View Full Isaca AAIR Exam Dumps and Practice Test Dumps.

Q101. An enterprise has created an AI governance council, but project teams remain uncertain about which matters must be escalated to it. What is the BEST improvement?

1. Require the council to review every technical change
   2. Allow project managers to decide informally
   3. Eliminate escalation to speed development
   4. Define a governance charter with decision rights, escalation criteria, scope, and authority

Correct Answer: 4. Define a governance charter with decision rights, escalation criteria, scope, and authority

Explanation: A governance body is effective only when stakeholders understand its purpose and authority. A formal charter should define which AI decisions require council review, what thresholds trigger escalation, which roles participate, and how exceptions are handled. Examples may include high-risk use cases, policy exceptions, material model changes, regulatory concerns, or significant residual-risk acceptance. Requiring review of every minor technical change can create unnecessary bottlenecks, while informal escalation produces inconsistent governance. Clear decision rights enable project teams to operate efficiently while ensuring that material AI risks receive appropriate oversight.

Q102. An organization maintains an AI inventory, but several deployed AI-enabled applications are discovered outside the inventory during a review. What is the MOST appropriate response?

1. Remove the inventory because it is incomplete
   2. Reconcile the inventory with discovery sources and establish a recurring completeness-control process
   3. Document only high-risk systems going forward
   4. Assume application owners will eventually update the inventory

Correct Answer: 2. Reconcile the inventory with discovery sources and establish a recurring completeness-control process

Explanation: An AI inventory is useful only if it is sufficiently complete and current. The organization should reconcile declared systems with procurement records, cloud usage, software inventories, model repositories, vendor lists, and other discovery sources. Identified gaps should be corrected and root causes addressed through recurring controls and ownership responsibilities. Simply documenting high-risk systems may leave unknown lower-risk systems that later become material. Periodic reconciliation helps detect shadow AI and supports risk classification, regulatory compliance, lifecycle management, supply-chain oversight, incident response, and reporting.

Q103. A new AI regulation becomes effective in six months and introduces additional documentation obligations. What should the risk function do FIRST?

1. Perform a gap assessment against current AI processes and controls
   2. Wait until the effective date before acting
   3. Immediately stop all AI development
   4. Assume existing policies already satisfy the regulation

Correct Answer: 1. Perform a gap assessment against current AI processes and controls

Explanation: When new regulatory requirements are identified, the organization should first determine how current governance, processes, controls, and documentation compare with those obligations. A gap assessment identifies which requirements are already satisfied and which require remediation. This enables management to prioritize work, assign owners, estimate resources, and complete changes before the regulation becomes effective. Waiting until enforcement begins increases compliance risk, while suspending all AI activity may be disproportionate. Regulatory change management should be proactive, structured, documented, and integrated with existing compliance and AI governance processes.

Q104. An organization wants to reduce the amount of sensitive customer information exposed during centralized model training. Which privacy-enhancing approach MOST directly allows model training to occur closer to where data resides?

1. Increasing the model’s parameter count
   2. Keeping one central raw-data repository
   3. Federated learning
   4. Extending data-retention periods

Correct Answer: 3. Federated learning

Explanation: Federated learning enables participating systems or locations to train models locally and share model updates rather than requiring all raw data to be centralized. This can reduce certain privacy and data-transfer risks, although it does not eliminate them. Model updates can still potentially leak information, and additional controls may be required for secure aggregation, authentication, poisoning resistance, and privacy. The suitability of federated learning depends on architecture and use case. It should therefore be viewed as a privacy-enhancing design option rather than a guarantee that sensitive information can never be exposed.

Q105. An organization wants to publish statistical insights from a sensitive dataset while reducing the risk that individual records can be inferred from the results. Which technique should it consider?

1. Differential privacy
   2. Increasing retention indefinitely
   3. Removing access logging
   4. Using larger model files

Correct Answer: 1. Differential privacy

Explanation: Differential privacy introduces carefully controlled statistical noise so aggregate analysis can remain useful while reducing the ability to infer whether a specific individual’s data contributed to the result. The effectiveness depends on parameters such as the privacy budget and how queries are composed over time. It is not appropriate for every use case and may involve a tradeoff between utility and privacy. Differential privacy should be combined with broader privacy governance, access controls, minimization, and retention practices. It is a technical privacy-enhancing mechanism rather than a replacement for legal or organizational privacy obligations.

