. Correct. It yields measurable evidence that the case-specific safeguard works. A probabilistic output is not a deterministic guarantee, especially in high-impact applications.
Incorrect Answers
A:Trace test cases from model output to the human decision, including cases in which the output is ignored. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with AI system boundary: AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions. The present case instead concerns Probabilistic uncertainty.
B:Test whether affected users and reviewers can understand and contest representative decisions. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with AI opacity and explanation: Transparency and useful explainability must match the affected audience and impact, not simply display a score. The present case instead concerns Probabilistic uncertainty.
C:Identify which component infers patterns from data and where outputs affect applicants, rather than treating every rule as machine learning. Not selected. This action relates to AI system boundary. AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions. For the validate decision, it does not resolve the specifically described issue of Probabilistic uncertainty in A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. The system owner is requesting a documented approval of this existing governance concern.
Question 16
A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. The system owner is requesting a documented approval of this existing governance concern. What should the owner do to prevent further harm from this specific failure?
- Disable high-impact tool actions and revoke credentials if the agent exceeds its approved autonomy.
- Route unreliable classifications to qualified human assessment while recalibrating and validating the model.
- Throttle rollout and provide manual review where accumulated harm exceeds tolerance.
- Stop the unapproved collection and reconsider privacy-preserving evaluation methods if proportionality cannot be shown.
Explanation
Correct Answer
B:Route unreliable classifications to qualified human assessment while recalibrating and validating the model. Correct. It limits further reliance or exposure while the risk remains unresolved. A probabilistic output is not a deterministic guarantee, especially in high-impact applications.
Incorrect Answers
A:Disable high-impact tool actions and revoke credentials if the agent exceeds its approved autonomy. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Autonomy and agentic scope: Agentic AI requires controls over actions and delegated authority as well as generated text. The present case instead concerns Probabilistic uncertainty.
C:Throttle rollout and provide manual review where accumulated harm exceeds tolerance. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Scale-driven harm: Even a low individual error rate may create material harm when AI is applied at scale. The present case instead concerns Probabilistic uncertainty.
D:Stop the unapproved collection and reconsider privacy-preserving evaluation methods if proportionality cannot be shown. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Responsible AI tradeoffs: Responsible AI principles interact; governance should record and justify risk tradeoffs rather than invoke a single slogan. The present case instead concerns Probabilistic uncertainty.
Question 17
A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. A separate assurance review finds the concern is relevant to an expanded user group. What should the AI governance lead examine before changing the system?
- Inspect calibration curves and subgroup error rates, not just the displayed confidence percentage.
- Identify whether explanations describe case-specific factors, system limits and meaningful review routes.
- Identify which component infers patterns from data and where outputs affect applicants, rather than treating every rule as machine learning.
- Document the model, its inputs, predictions, decision points and accountable humans in the system inventory.
Explanation
Correct Answer
A:Inspect calibration curves and subgroup error rates, not just the displayed confidence percentage. Correct. It establishes the relevant factual risk boundary before implementation. A probabilistic output is not a deterministic guarantee, especially in high-impact applications.
Incorrect Answers
B:Identify whether explanations describe case-specific factors, system limits and meaningful review routes. Not selected. That check primarily diagnoses a different governance issue. It belongs with AI opacity and explanation: Transparency and useful explainability must match the affected audience and impact, not simply display a score. The present case instead concerns Probabilistic uncertainty.
C:Identify which component infers patterns from data and where outputs affect applicants, rather than treating every rule as machine learning. Not selected. That check primarily diagnoses a different governance issue. It belongs with AI system boundary: AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions. The present case instead concerns Probabilistic uncertainty.
D:Document the model, its inputs, predictions, decision points and accountable humans in the system inventory. Not selected. This action relates to AI system boundary. AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions. For the diagnose decision, it does not resolve the specifically described issue of Probabilistic uncertainty in A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. A separate assurance review finds the concern is relevant to an expanded user group.
Question 18
A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. A separate assurance review finds the concern is relevant to an expanded user group. What is the most directly applicable governance control for this situation?
- Set confidence thresholds and an escalation pathway for uncertainty, with clinician review for consequential cases.
- Document the model, its inputs, predictions, decision points and accountable humans in the system inventory.
- Select a documented proportionate mitigation that tests fairness while minimizing sensitive-data access and securing any justified processing.
