I85PREP
⌘ K
Lead AI Risk Manager study materials
Lead AI Risk Manager study resources

Lead AI Risk Manager practice questions

One question at a time. Build your Lead AI Risk Manager understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI Risk Management

In an enterprise AI risk management framework, which artifact records the acceptable level of harm or loss an organization is willing to tolerate for a specific AI use case?

  1. Data lineage record
  2. Risk appetite statement Correct answer.
  3. Model card
  4. Incident postmortem

Explanation

A risk appetite (or risk tolerance) statement defines the amount and type of risk an organization is willing to accept in pursuit of its objectives, and is set per use case or domain. Model cards document model behavior and intended use, lineage records trace data, and postmortems analyze past incidents; none of these establish a tolerance threshold.

Question 2AI Risk Management

A team is classifying an AI system that autonomously approves consumer loan applications with no human review. Under most AI risk tiering schemes, which factor most strongly drives the system's risk classification?

  1. The number of engineers who reviewed the code
  2. The size of the training dataset in terabytes
  3. The combination of high-impact decisions and absence of human oversight Correct answer.
  4. The programming language used to build the model

Explanation

Risk tiering typically weighs the severity of potential harm (high-impact domain such as credit) together with the degree of autonomy and human oversight. A system making consequential decisions without human review concentrates both factors, raising its tier. Dataset size, language choice, and code review headcount are not the primary drivers of harm severity in tiering schemes.

Question 3AI Risk Management

Which practice best supports the traceability principle in AI risk management across a model's lifecycle?

  1. Maintaining versioned records of data, code, configuration, and approvals Correct answer.
  2. Relying on individual engineers' memory of design decisions
  3. Storing only the final production model weights
  4. Deleting intermediate experiment logs after deployment

Explanation

Traceability requires reconstructing how a model was built, tested, and approved, which demands versioned records spanning data, code, configuration, and sign-offs. Retaining only final weights or deleting experiment logs breaks the audit chain, and relying on memory is not a verifiable record.

Question 4AI Risk Management

A risk manager is building a monitoring plan for a deployed credit-scoring model. Which metric set most directly addresses model degradation risk over time?

  1. Feature drift, prediction distribution shift, and subgroup performance trends Correct answer.
  2. Marketing conversion rate of the lending website
  3. Number of pull requests merged per sprint
  4. Server CPU utilization and disk I/O

Explanation

Model degradation is detected through drift in input features, shifts in the prediction distribution, and changes in performance across subgroups. These signals reveal when the model's learned relationships no longer match reality. Infrastructure and engineering metrics describe operational health, not model validity.

Question 5AI Risk Management

Which statement correctly distinguishes an AI risk from an AI incident?

  1. A risk is a realized harm; an incident is a forecast of possible harm.
  2. Risks apply only to generative models; incidents apply only to classical models.
  3. Risks and incidents are synonyms in AI governance terminology.
  4. A risk is a potential future harm; an incident is a realized event that caused or could have caused harm. Correct answer.

Explanation

Risk refers to the possibility of harm weighted by likelihood and severity, while an incident is an event where harm or a near miss actually occurred. Incident management feeds back into risk assessment as evidence. The two are distinct stages in the risk lifecycle, not synonyms.

Question 6AI Risk Management

A vendor supplies a third-party foundation model that a bank fine-tunes for customer service. Which control most directly addresses the bank's residual risk from the vendor's upstream model behavior?

  1. Assuming the vendor's public benchmark scores transfer unchanged to the bank's use case
  2. Removing the bank's own content filters to reduce latency
  3. Outsourcing all customer-facing decisions to the vendor's default configuration
  4. Requiring the vendor to disclose training data provenance and known limitations, then testing the fine-tuned model against bank-specific evaluation suites Correct answer.

Explanation

Third-party model risk is mitigated by obtaining transparency about provenance and limitations and by independently evaluating the model in the deployer's own context. Vendor benchmarks do not automatically transfer, and the deployer retains responsibility for downstream behavior. Removing safeguards increases rather than reduces residual risk.

Question 7AI Risk Management

In a risk register entry for an AI hiring-screening tool, which field is essential for prioritizing remediation work?

  1. The color scheme of the dashboard
  2. The inherent risk rating combining likelihood and severity of harm Correct answer.
  3. The name of the model's original author
  4. The number of training epochs used

Explanation

Prioritization depends on a risk rating that combines how likely a harm is with how severe it would be, often adjusted for existing controls. This score lets leadership rank and fund remediation. Author names, UI styling, and epoch counts do not convey risk magnitude.

