View Full Isaca AAIR Exam Dumps and Practice Test Dumps.

Q41. A business unit proposes an AI initiative because competitors are deploying similar technology. What should the risk professional evaluate FIRST?

1. Whether the competitor uses the same AI vendor
   2. Whether the model contains more parameters than competing systems
   3. Whether implementation can begin before governance review
   4. Whether the proposed AI use case creates measurable value aligned with enterprise objectives

Correct Answer: 4. Whether the proposed AI use case creates measurable value aligned with enterprise objectives

Explanation: AI adoption should begin with a legitimate business problem or opportunity rather than competitive pressure alone. The risk professional should determine whether the use case supports enterprise strategy, expected value, stakeholder needs, and organizational risk appetite. Once the business rationale is established, technical, legal, ethical, security, and operational risks can be evaluated in context. Deploying AI simply because competitors have done so can create unnecessary exposure without meaningful benefit. ISACA’s AAIR approach emphasizes evaluating AI use cases in relation to organizational goals, risk appetite, and value creation rather than treating AI implementation as an objective by itself.

Q42. An organization is developing an AI-specific risk taxonomy. What is the MOST important characteristic of the taxonomy?

1. It should use terminology unrelated to the enterprise risk taxonomy
   2. It should map AI risks into existing enterprise risk categories while capturing AI-specific characteristics
   3. It should contain only cybersecurity threats
   4. It should exclude operational and reputational impacts

Correct Answer: 2. It should map AI risks into existing enterprise risk categories while capturing AI-specific characteristics

Explanation: An AI risk taxonomy should help the organization describe and classify AI-specific exposures while remaining compatible with existing enterprise risk management. Categories may include model, data, privacy, legal, security, fairness, reliability, third-party, operational, and societal risks, but these should connect to the broader risk taxonomy used by management. This supports aggregation, consistent reporting, ownership, and comparison across risk types. Creating an entirely separate vocabulary can fragment governance. Limiting the taxonomy to cybersecurity also overlooks major AI concerns such as discrimination, explainability, regulatory compliance, intellectual property, reliability, and business-process impact.

Q43. What is the PRIMARY purpose of defining quantitative or qualitative AI risk tolerance thresholds?

1. To identify the level of variation or exposure management is willing to accept before escalation or treatment is required
   2. To guarantee that all AI risks are eliminated
   3. To determine the programming language used for AI models
   4. To replace executive risk appetite statements

Correct Answer: 1. To identify the level of variation or exposure management is willing to accept before escalation or treatment is required

Explanation: Risk tolerance translates broader risk appetite into more operational boundaries. For an AI system, tolerance might specify maximum acceptable error rates, fairness deviations, service downtime, privacy incidents, or other measurable exposure thresholds. When actual performance approaches or exceeds these limits, defined escalation, remediation, or suspension procedures can be triggered. Tolerances do not eliminate risk and should remain aligned with enterprise risk appetite rather than replacing it. Clear thresholds make AI governance actionable because teams know when normal monitoring is sufficient and when management attention or additional risk treatment becomes necessary.

Q44. An AI governance committee is being established. Which responsibility is MOST appropriate for the committee?

1. Writing every line of model source code
   2. Performing all daily production support tasks
   3. Providing cross-functional oversight of material AI risks, policies, exceptions, and strategic alignment
   4. Approving employee vacation requests

Correct Answer: 3. Providing cross-functional oversight of material AI risks, policies, exceptions, and strategic alignment

Explanation: AI governance usually requires input from business leadership, risk, legal, privacy, security, compliance, data, technology, and other stakeholders. A governance committee can provide enterprise-wide oversight by reviewing significant AI use cases, policy exceptions, material risk exposures, accountability, regulatory issues, and alignment with organizational objectives. It should not replace day-to-day technical or operational responsibilities assigned to accountable teams. Effective governance creates clear decision rights and escalation paths while maintaining separation between oversight and execution. Cross-functional oversight is particularly important because AI risk often spans multiple traditional organizational boundaries.

Q45. An organization is assessing the environmental impact of training a large AI model. Which issue is MOST relevant to its AI risk evaluation?