Q106. A high-impact AI decision system allows a reviewer to override the model’s recommendation. What makes this control MOST effective?

1. Reviewers always approve the model output
   2. Overrides are undocumented to protect reviewer privacy
   3. Review occurs only after irreversible action
   4. Reviewers have authority, sufficient information, training, and a practical ability to challenge the model**

Correct Answer: 4. Reviewers have authority, sufficient information, training, and a practical ability to challenge the model

Explanation: Human oversight is meaningful only when reviewers can genuinely understand and challenge AI recommendations. They need appropriate authority, contextual information, training, sufficient time, and clear escalation mechanisms. A nominal approval step where people routinely accept the model without scrutiny creates automation bias rather than effective oversight. Overrides should generally be logged and analyzed because patterns may reveal model weaknesses or unclear procedures. Human-in-the-loop controls should therefore be designed around actual decision quality rather than merely inserting a person into the workflow.

Q107. A safety-critical AI service begins producing clearly harmful outputs. Which capability is MOST important for immediate containment?

1. A longer model training cycle
   2. A tested mechanism to disable, isolate, or switch away from the AI function
   3. Increasing the model’s autonomy
   4. Delaying response until root cause is confirmed

Correct Answer: 2. A tested mechanism to disable, isolate, or switch away from the AI function

Explanation: High-impact AI systems should have practical containment mechanisms so harmful behavior can be stopped quickly. Depending on architecture, this may include disabling a model endpoint, switching to a prior version, invoking a manual process, removing tool permissions, or isolating the affected component. The mechanism must be tested and accompanied by defined decision authority. Root-cause investigation remains important but should not delay containment when ongoing harm is possible. Designing for safe shutdown and fallback is part of lifecycle risk management and operational resilience.

Q108. A risk manager wants evidence that only authorized personnel have accessed a sensitive model repository. Which control provides the MOST useful evidence?

1. Model documentation only
   2. Employee awareness posters
   3. Tamper-resistant access logs linked to authenticated identities
   4. The model’s accuracy score

Correct Answer: 3. Tamper-resistant access logs linked to authenticated identities

Explanation: Access logs tied to authenticated identities provide evidence about who accessed sensitive model assets, when access occurred, and potentially what actions were performed. Protecting those logs from unauthorized modification strengthens their value for monitoring, investigations, and assurance. Access controls themselves are preventive, while logging provides detective and evidentiary capability. Awareness materials and model documentation do not demonstrate actual access behavior. For sensitive model repositories, organizations should combine identity management, least privilege, logging, periodic access review, and alerting for anomalous activity.

Q109. An AI model requires specialized GPU infrastructure from one provider, and migrating to another provider would take more than a year. Which risk characteristic is MOST relevant?

1. Explainability risk
   2. Vendor lock-in and exit risk
   3. Model calibration risk
   4. Data labeling risk

Correct Answer: 2. Vendor lock-in and exit risk

Explanation: When an organization cannot readily move an AI workload away from a provider, it faces lock-in and exit risk. A provider outage, contract dispute, price increase, geopolitical issue, regulatory problem, or strategic change could leave the organization with limited alternatives. Risk management should consider portability, migration time, proprietary interfaces, data-transfer requirements, skills, replacement providers, and contractual exit terms. The organization may accept some lock-in if benefits justify it, but the exposure should be understood and reflected in continuity and supply-chain planning.

Q110. An organization identifies several AI risks that could materialize very quickly once triggered, leaving little time for management response. Which characteristic should be emphasized in prioritization?

1. Risk velocity
   2. Model size
   3. Number of developers
   4. Document length

Correct Answer: 1. Risk velocity

Explanation: Risk velocity describes how quickly a risk can move from trigger to material impact. Two risks may have similar likelihood and impact, but the one that develops within minutes may require more automated detection and predefined response than one that unfolds over months. AI risks such as rapidly propagating harmful outputs, automated decisions, or compromised autonomous tools can have high velocity. Including velocity in risk analysis can improve monitoring, escalation, and incident preparedness. It does not replace likelihood and impact but provides additional information about the time available to respond effectively.