Question 8AI Risk Management

A red team finds that a customer-facing chatbot can be prompted to reveal its system instructions. Which risk treatment is most appropriate first?

  1. Apply mitigations such as instruction-hierarchy defenses and output filtering, then retest Correct answer.
  2. Eliminate the product line immediately without further assessment
  3. Transfer the risk entirely by adding a disclaimer to the website
  4. Accept the risk because system prompts contain no personal data

Explanation

Prompt-extraction findings are typically treated by strengthening defenses, such as instruction hierarchy and output filtering, followed by retesting to confirm the fix. Acceptance requires a documented rationale tied to impact, and blanket elimination is disproportionate before assessing severity. Transfer via disclaimer does not remove the technical exposure.

Question 9AI Risk Management

Which is the primary purpose of a human-in-the-loop control in a high-risk AI workflow?

  1. To replace the need for monitoring after deployment
  2. To satisfy open-source licensing requirements
  3. To increase model throughput
  4. To ensure a qualified person can review, override, or halt consequential AI outputs Correct answer.

Explanation

Human-in-the-loop controls place a competent reviewer in the decision path so that consequential outputs can be checked, overridden, or stopped. This reduces the risk of automated errors causing harm. It is not a performance optimization, a substitute for ongoing monitoring, or a licensing mechanism.

Question 10AI Risk Management

A risk manager must estimate the likelihood of a generative model producing defamatory statements about public figures. Which approach yields the most defensible estimate?

  1. Asking a single engineer for a gut-feel percentage
  2. Sampling adversarial and natural prompts, measuring harmful-output rates, and extrapolating with documented uncertainty Correct answer.
  3. Assuming zero likelihood because outputs are probabilistic
  4. Relying on the vendor's claim that the model is safe

Explanation

Likelihood estimates should rest on structured testing that quantifies harmful-output rates across realistic and adversarial prompts, with uncertainty stated explicitly. Vendor assurances and single opinions are not evidence, and assuming zero likelihood ignores known failure modes of generative models.

Question 11AI Risk Management

Which element is central to an AI system's accountability framework?

  1. A named owner responsible for the system's risks and decisions Correct answer.
  2. An anonymous escalation channel with no defined responders
  3. A policy that assigns responsibility to 'the organization' collectively
  4. A requirement that no one may question model outputs

Explanation

Accountability requires identifiable owners who are answerable for a system's risk decisions and outcomes. Diffuse or anonymous responsibility undermines enforcement, and prohibiting challenge removes the checks that accountability depends on.

Question 12AI Risk Management

When assessing bias risk in a resume-screening model, which analysis provides the strongest evidence of disparate impact?

  1. Overall accuracy on the full applicant pool
  2. Selection rates across protected groups relative to a reference group, with confidence intervals Correct answer.
  3. The model's training loss curve
  4. The average resume length in the dataset

Explanation

Disparate impact is assessed by comparing selection or favorable-outcome rates between protected groups and a reference group, with statistical uncertainty reported. Aggregate accuracy can mask subgroup disparities, and training diagnostics or resume length do not measure impact on groups.

Question 13AI Risk Management

A risk manager documents that an AI system's residual risk exceeds the organization's tolerance even after controls. What is the most appropriate governance action?

  1. Escalate the decision to the accountable executive for an explicit go/no-go determination Correct answer.
  2. Silently adjust the risk rating downward to fit the tolerance
  3. Remove the risk from the register to avoid scrutiny
  4. Deploy the system and revisit the issue after launch

Explanation

When residual risk exceeds tolerance, the decision to proceed, pause, or reject must be made and recorded at the appropriate authority level. This preserves accountability and ensures the exception is deliberate. Manipulating ratings or hiding the risk defeats the register's purpose.

Question 14AI Risk Management

Which describes the correct relationship between an AI impact assessment and a data protection impact assessment?

  1. They are identical and one always replaces the other
  2. An AI impact assessment is broader in scope and may incorporate or reference a data protection impact assessment where personal data is processed Correct answer.
  3. A data protection impact assessment always supersedes any AI-specific assessment
  4. Neither applies to systems that process only synthetic data

Explanation

An AI impact assessment addresses a wide range of harms including safety, fairness, and autonomy, while a data protection impact assessment focuses on privacy risks to personal data. Where both apply, the AI assessment can reference the privacy analysis rather than duplicate it. They are complementary, not interchangeable.