1. The color of the development interface
   2. Energy consumption and associated environmental sustainability impacts
   3. The number of project meetings
   4. Whether developers use identical laptops

Correct Answer: 2. Energy consumption and associated environmental sustainability impacts

Explanation: Responsible AI risk management can include environmental, social, and governance considerations. Large-scale model training and inference can consume significant energy and computing resources, potentially affecting sustainability commitments, operating costs, carbon targets, and stakeholder expectations. The significance varies by model, workload, infrastructure source, and organizational objectives. Environmental impacts should therefore be assessed proportionately alongside other AI risks rather than ignored as purely technical considerations. ISACA’s AAIR outline explicitly includes ethical, societal, and ESG implications within trustworthy AI governance, making sustainability a legitimate consideration in enterprise AI risk assessment.

Q46. A company plans to use copyrighted material to train an internal generative AI model. What should be evaluated before training begins?

1. Only whether the data improves model accuracy
   2. Only the storage cost of the dataset
   3. Whether users prefer the training material
   4. Intellectual-property rights, licenses, permitted use, and applicable legal obligations

Correct Answer: 4. Intellectual-property rights, licenses, permitted use, and applicable legal obligations

Explanation: Training data can create intellectual-property risk when copyrighted, licensed, proprietary, or otherwise restricted material is used without appropriate rights. The organization should determine the data’s ownership, licensing terms, permitted uses, contractual restrictions, and relevant law before incorporating it into model training. The assessment should also consider whether generated outputs could reproduce protected material. Model quality does not override legal rights. Documented provenance and legal review help demonstrate responsible use and reduce the risk of disputes, regulatory issues, or requirements to remove data or retrain models later.

Q47. During model selection, two algorithms provide similar accuracy, but one is significantly easier to explain to regulators and affected users. What should the organization do?

1. Select the more complex model automatically
   2. Ignore explainability because accuracy is similar
   3. Consider explainability together with performance and the risk requirements of the use case
   4. Select whichever model was developed first

Correct Answer: 3. Consider explainability together with performance and the risk requirements of the use case

Explanation: Model selection should reflect the entire risk and business context, not accuracy alone. When an AI system supports consequential decisions, explainability may be important for regulatory compliance, challenge processes, customer communication, validation, and incident investigation. If two models offer comparable predictive performance, the model providing sufficient transparency may offer better overall risk characteristics. However, explainability is only one factor; robustness, fairness, privacy, security, maintainability, and business value should also be considered. Risk-based model selection seeks the solution most suitable for the intended use rather than simply the technically most sophisticated algorithm.

Q48. What is the PRIMARY value of maintaining a model card or equivalent structured AI system documentation?

1. It documents intended use, limitations, performance characteristics, assumptions, and other information needed for governance
   2. It guarantees that the AI system is legally compliant
   3. It replaces model monitoring after deployment
   4. It eliminates the need for independent validation

Correct Answer: 1. It documents intended use, limitations, performance characteristics, assumptions, and other information needed for governance

Explanation: Structured model documentation helps stakeholders understand what a model was designed to do, how it was developed and evaluated, important limitations, expected inputs, performance characteristics, known risks, and inappropriate uses. This supports validation, approval, monitoring, incident response, change management, and eventual decommissioning. Documentation can also improve transparency between developers, risk teams, business owners, and external stakeholders. A model card does not itself prove compliance or effectiveness and must remain current as the system changes. Documentation is therefore an important governance artifact rather than a substitute for controls and ongoing oversight.

Q49. Before deploying a generative AI application, the organization intentionally tests whether adversarial users can manipulate it into revealing protected information. What activity is being performed?

1. AI red-team testing
   2. Data archival
   3. Business continuity testing only
   4. Risk acceptance

Correct Answer: 4. AI red-team testing

Explanation: AI red teaming deliberately challenges an AI system using adversarial scenarios to discover weaknesses before attackers or ordinary users encounter them. Testing may examine prompt injection, sensitive-data disclosure, policy bypass, unsafe content, tool misuse, excessive agency, or other failure modes relevant to the use case. Findings should feed into control improvements, residual risk assessment, and deployment decisions. Red teaming complements ordinary functional testing because it assumes hostile or unexpected behavior rather than normal intended use. High-impact or externally exposed AI applications may require particularly robust adversarial testing before deployment and after material changes.

Q50. An image-recognition AI performs accurately on clean images but fails when very small intentionally crafted changes are introduced into those images. Which issue does this demonstrate?

