IAPP AIGP Practice Test 98
Question 970
A public agency proposes fine-tuning a language model on volatile benefit rules that change every month. A new business unit wants to scale the system without a further governance review. Which specific implementation best addresses the identified gap?
- Prefer controlled retrieval of versioned policy sources for rapidly changing factual rules; fine-tune only for justified behavior goals.
- Define a site-readiness gate with infrastructure, workforce and performance acceptance criteria.
- Use controlled retrieval from approved current documents with access filtering and source-aware answers.
- Assess business objectives, data availability, operator skills, network conditions and failure impact at each deployment site.
Explanation
Correct Answer
A:Prefer controlled retrieval of versioned policy sources for rapidly changing factual rules; fine-tune only for justified behavior goals. Correct. It directly reduces or governs the specific documented risk. Fine-tuning adapts model behavior but is not always suitable for rapidly changing facts.
Incorrect Answers
B:Define a site-readiness gate with infrastructure, workforce and performance acceptance criteria. Not selected. That action treats another control objective, not the specific gap described. It belongs with Deployment-context feasibility: The deployment environment and workforce can create new risk even when the model was sound in development. The present case instead concerns Fine-tuning versus prompt or RAG.
C:Use controlled retrieval from approved current documents with access filtering and source-aware answers. Not selected. That action treats another control objective, not the specific gap described. It belongs with Retrieval augmented generation: RAG can improve grounding but does not eliminate hallucination, leakage or access-control obligations. The present case instead concerns Fine-tuning versus prompt or RAG.
D:Assess business objectives, data availability, operator skills, network conditions and failure impact at each deployment site. Not selected. This action relates to Deployment-context feasibility. The deployment environment and workforce can create new risk even when the model was sound in development. For the control decision, it does not resolve the specifically described issue of Fine-tuning versus prompt or RAG in A public agency proposes fine-tuning a language model on volatile benefit rules that change every month. A new business unit wants to scale the system without a further governance review.
Question 971
A public agency proposes fine-tuning a language model on volatile benefit rules that change every month. A new business unit wants to scale the system without a further governance review. Which acceptance test most directly verifies the relevant control?
- Test version-changed policy questions and confirm citations, permissions and abstention behavior.
- Run field pilots at representative stores and review user comprehension and prediction failure rates.
- Change a policy rule and verify the system retrieves the latest approved version without retraining.
- Measure tail latency, operational energy use, robustness and error rates on representative meters.
Explanation
Correct Answer
C:Change a policy rule and verify the system retrieves the latest approved version without retraining. Correct. It yields measurable evidence that the case-specific safeguard works. Fine-tuning adapts model behavior but is not always suitable for rapidly changing facts.
Incorrect Answers
A:Test version-changed policy questions and confirm citations, permissions and abstention behavior. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Retrieval augmented generation: RAG can improve grounding but does not eliminate hallucination, leakage or access-control obligations. The present case instead concerns Fine-tuning versus prompt or RAG.
B:Run field pilots at representative stores and review user comprehension and prediction failure rates. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Deployment-context feasibility: The deployment environment and workforce can create new risk even when the model was sound in development. The present case instead concerns Fine-tuning versus prompt or RAG.
D:Measure tail latency, operational energy use, robustness and error rates on representative meters. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Small versus large model: Model size is one factor in suitability and can increase cost and exposure without adding needed value. The present case instead concerns Fine-tuning versus prompt or RAG.
Question 972
A public agency proposes fine-tuning a language model on volatile benefit rules that change every month. A new business unit wants to scale the system without a further governance review. For a defensible decision, choose TWO: one diagnosis and one controlled response.
- Compare fine-tuning, prompt design and retrieval on update frequency, provenance, cost and rights.
- Prefer controlled retrieval of versioned policy sources for rapidly changing factual rules; fine-tune only for justified behavior goals.
- Compare task quality, compute constraints, energy, privacy exposure and response time for appropriately sized models.
- Deploy an appropriately validated edge or on-prem design with bounded synchronization when cloud reliance is unsafe.
Explanation
Correct Answers
A:Compare fine-tuning, prompt design and retrieval on update frequency, provenance, cost and rights. Correct. As the first, evidence-gathering half of the requested pair, this establishes the specific risk boundary. Fine-tuning adapts model behavior but is not always suitable for rapidly changing facts.
B:Prefer controlled retrieval of versioned policy sources for rapidly changing factual rules; fine-tune only for justified behavior goals. Correct. As the second, implementation half of the requested pair, this directly applies the needed safeguard. Fine-tuning adapts model behavior but is not always suitable for rapidly changing facts.