Q111. A generative AI application uses retrieval from an internal knowledge base. Attackers add malicious documents intended to manipulate future model responses. What threat is MOST directly illustrated?

1. Ordinary model drift
   2. Business interruption
   3. Data minimization
   4. Retrieval or knowledge-base poisoning**

Correct Answer: 4. Retrieval or knowledge-base poisoning

Explanation: Retrieval poisoning occurs when an attacker introduces manipulated or malicious content into a knowledge source that an AI system later retrieves. The content can distort answers, mislead users, or contain indirect prompt-injection instructions. Controls may include source approval, write-access restrictions, integrity monitoring, document provenance, content review, isolation of untrusted text, and retrieval filtering. This differs from traditional training-data poisoning because the compromised material may affect outputs dynamically without retraining the model. Retrieval pipelines therefore require security and governance controls similar to other important enterprise data sources.

Q112. An organization accepts a moderate AI risk for six months while a replacement control is implemented. What should happen when the acceptance period expires?

1. The risk should automatically disappear from the register
   2. Acceptance should become permanent
   3. The risk and current residual exposure should be reassessed and reauthorized if still necessary
   4. Monitoring should stop

Correct Answer: 3. The risk and current residual exposure should be reassessed and reauthorized if still necessary

Explanation: Time-limited risk acceptance should include an expiration or review date. At that point, management should determine whether planned remediation was completed, whether the risk changed, and whether remaining exposure is still within tolerance. If continued acceptance is necessary, the appropriate risk owner should explicitly reauthorize it rather than allowing the decision to roll forward automatically. This prevents temporary exceptions from becoming permanent unmanaged exposures. The risk register should retain relevant history so stakeholders can understand prior decisions, treatment progress, and current accountability.

Q113. A risk team wants to understand how failure of one AI service could trigger failures in dependent business processes and other AI systems. Which technique is MOST useful?

1. Dependency and interdependency mapping
   2. Increasing model accuracy only
   3. Removing asset ownership information
   4. Reviewing model source code alone

Correct Answer: 1. Dependency and interdependency mapping

Explanation: AI systems often depend on shared data, APIs, identity services, cloud infrastructure, models, and business processes. Dependency mapping reveals how failure of one component may cascade into other services and can expose concentration or common-cause risks. This supports business impact analysis, continuity planning, incident response, and architecture decisions. A model may perform accurately yet still create enterprise exposure if critical dependencies are poorly understood. Mapping should include both upstream dependencies required by the AI service and downstream processes that rely on its output.

Q114. A high-risk AI control is self-assessed as effective by the team that operates it. What would provide the STRONGEST additional assurance?

1. Accept the self-assessment without evidence
   2. Independent testing of control design and operating effectiveness
   3. Increase the number of policy documents
   4. Ask the AI model to evaluate itself

Correct Answer: 3. Independent testing of control design and operating effectiveness

Explanation: Independent testing provides greater assurance because it reduces the risk of bias or oversight by the team responsible for designing and operating the control. The assessor should evaluate whether the control is appropriately designed to address the stated AI risk and whether it operates consistently in practice. Evidence might include logs, samples, configuration, approvals, monitoring results, and exceptions. Self-assessment remains valuable for ongoing ownership, but material risks often warrant additional challenge from a sufficiently independent and competent party.

Q115. An AI vendor’s service-level agreement measures uptime but contains no commitment regarding material model changes. Which additional requirement would BEST address lifecycle risk?

1. Require all vendor employees to work from one office
   2. Increase the contract font size
   3. Allow unlimited undocumented model updates
   4. Require notification and appropriate assessment of material model or service changes**

Correct Answer: 4. Require notification and appropriate assessment of material model or service changes

Explanation: AI services can change behavior when providers modify models, safety controls, data handling, APIs, or underlying infrastructure. These changes can invalidate prior validation or alter compliance, performance, fairness, or security characteristics. Contractual requirements should therefore address notification of material changes and may provide time for customer assessment, testing, or approval depending on criticality. Uptime alone does not address behavioral or lifecycle risk. Strong third-party agreements should reflect the full set of material dependencies and risks associated with the AI service.