Question 15AI Risk Management

A monitoring dashboard shows a sudden drop in a fraud model's precision alongside stable input distributions. Which explanation should the risk manager investigate first?

  1. Concept drift, where the relationship between features and the fraud label has changed Correct answer.
  2. A change in the office floor plan
  3. A hardware failure in the dashboard's display
  4. An increase in the number of engineers on the team

Explanation

When input distributions are stable but predictive performance drops, the likely cause is concept drift: the mapping from features to outcomes has shifted, for example because fraud tactics changed. This is distinct from data drift, which would show up as changed inputs. The other options are unrelated to model behavior.

Question 16AI Risk Management

Which control most effectively limits the risk of sensitive data leaking through a model's memorized training content?

  1. Disabling all logging of model outputs
  2. Publishing the training dataset openly for transparency
  3. Increasing the model's parameter count
  4. Applying data minimization, deduplication, and privacy-enhancing techniques during training, plus output monitoring Correct answer.

Explanation

Memorization risk is reduced by limiting and deduplicating sensitive training data and by privacy-enhancing techniques, with output monitoring to catch leakage. Larger models can memorize more, publishing sensitive data increases exposure, and disabling logging removes detection capability.

Question 17AI Risk Management

In AI risk governance, what is the function of a model inventory?

  1. To track which models exist, their owners, purposes, and risk classifications so they can be governed consistently Correct answer.
  2. To replace the need for periodic risk assessments
  3. To store the source code of every application in the company
  4. To measure employee satisfaction with AI tools

Explanation

A model inventory is a register of deployed and in-development models with metadata such as ownership, purpose, and risk tier, enabling consistent oversight and periodic review. It is not a code repository, does not substitute for assessments, and is unrelated to employee sentiment.

Question 18AI Risk Management

A team proposes using an LLM to summarize employee performance reviews that will inform promotion decisions. Which risk consideration is most salient for the risk manager?

  1. The summarizer will certainly reduce reading time
  2. The summarizer may introduce or amplify biased characterizations that affect consequential employment outcomes Correct answer.
  3. The font used in the generated summary
  4. The model's token cost per request

Explanation

Because summaries feed into promotion decisions, distortions or biased framing can cause unfair employment outcomes, which is a high-severity risk. Cost and formatting are secondary operational concerns. Efficiency gains do not offset the need to control decision quality.

Question 19AI Risk Management

Which metric best captures whether an AI system's explanations are useful to human reviewers making decisions?

  1. The number of features displayed in the interface
  2. The model's inference latency in milliseconds
  3. Agreement between reviewer decisions aided by explanations and ground-truth outcomes, plus reviewer comprehension measures Correct answer.
  4. The count of API calls per day

Explanation

Explanation quality is judged by whether it improves human understanding and decision accuracy, measured through comprehension studies and decision-agreement with ground truth. Feature counts, latency, and call volume describe interface or operational properties rather than explanatory value.

Question 20AI Risk Management

What is the primary role of an AI incident response plan?

  1. To replace the organization's overall business continuity plan
  2. To guarantee that no AI incident will ever occur
  3. To define detection, escalation, containment, and communication steps when an AI system causes or nearly causes harm Correct answer.
  4. To document the marketing strategy for AI products

Explanation

An incident response plan specifies how incidents are detected, escalated, contained, and communicated so that harm is limited and lessons are captured. It cannot prevent all incidents, and it complements rather than replaces broader continuity planning. It is not a marketing artifact.

Question 21AI Risk Management

A regulator asks how the organization ensures that an AI system remains within approved operating conditions after deployment. Which evidence is most relevant?

  1. A one-time pre-deployment test report with no follow-up
  2. The model's original design sketch
  3. A list of team members' academic degrees
  4. Continuous monitoring records, alert thresholds, and documented actions taken when thresholds were breached Correct answer.

Explanation

Demonstrating ongoing compliance requires evidence of continuous monitoring against defined thresholds and records of interventions when those thresholds were crossed. A one-time report only shows the state at launch. Design sketches and credentials do not evidence post-deployment control.

Question 22AI Risk Management

Which situation best illustrates risk transfer in AI risk management?