Incorrect Answers
C:Compare task quality, compute constraints, energy, privacy exposure and response time for appropriately sized models. Not selected. It addresses a different governance mechanism, not the required assessment-and-control pair. It belongs with Small versus large model: Model size is one factor in suitability and can increase cost and exposure without adding needed value. The present case instead concerns Fine-tuning versus prompt or RAG.
D:Deploy an appropriately validated edge or on-prem design with bounded synchronization when cloud reliance is unsafe. Not selected. It addresses a different governance mechanism, not the required assessment-and-control pair. It belongs with Cloud edge on-prem choice: AI deployment topology should match operational hazards and data obligations. The present case instead concerns Fine-tuning versus prompt or RAG.
Question 973
A public agency proposes fine-tuning a language model on volatile benefit rules that change every month. Postdeployment review has identified the same issue after a material service change. Which diagnostic task establishes the risk relevant to this described system?
- Compare task quality, compute constraints, energy, privacy exposure and response time for appropriately sized models.
- Compare fine-tuning, prompt design and retrieval on update frequency, provenance, cost and rights.
- Compare cloud, on-premises and edge deployment against latency, resilience, data location and incident-recovery needs.
- Assess licensing, hosting control, security patching, model access, cost, supplier dependence and data terms.
Explanation
Correct Answer
B:Compare fine-tuning, prompt design and retrieval on update frequency, provenance, cost and rights. Correct. It establishes the relevant factual risk boundary before implementation. Fine-tuning adapts model behavior but is not always suitable for rapidly changing facts.
Incorrect Answers
A:Compare task quality, compute constraints, energy, privacy exposure and response time for appropriately sized models. Not selected. That check primarily diagnoses a different governance issue. It belongs with Small versus large model: Model size is one factor in suitability and can increase cost and exposure without adding needed value. The present case instead concerns Fine-tuning versus prompt or RAG.
C:Compare cloud, on-premises and edge deployment against latency, resilience, data location and incident-recovery needs. Not selected. That check primarily diagnoses a different governance issue. It belongs with Cloud edge on-prem choice: AI deployment topology should match operational hazards and data obligations. The present case instead concerns Fine-tuning versus prompt or RAG.
D:Assess licensing, hosting control, security patching, model access, cost, supplier dependence and data terms. Not selected. That check primarily diagnoses a different governance issue. It belongs with Open versus proprietary model: Open or proprietary licensing changes governance tradeoffs but neither eliminates accountability. The present case instead concerns Fine-tuning versus prompt or RAG.
Question 974
A public agency proposes fine-tuning a language model on volatile benefit rules that change every month. Postdeployment review has identified the same issue after a material service change. What is the strongest proportionate corrective design decision?
- Deploy an appropriately validated edge or on-prem design with bounded synchronization when cloud reliance is unsafe.
- Pilot a small local model or simple algorithm when it meets safety and accuracy constraints.
- Use controlled retrieval from approved current documents with access filtering and source-aware answers.
- Prefer controlled retrieval of versioned policy sources for rapidly changing factual rules; fine-tune only for justified behavior goals.
Explanation
Correct Answer
D:Prefer controlled retrieval of versioned policy sources for rapidly changing factual rules; fine-tune only for justified behavior goals. Correct. It directly reduces or governs the specific documented risk. Fine-tuning adapts model behavior but is not always suitable for rapidly changing facts.
Incorrect Answers
A:Deploy an appropriately validated edge or on-prem design with bounded synchronization when cloud reliance is unsafe. Not selected. That action treats another control objective, not the specific gap described. It belongs with Cloud edge on-prem choice: AI deployment topology should match operational hazards and data obligations. The present case instead concerns Fine-tuning versus prompt or RAG.
B:Pilot a small local model or simple algorithm when it meets safety and accuracy constraints. Not selected. That action treats another control objective, not the specific gap described. It belongs with Small versus large model: Model size is one factor in suitability and can increase cost and exposure without adding needed value. The present case instead concerns Fine-tuning versus prompt or RAG.
C:Use controlled retrieval from approved current documents with access filtering and source-aware answers. Not selected. That action treats another control objective, not the specific gap described. It belongs with Retrieval augmented generation: RAG can improve grounding but does not eliminate hallucination, leakage or access-control obligations. The present case instead concerns Fine-tuning versus prompt or RAG.
Question 975
A public agency proposes fine-tuning a language model on volatile benefit rules that change every month. Postdeployment review has identified the same issue after a material service change. Which observable evidence best demonstrates the chosen safeguard is working?
- Test corrupted images, accented speech and conflicting text for safe handling and escalation.