Q116. A critical AI vendor must be replaced. Which activity provides the BEST evidence that the organization’s exit plan is viable?

1. A tested migration or transition exercise
   2. A statement that migration should be possible
   3. Vendor marketing material
   4. An automatically renewed contract

Correct Answer: 2. A tested migration or transition exercise

Explanation: An exit plan is most credible when key aspects have been tested rather than merely documented. A transition exercise can reveal unexpected dependencies, proprietary formats, data-transfer limitations, retraining requirements, licensing constraints, skills gaps, and realistic migration timelines. Not every organization needs to perform a full provider migration regularly, but critical components should be tested sufficiently to support confidence in the plan. Exit testing also helps estimate recovery or transition effort if a vendor suddenly becomes unavailable or unacceptable.

Q117. During an AI incident investigation, potentially relevant model prompts and outputs may later be needed for regulatory review. What is the MOST important action?

1. Preserve relevant evidence with integrity and chain-of-custody considerations
   2. Delete prompts immediately to reduce storage
   3. Allow unrestricted editing of incident records
   4. Store only verbal recollections from responders

Correct Answer: 3. Preserve relevant evidence with integrity and chain-of-custody considerations

Explanation: Incident evidence should be preserved in a way that supports later investigation, legal review, regulatory inquiry, or disciplinary action where relevant. For AI incidents, evidence may include prompts, outputs, model versions, logs, configuration, retrieved content, approval records, and data lineage. Access and transfer should be documented, and integrity controls should reduce the risk of accidental or deliberate modification. Retention must also comply with privacy and legal requirements. Deleting evidence prematurely can prevent the organization from reconstructing what occurred and why.

Q118. A public-facing AI incident may materially affect customers and attract media attention. Which planning activity is MOST important in addition to technical containment?

1. Increase model training frequency
   2. Prepare coordinated crisis communication and stakeholder-notification procedures
   3. Disable all incident documentation
   4. Allow individual employees to issue public statements independently

Correct Answer: 1. Prepare coordinated crisis communication and stakeholder-notification procedures

Explanation: Material AI incidents can have reputational, legal, customer, and regulatory consequences in addition to technical impact. Organizations should define who communicates, which stakeholders must be notified, how information is verified, and what regulatory or contractual timelines apply. Messaging should be coordinated with legal, privacy, risk, communications, and business leaders to avoid contradictory or premature statements. Technical containment remains essential, but poor communication can amplify harm. Crisis communication should therefore be incorporated into AI incident response planning for significant scenarios.

Q119. An AI system automatically processes applications, but policy requires human intervention whenever confidence falls below an approved threshold. What does this threshold primarily provide?

1. A model-training schedule
   2. A defined escalation boundary for uncertain decisions
   3. A vendor exit clause
   4. A data-retention period

Correct Answer: 2. A defined escalation boundary for uncertain decisions

Explanation: A confidence-based escalation threshold determines when the AI system should stop making autonomous decisions and route the case for human review. This can reduce the risk of acting on uncertain model outputs, particularly when consequences are significant. The threshold should be validated against actual model behavior because confidence scores may not be perfectly calibrated. Monitoring should also evaluate whether too many or too few cases are escalated. The control is useful only when reviewers have sufficient authority, training, information, and time to make a meaningful independent decision.

Q120. An organization uses an AI system to recommend enterprise risk treatments. What is the BEST governance approach?

1. Allow the AI to approve all treatments automatically
   2. Replace risk owners with the recommendation engine
   3. Ignore the rationale behind recommendations
   4. Treat AI recommendations as decision support and retain accountable human approval for material risk responses**

Correct Answer: 4. Treat AI recommendations as decision support and retain accountable human approval for material risk responses

Explanation: AI can analyze risk data, compare scenarios, and suggest treatment alternatives, but accountability for material enterprise risk decisions should remain with authorized human risk owners and governance bodies. Decision-makers should understand the inputs, limitations, uncertainties, and rationale behind recommendations and verify that proposed treatments align with organizational risk appetite. Automated recommendations may improve consistency and speed, but they can also reflect incomplete data or flawed assumptions. The strongest governance model uses AI to support professional judgment rather than replacing accountable risk ownership.