  1. Retraining a model on fresher data
  2. Adding a second reviewer to a high-stakes decision
  3. Purchasing cyber-liability insurance that covers harms caused by AI-driven systems Correct answer.
  4. Deciding to stop using an AI feature entirely

Explanation

Risk transfer shifts the financial consequences of a risk to a third party, as with insurance. Adding reviewers is mitigation, retraining addresses model quality, and discontinuing a feature is risk avoidance. Each is a distinct treatment option.

Question 23AI Risk Management

A risk manager is told that a model's evaluation used a benchmark that overlaps with its training data. What is the primary concern?

  1. The evaluation will violate open-source licensing
  2. The benchmark will run too slowly
  3. The evaluation results may be inflated and overstate real-world performance Correct answer.
  4. The model will require more storage

Explanation

Benchmark contamination occurs when evaluation items appear in training data, causing memorized answers to inflate scores. This overstates capability and understates deployment risk. Speed, storage, and licensing are separate concerns that do not follow from contamination.

Question 24AI Risk Management

Which is the most appropriate way to set risk thresholds for an AI system used in medical triage support?

  1. Derive thresholds from the severity of potential patient harm and the clinical evidence, with clinician input Correct answer.
  2. Set thresholds to whatever the model achieves by default
  3. Choose thresholds based on the vendor's pricing tier
  4. Reuse thresholds from an unrelated marketing model

Explanation

Thresholds for clinical decision support must reflect the severity of harm to patients and be grounded in clinical evidence with practitioner involvement. Copying thresholds from an unrelated domain or accepting defaults ignores context-specific risk. Pricing tiers have no bearing on safety limits.

Question 25AI Risk Management

An organization wants to reduce automation bias among staff who review AI recommendations. Which intervention most directly targets this risk?

  1. Increasing the model's confidence score display size
  2. Requiring reviewers to record independent judgments before seeing the AI output and training them on over-reliance Correct answer.
  3. Removing all documentation of reviewer decisions
  4. Making the AI recommendation more visually prominent

Explanation

Automation bias is the tendency to over-trust automated suggestions. Requiring an independent judgment before exposure to the AI output, combined with training on over-reliance, directly counteracts anchoring on the recommendation. Emphasizing the AI output or its confidence score increases deference rather than reducing it.

Question 26AI Risk Management

Which document defines the acceptable level of risk an organization is willing to bear for a specific AI use case before escalation is required?

  1. The model card
  2. The AI risk appetite statement Correct answer.
  3. The data processing agreement
  4. The system architecture diagram

Explanation

A risk appetite statement sets the boundaries of acceptable residual risk for AI use cases and drives escalation and approval decisions. Model cards and architecture diagrams describe the system rather than the organization's tolerance for its risk.

Question 27AI Risk Management

An AI system classified as high-risk under a regulatory framework is about to be deployed. Which artifact is most important to complete before go-live?

  1. A completed conformity assessment with documented evidence Correct answer.
  2. A list of all open-source libraries used
  3. An estimate of future cloud costs
  4. A marketing plan for the model's capabilities

Explanation

High-risk classification typically triggers an obligation to perform a conformity assessment and retain the supporting evidence before deployment. The other artifacts may be useful operationally but do not satisfy the pre-deployment assurance obligation.

Question 28AI Risk Management

A risk manager needs to record how a specific AI risk will be handled. Which set of treatment categories is standard in AI risk management?

  1. Identify, quantify, monitor, retire
  2. Avoid, mitigate, transfer, accept Correct answer.
  3. Design, build, test, deploy
  4. Detect, contain, recover, report

Explanation

Risk treatment is conventionally expressed as avoiding, mitigating, transferring, or accepting the risk. The other lists describe lifecycle phases or incident handling activities rather than treatment choices.

Question 29AI Risk Management

Which role is accountable for ensuring that AI risks are escalated to the appropriate governance body?

  1. The cloud infrastructure provider
  2. The data labeling vendor
  3. The AI risk owner designated for the use case Correct answer.
  4. The end user of the AI system

Explanation

Each AI use case should have a named risk owner accountable for monitoring and escalation. Vendors, end users, and infrastructure providers contribute information but do not own the escalation duty.

Question 30AI Risk Management

A risk manager is building a taxonomy for AI-specific risks. Which item is an AI-specific risk rather than a conventional IT risk?