1. Adversarial-example vulnerability
   2. Data retention failure
   3. Third-party concentration
   4. Business continuity failure

Correct Answer: 1. Adversarial-example vulnerability

Explanation: Adversarial examples are inputs deliberately modified in subtle ways to cause an AI model to produce incorrect predictions. The changes may be nearly imperceptible to humans while exploiting weaknesses in the model’s learned decision boundaries. Organizations deploying AI in adversarial environments should assess robustness against these manipulations and consider controls such as adversarial testing, input validation, monitoring, model hardening, and human review. The significance depends on the use case; an error in a low-impact recommendation engine differs greatly from an error in a safety-critical or security-related AI system.

Q51. An AI project collects significantly more personal information than is necessary for its stated purpose. Which risk-management principle is MOST directly being violated?

1. Vendor concentration management
   2. Data minimization
   3. Disaster recovery
   4. Model explainability

Correct Answer: 2. Data minimization

Explanation: Data minimization means collecting and processing only the personal or sensitive information reasonably necessary for the defined purpose. Excess data increases privacy, security, legal, retention, and breach exposure without necessarily providing proportional business value. AI teams should define data requirements during design and periodically reassess whether each data element remains necessary. Minimization can also reduce the impact of an incident and simplify compliance obligations. AI systems can create incentives to gather large datasets “just in case,” so strong governance is needed to ensure data collection remains proportional to legitimate business needs.

Q52. A production model is scheduled for retraining whenever a predefined performance threshold is breached. What governance activity is MOST important when retraining occurs?

1. Treat the retrained model as automatically approved because the original version was approved
   2. Skip documentation if accuracy improves
   3. Apply change management, versioning, validation, and approval proportionate to the change
   4. Delete the previous model immediately before testing

Correct Answer: 3. Apply change management, versioning, validation, and approval proportionate to the change

Explanation: Retraining can materially change model behavior even when the architecture and source code remain the same. New data can alter accuracy, fairness, explainability, security characteristics, or business outcomes. The retrained model should therefore be versioned, documented, tested, compared with the current version, and approved according to the organization’s risk-based change process. The extent of review can depend on materiality. Retaining the previous known-good model can also support rollback. Treating every retraining event as automatically acceptable undermines lifecycle governance and can allow untested behavioral changes into production.

Q53. Why is model versioning important in AI lifecycle management?

1. It enables traceability of which model version produced particular outcomes and supports controlled rollback
   2. It guarantees that newer models are always better
   3. It eliminates the need to retain documentation
   4. It prevents all AI-related incidents

Correct Answer: 1. It enables traceability of which model version produced particular outcomes and supports controlled rollback

Explanation: Model versioning allows organizations to identify which model was active at a particular time and link decisions or incidents to the correct training data, code, parameters, validation evidence, and approvals. It also makes rollback practical when a newer version performs poorly. Without version control, teams may be unable to reproduce outcomes or determine which model generated a disputed decision. Versioning should be combined with documentation, change management, data lineage, and deployment records. A higher version number does not guarantee improved quality; each material version still requires appropriate evaluation.

Q54. A new AI model has been deployed, but the organization keeps the previous validated model available for rapid restoration if unexpected problems occur. What risk treatment capability does this provide?

1. Risk transfer
   2. Data poisoning
   3. Rollback capability
   4. Shadow AI detection

Correct Answer: 3. Rollback capability

Explanation: Rollback allows an organization to return quickly to a previous known-good version when a newly deployed model behaves unexpectedly. This reduces operational impact and is especially valuable when post-deployment monitoring detects accuracy, fairness, security, or reliability problems. Effective rollback requires preserved prior artifacts, compatible dependencies, deployment procedures, and clear decision authority. It should be tested rather than assumed to work. Rollback does not remove the need for predeployment validation; it complements preventive controls by providing a recovery mechanism when production conditions reveal issues that earlier testing did not identify.

Q55. When developing an AI risk scenario, which combination provides the MOST complete basis for assessment?

1. Model name and development date only
   2. Threat event, vulnerability or condition, affected asset, and potential business impact
   3. Vendor revenue and employee count only
   4. Number of training records alone

Correct Answer: 4. Threat event, vulnerability or condition, affected asset, and potential business impact

Explanation: A useful risk scenario describes how a threat or adverse event could exploit a vulnerability or risky condition, which assets or processes would be affected, and what consequences could result. This structure supports evaluation of likelihood, impact, controls, and treatment options. For AI, scenarios may involve malicious manipulation, biased outputs, model failure, privacy leakage, vendor outages, or regulatory violations. Simply naming the model does not explain the pathway to harm. Well-defined scenarios help management understand risk in business terms and make treatment decisions more consistent and actionable.