- Test version-changed policy questions and confirm citations, permissions and abstention behavior.
- Change a policy rule and verify the system retrieves the latest approved version without retraining.
- Assess business objectives, data availability, operator skills, network conditions and failure impact at each deployment site.
Explanation
Correct Answer
C:Change a policy rule and verify the system retrieves the latest approved version without retraining. Correct. It yields measurable evidence that the case-specific safeguard works. Fine-tuning adapts model behavior but is not always suitable for rapidly changing facts.
Incorrect Answers
A:Test corrupted images, accented speech and conflicting text for safe handling and escalation. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Multimodal model assessment: Multimodal AI adds failure channels that text-only benchmarks cannot substantiate. The present case instead concerns Fine-tuning versus prompt or RAG.
B:Test version-changed policy questions and confirm citations, permissions and abstention behavior. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Retrieval augmented generation: RAG can improve grounding but does not eliminate hallucination, leakage or access-control obligations. The present case instead concerns Fine-tuning versus prompt or RAG.
D:Assess business objectives, data availability, operator skills, network conditions and failure impact at each deployment site. Not selected. This action relates to Deployment-context feasibility. The deployment environment and workforce can create new risk even when the model was sound in development. For the validate decision, it does not resolve the specifically described issue of Fine-tuning versus prompt or RAG in A public agency proposes fine-tuning a language model on volatile benefit rules that change every month. Postdeployment review has identified the same issue after a material service change.
Question 976
A public agency proposes fine-tuning a language model on volatile benefit rules that change every month. Postdeployment review has identified the same issue after a material service change. A material exception remains unresolved. Which containment decision best protects affected parties?
- Remove stale model-baked claims and fall back to verified source documents.
- Limit deployment to sites with validated readiness and manual fallback.
- Disable unsupported policy advice and provide verified staff assistance during remediation.
- Move to a safe manual or local mode if connectivity loss undermines reliable inference.
Explanation
Correct Answer
A:Remove stale model-baked claims and fall back to verified source documents. Correct. It limits further reliance or exposure while the risk remains unresolved. Fine-tuning adapts model behavior but is not always suitable for rapidly changing facts.
Incorrect Answers
B:Limit deployment to sites with validated readiness and manual fallback. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Deployment-context feasibility: The deployment environment and workforce can create new risk even when the model was sound in development. The present case instead concerns Fine-tuning versus prompt or RAG.
C:Disable unsupported policy advice and provide verified staff assistance during remediation. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Retrieval augmented generation: RAG can improve grounding but does not eliminate hallucination, leakage or access-control obligations. The present case instead concerns Fine-tuning versus prompt or RAG.
D:Move to a safe manual or local mode if connectivity loss undermines reliable inference. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Cloud edge on-prem choice: AI deployment topology should match operational hazards and data obligations. The present case instead concerns Fine-tuning versus prompt or RAG.
Question 977
An HR firm buys a model tested for internal scheduling and now wants it to rank external job applicants. The operational owner must decide whether the current deployment can safely proceed. Which diagnostic task establishes the risk relevant to this described system?
- Compare the supplier's tested purpose and impact assessment with the new high-stakes employment use.
- Review prompt retention, training reuse, subprocessors, transfer locations and deletion rights.
- Review indemnities, liability caps, exclusions, claims procedures and jurisdiction-specific obligations.
- Assess vendor concentration, portability, access continuity, model dependency and exit obligations.
Explanation
Correct Answer
A:Compare the supplier's tested purpose and impact assessment with the new high-stakes employment use. Correct. It establishes the relevant factual risk boundary before implementation. A vendor's original impact assessment does not cover every downstream purpose and population.
Incorrect Answers
B:Review prompt retention, training reuse, subprocessors, transfer locations and deletion rights. Not selected. That check primarily diagnoses a different governance issue. It belongs with Vendor data-use clauses: Vendor AI contract terms can change the privacy and confidentiality risk of deployment. The present case instead concerns Predeployment impact reassessment.
C:Review indemnities, liability caps, exclusions, claims procedures and jurisdiction-specific obligations. Not selected. That check primarily diagnoses a different governance issue. It belongs with Indemnity and liability allocation: Legal and commercial risk allocation requires scrutiny beyond marketing promises of responsible AI. The present case instead concerns Predeployment impact reassessment.
D:Assess vendor concentration, portability, access continuity, model dependency and exit obligations. Not selected. That check primarily diagnoses a different governance issue. It belongs with Supplier exit strategy: Third-party AI risk management includes practical exit and concentration resilience. The present case instead concerns Predeployment impact reassessment.