  1. Expiration of a TLS certificate
  2. Unauthorized physical access to a server room
  3. Harmful output drift caused by distribution shift in model inputs Correct answer.
  4. Server hardware failure in the primary data center

Explanation

Distribution shift degrading model behavior is a risk arising from learned model behavior and is specific to AI systems. Hardware failure, certificate expiry, and physical intrusion are conventional IT and security risks.

Question 31AI Risk Management

What is the main purpose of documenting a model's intended use and out-of-scope uses in its risk documentation?

  1. To reduce the amount of training data required
  2. To guarantee the model will not fail in production
  3. To bound the context in which the model's risk assessment remains valid Correct answer.
  4. To satisfy marketing requirements for product positioning

Explanation

Intended-use documentation defines the boundary within which the risk assessment and validation evidence apply. Using the model outside that boundary invalidates assumptions about its performance and harms.

Question 32AI Risk Management

Which activity belongs to the monitoring phase rather than the design phase of an AI system's risk lifecycle?

  1. Tracking production error rates against predefined thresholds Correct answer.
  2. Selecting the model architecture
  3. Drafting the initial risk assessment
  4. Defining the business objective

Explanation

Monitoring concerns observing a deployed system against thresholds and triggers. Architecture selection, objective definition, and initial assessment occur during design and development.

Question 33AI Risk Management

A risk manager must ensure that an AI system's risk assessment is refreshed when material changes occur. Which event most clearly triggers a reassessment?

  1. Renaming the deployment environment
  2. Updating the internal wiki page describing the system
  3. Retraining the model on a new data source that changes its population coverage Correct answer.
  4. Rotating a service account password

Explanation

A change in training data that alters the population the model serves can change both performance and harm profiles, so the assessment must be revisited. Administrative changes such as passwords, wiki edits, or environment names do not alter model risk.

Question 34AI Risk Management

A risk manager learns that a vendor's model update silently changed the model behind a stable API version. Which risk is most directly realized?

  1. Increased cloud storage cost
  2. Shortened data retention period
  3. Loss of version control over the assessed system, invalidating prior assurance Correct answer.
  4. Reduced network latency

Explanation

If the underlying model changes without notice, the evidence gathered for the assessed version may no longer apply, which undermines assurance. Cost, latency, and retention are unrelated to silent model substitution.

Question 35AI Risk Management

When scoring AI risks on a likelihood-by-impact matrix, which practice improves consistency across assessors?

  1. Allowing each assessor to choose their own scoring scale
  2. Averaging scores from unrelated use cases
  3. Publishing calibrated scale definitions with concrete examples for each level Correct answer.
  4. Scoring only impact and ignoring likelihood

Explanation

Calibrated scale definitions with examples reduce subjective variation between assessors and make scores comparable. Letting assessors invent scales or ignoring a dimension destroys comparability.

Question 36AI Risk Management

A generative AI feature is planned for a public-facing product, and the team proposes accepting the risk without any mitigation. Which precondition is most necessary for that acceptance to be defensible?

  1. The team has completed the product backlog
  2. The residual risk is documented and formally approved by an accountable authority within appetite Correct answer.
  3. The launch date has already been announced
  4. The model has the largest possible parameter count

Explanation

Accepting risk is defensible only when the residual level is documented and approved by someone with authority, within the stated appetite. Backlog status, model size, and marketing commitments are irrelevant to the validity of acceptance.

Question 37AI Risk Management

A risk manager is quantifying the potential severity of an AI system that assists with credit limit decisions. Which factor most appropriately anchors the impact rating?

  1. The size of the engineering team
  2. The programming language used for serving
  3. The magnitude of financial harm and number of affected customers if the model errs systematically Correct answer.
  4. The number of features in the model

Explanation

Impact ratings should reflect the scale of harm to affected parties, such as financial loss and the number of people affected. Technical implementation details and team size do not measure harm severity.

Question 38AI Risk Management

An AI system's risk register entry lists a control that has never been tested. How should the risk manager treat that control when estimating residual risk?

  1. As unnecessary and removed from the register
  2. As unverified, so residual risk should not be reduced until effectiveness is evidenced Correct answer.
  3. As fully effective because it is documented
  4. As transferred to the vendor automatically

Explanation

Residual risk estimates should only credit controls whose effectiveness has been demonstrated. Documenting a control does not prove it works, and untested controls cannot justify reducing the residual estimate.

Question 39AI Risk Management

Which approach best aligns AI risk reporting with an organization's enterprise risk management framework?