Q56. An AI use case has unacceptable safety risk that cannot be reduced to within organizational tolerance using feasible controls. Which treatment strategy is MOST appropriate?

1. Risk avoidance by not proceeding with or discontinuing the use case
   2. Risk acceptance regardless of tolerance
   3. Removing the risk from the register
   4. Reporting the risk less frequently

Correct Answer: 2. Risk avoidance by not proceeding with or discontinuing the use case

Explanation: Risk avoidance is appropriate when the organization determines that an activity creates unacceptable exposure and practical controls cannot reduce it to within tolerance. Avoidance may involve not deploying the AI solution, removing a particular autonomous function, or discontinuing an existing use case. Risk acceptance should occur only when an authorized risk owner determines that residual exposure is acceptable under governance requirements. Removing the risk from documentation does not change the actual exposure. AI innovation should not override established safety thresholds or enterprise risk appetite when the potential consequences remain unacceptable.

Q57. A required AI control cannot be implemented immediately because of a technical limitation. What is the BEST interim approach?

1. Ignore the risk until the preferred control becomes available
   2. Declare the risk eliminated
   3. Implement suitable compensating controls and reassess residual risk
   4. Remove the requirement from all documentation

Correct Answer: 3. Implement suitable compensating controls and reassess residual risk

Explanation: Compensating controls can provide alternative risk reduction when the preferred control cannot be implemented immediately. They should address the same underlying risk as effectively as practical and should be documented, monitored, and reviewed. For example, stronger human approval and monitoring may temporarily compensate for an unavailable automated restriction. Management must then reassess the residual risk to determine whether it remains within tolerance. Compensating controls should not become permanent by default; remediation plans and review dates help ensure the organization eventually implements the intended solution where appropriate.

Q58. A model’s approved accuracy threshold is 95%. Which indicator provides the BEST early warning that model performance risk is increasing?

1. A trend showing accuracy declining toward the 95% threshold
   2. Number of chairs in the data science office
   3. Total model documentation pages
   4. Number of meetings scheduled this month

Correct Answer: 1. A trend showing accuracy declining toward the 95% threshold

Explanation: A useful key risk indicator provides early warning before an exposure breaches an established limit. Monitoring a declining accuracy trend allows management to investigate model drift, data changes, system integration problems, or other causes before performance falls below the approved threshold. A single breach is important, but trend information can support proactive intervention. Thresholds should be linked to escalation procedures, owners, and treatment options. Metrics unrelated to the risk—such as meeting counts or office resources—do not help management understand whether model performance is approaching an unacceptable level.

Q59. A third-party AI solution depends on several external model, data, and infrastructure providers. What information would MOST improve supply-chain risk visibility?

1. Only the name of the direct vendor
   2. Only the direct vendor’s annual revenue
   3. A mapping of material upstream dependencies and critical subcontractors
   4. The vendor’s advertising strategy

Correct Answer: 4. A mapping of material upstream dependencies and critical subcontractors

Explanation: AI supply chains can contain multiple layers, including cloud services, foundation models, datasets, open-source components, inference providers, and subcontractors. Understanding material upstream dependencies helps identify concentration, availability, legal, privacy, security, and geopolitical risks that may not be visible when evaluating only the direct supplier. Contracts may need to address subcontractor notification and material changes. The organization does not necessarily need exhaustive information about every minor component, but critical dependencies should be understood well enough to assess resilience and determine whether supplier failures could affect important business services.

Q60. After an AI incident is contained and services are restored, what activity provides the MOST value for strengthening the AI risk program?

1. Delete all incident records
   2. Conduct a lessons-learned review and update controls, scenarios, procedures, and training as needed
   3. Avoid discussing the incident with risk owners
   4. Assume the same incident cannot happen again

Correct Answer: 2. Conduct a lessons-learned review and update controls, scenarios, procedures, and training as needed

Explanation: Post-incident review converts experience into risk-program improvement. The organization should evaluate root causes, control performance, detection effectiveness, escalation, communications, recovery, decision authority, and any unexpected dependencies. Findings can lead to updated risk scenarios, controls, monitoring thresholds, response playbooks, business continuity plans, vendor requirements, and staff training. Evidence should be retained according to applicable requirements. Treating restoration as the end of the incident misses an important opportunity to strengthen resilience. AAIR emphasizes integration of AI risk into incident response, BIA, business continuity, and disaster recovery processes.