Question 978
An HR firm buys a model tested for internal scheduling and now wants it to rank external job applicants. The operational owner must decide whether the current deployment can safely proceed. What is the strongest proportionate corrective design decision?
- Negotiate change-notice, data-return, deletion and service-transition provisions and test an alternative.
- Agree proportionate allocation of intellectual-property, security and incident cooperation duties with counsel.
- Conduct a new context-specific impact assessment and obtain required safeguards before deployment.
- Maintain end-to-end assurance and responsibility across model creation and deployment.
Explanation
Correct Answer
C:Conduct a new context-specific impact assessment and obtain required safeguards before deployment. Correct. It directly reduces or governs the specific documented risk. A vendor's original impact assessment does not cover every downstream purpose and population.
Incorrect Answers
A:Negotiate change-notice, data-return, deletion and service-transition provisions and test an alternative. Not selected. That action treats another control objective, not the specific gap described. It belongs with Supplier exit strategy: Third-party AI risk management includes practical exit and concentration resilience. The present case instead concerns Predeployment impact reassessment.
B:Agree proportionate allocation of intellectual-property, security and incident cooperation duties with counsel. Not selected. That action treats another control objective, not the specific gap described. It belongs with Indemnity and liability allocation: Legal and commercial risk allocation requires scrutiny beyond marketing promises of responsible AI. The present case instead concerns Predeployment impact reassessment.
D:Maintain end-to-end assurance and responsibility across model creation and deployment. Not selected. That action treats another control objective, not the specific gap described. It belongs with Own-model development obligations: Owning a model increases control but also concentrates lifecycle and potential legal responsibilities. The present case instead concerns Predeployment impact reassessment.
Question 979
An HR firm buys a model tested for internal scheduling and now wants it to rank external job applicants. The operational owner must decide whether the current deployment can safely proceed. Which observable evidence best demonstrates the chosen safeguard is working?
- Run a tabletop scenario and confirm that contract obligations support timely notice and claim handling.
- Exercise a supplier failure scenario and verify continuity and secure data deletion.
- Verify affected applicants, bias risks, appeal routes and operators were assessed for the changed purpose.
- Audit reproducible data, model testing, approvals and consumer-facing safeguards.
Explanation
Correct Answer
C:Verify affected applicants, bias risks, appeal routes and operators were assessed for the changed purpose. Correct. It yields measurable evidence that the case-specific safeguard works. A vendor's original impact assessment does not cover every downstream purpose and population.
Incorrect Answers
A:Run a tabletop scenario and confirm that contract obligations support timely notice and claim handling. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Indemnity and liability allocation: Legal and commercial risk allocation requires scrutiny beyond marketing promises of responsible AI. The present case instead concerns Predeployment impact reassessment.
B:Exercise a supplier failure scenario and verify continuity and secure data deletion. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Supplier exit strategy: Third-party AI risk management includes practical exit and concentration resilience. The present case instead concerns Predeployment impact reassessment.
D:Audit reproducible data, model testing, approvals and consumer-facing safeguards. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Own-model development obligations: Owning a model increases control but also concentrates lifecycle and potential legal responsibilities. The present case instead concerns Predeployment impact reassessment.
Question 980
An HR firm buys a model tested for internal scheduling and now wants it to rank external job applicants. The operational owner must decide whether the current deployment can safely proceed. A material exception remains unresolved. Which containment decision best protects affected parties?
- Halt distribution of noncompliant derivatives until counsel approves corrective terms.
- Reject the unsupported employment use while suitable assessment and controls are missing.
- Stop sending protected prompts and move to a compliant service if terms cannot be controlled.
- Delay rollout of the proprietary model where internal assurance artifacts are missing.
Explanation
Correct Answer
B:Reject the unsupported employment use while suitable assessment and controls are missing. Correct. It limits further reliance or exposure while the risk remains unresolved. A vendor's original impact assessment does not cover every downstream purpose and population.
Incorrect Answers
A:Halt distribution of noncompliant derivatives until counsel approves corrective terms. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Licensing and downstream rights: Open weights and permissive access do not mean there are no licensing obligations. The present case instead concerns Predeployment impact reassessment.
C:Stop sending protected prompts and move to a compliant service if terms cannot be controlled. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Vendor data-use clauses: Vendor AI contract terms can change the privacy and confidentiality risk of deployment. The present case instead concerns Predeployment impact reassessment.
D:Delay rollout of the proprietary model where internal assurance artifacts are missing. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Own-model development obligations: Owning a model increases control but also concentrates lifecycle and potential legal responsibilities. The present case instead concerns Predeployment impact reassessment.