- Reassess risk classification, capacity for appeals and monitoring thresholds before scaling.
Explanation
Correct Answer
A:Set confidence thresholds and an escalation pathway for uncertainty, with clinician review for consequential cases. Correct. It directly reduces or governs the specific documented risk. A probabilistic output is not a deterministic guarantee, especially in high-impact applications.
Incorrect Answers
B:Document the model, its inputs, predictions, decision points and accountable humans in the system inventory. Not selected. That action treats another control objective, not the specific gap described. It belongs with AI system boundary: AI governance starts with a precise inventory of what the AI system does and how its outputs affect decisions. The present case instead concerns Probabilistic uncertainty.
C:Select a documented proportionate mitigation that tests fairness while minimizing sensitive-data access and securing any justified processing. Not selected. That action treats another control objective, not the specific gap described. It belongs with Responsible AI tradeoffs: Responsible AI principles interact; governance should record and justify risk tradeoffs rather than invoke a single slogan. The present case instead concerns Probabilistic uncertainty.
D:Reassess risk classification, capacity for appeals and monitoring thresholds before scaling. Not selected. That action treats another control objective, not the specific gap described. It belongs with Scale-driven harm: Even a low individual error rate may create material harm when AI is applied at scale. The present case instead concerns Probabilistic uncertainty.
Question 19
A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. A separate assurance review finds the concern is relevant to an expanded user group. What should an independent reviewer test before signing off on this specific risk?
- Measure calibration and false-negative performance by subgroup on representative evaluation data.
- Verify the fairness test improves relevant subgroup outcomes without exceeding approved data purposes.
- Test whether affected users and reviewers can understand and contest representative decisions.
- Attempt unauthorized and irreversible test actions and confirm the agent stops or requests approval.
Explanation
Correct Answer
A:Measure calibration and false-negative performance by subgroup on representative evaluation data. Correct. It yields measurable evidence that the case-specific safeguard works. A probabilistic output is not a deterministic guarantee, especially in high-impact applications.
Incorrect Answers
B:Verify the fairness test improves relevant subgroup outcomes without exceeding approved data purposes. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Responsible AI tradeoffs: Responsible AI principles interact; governance should record and justify risk tradeoffs rather than invoke a single slogan. The present case instead concerns Probabilistic uncertainty.
C:Test whether affected users and reviewers can understand and contest representative decisions. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with AI opacity and explanation: Transparency and useful explainability must match the affected audience and impact, not simply display a score. The present case instead concerns Probabilistic uncertainty.
D:Attempt unauthorized and irreversible test actions and confirm the agent stops or requests approval. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Autonomy and agentic scope: Agentic AI requires controls over actions and delegated authority as well as generated text. The present case instead concerns Probabilistic uncertainty.
Question 20
A medical triage classifier returns confident-looking labels although its probability calibration differs sharply across patient groups. A separate assurance review finds the concern is relevant to an expanded user group. What is the most relevant fallback when the promised safeguard cannot be demonstrated?
- Route unreliable classifications to qualified human assessment while recalibrating and validating the model.
- Disable high-impact tool actions and revoke credentials if the agent exceeds its approved autonomy.
- Suspend unsupported automated adverse recommendations and route cases for explainable review.
- Stop the unapproved collection and reconsider privacy-preserving evaluation methods if proportionality cannot be shown.
Explanation
Correct Answer
A:Route unreliable classifications to qualified human assessment while recalibrating and validating the model. Correct. It limits further reliance or exposure while the risk remains unresolved. A probabilistic output is not a deterministic guarantee, especially in high-impact applications.
Incorrect Answers
B:Disable high-impact tool actions and revoke credentials if the agent exceeds its approved autonomy. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Autonomy and agentic scope: Agentic AI requires controls over actions and delegated authority as well as generated text. The present case instead concerns Probabilistic uncertainty.
C:Suspend unsupported automated adverse recommendations and route cases for explainable review. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with AI opacity and explanation: Transparency and useful explainability must match the affected audience and impact, not simply display a score. The present case instead concerns Probabilistic uncertainty.
D:Stop the unapproved collection and reconsider privacy-preserving evaluation methods if proportionality cannot be shown. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Responsible AI tradeoffs: Responsible AI principles interact; governance should record and justify risk tradeoffs rather than invoke a single slogan. The present case instead concerns Probabilistic uncertainty.