  1. Mapping AI risks to the enterprise risk taxonomy and reporting through existing governance channels Correct answer.
  2. Reporting AI risks only after an incident occurs
  3. Restricting AI risk reporting to the engineering team
  4. Creating a separate parallel risk universe with no linkage to enterprise categories

Explanation

Mapping AI risks into the enterprise taxonomy lets leadership compare and aggregate them alongside other risks. A parallel universe or incident-only reporting fragments oversight and hides exposure.

Question 40AI Risk Management

A risk manager is assessing a third-party AI service and needs to understand the vendor's own risk controls. Which request is most directly relevant?

  1. The vendor's executive travel schedule
  2. The vendor's independent audit reports and control attestations covering the model lifecycle Correct answer.
  3. The vendor's social media follower count
  4. The vendor's office floor plan

Explanation

Independent audit reports and attestations provide evidence about the vendor's controls across the model lifecycle. The other items carry no assurance value about risk management.

Question 41AI Risk Management

An AI system's risk assessment identifies a potential harm that is severe but highly uncertain. Which treatment approach is most appropriate while evidence is gathered?

  1. Ignore the harm until it is observed in production
  2. Close the risk because it is uncertain
  3. Apply interim controls and set a review trigger tied to the evidence gap Correct answer.
  4. Accept the risk permanently without further action

Explanation

When severity is high but evidence is incomplete, interim controls plus a defined review trigger manage exposure while uncertainty is reduced. Permanent acceptance or closure misrepresents an unresolved risk.

Question 42AI Risk Management

Which practice best supports consistent escalation of AI risks across business units?

  1. Defining quantitative escalation thresholds tied to risk appetite and severity Correct answer.
  2. Letting each unit escalate based on informal judgment alone
  3. Escalating only risks raised by external parties
  4. Escalating every risk regardless of severity

Explanation

Quantitative thresholds linked to appetite and severity make escalation predictable and comparable across units. Purely informal or indiscriminate escalation produces inconsistency and alert fatigue.

Question 43AI Risk Management

A risk manager reviews a model's evaluation and notes that performance was measured only on data from the same region as training. What is the most important limitation to record?

  1. Limited evidence of generalization to other populations the system will serve Correct answer.
  2. The evaluation was too expensive to be useful
  3. The model has too few parameters to be assessed
  4. The model is guaranteed to fail in all regions

Explanation

When evaluation data mirrors training conditions, the evidence says little about performance for other populations. The correct response is to record the generalization gap as a limitation, not to overstate failure or dismiss the evaluation.

Question 44AI Risk Management

A regulator asks how the organization determines which AI systems require enhanced governance. Which criterion is most defensible?

  1. Whether the system is written in a popular language
  2. The seniority of the requesting manager
  3. Whether the system was built internally or externally
  4. A documented classification based on use context, affected populations, and potential severity of harm Correct answer.

Explanation

Risk-based classification should rest on the use context, who is affected, and the severity of potential harm. Build origin, language, and requester seniority are not valid proxies for risk.

Question 45AI Risk Management

A risk manager must decide whether to approve a model whose subgroup performance gap exceeds policy limits but whose aggregate accuracy is strong. Which reasoning is most sound?

  1. Strong aggregate accuracy overrides subgroup gaps by definition
  2. Subgroup gaps are irrelevant if the model is accurate overall
  3. Aggregate accuracy can mask subgroup harm, so the gap should block approval until mitigated or explicitly accepted by an authorized body Correct answer.
  4. The gap should be ignored because it was measured after deployment

Explanation

Aggregate metrics can conceal concentrated harm in subgroups, so a gap beyond policy limits requires mitigation or explicit, authorized acceptance. Aggregate performance alone cannot justify approval.

Question 46AI Risk Management

An organization wants to rely on a vendor's safety evaluation for a high-risk AI system. Which condition most strengthens that reliance?

  1. The evaluation was performed once before general release
  2. The evaluation methodology, test data, and limitations are disclosed and independently reproducible Correct answer.
  3. The vendor summarizes results in a marketing brochure
  4. The vendor states the model is safe without detail

Explanation

Reliance on third-party evaluation is strongest when methodology, data, and limitations are transparent and the results can be independently reproduced. Marketing summaries and unqualified safety claims provide no verifiable assurance.

Question 47AI Risk Management

A risk manager observes that a model's risk profile has shifted because user behavior changed after a product redesign. Which response best reflects continuous risk management?

  1. Retire the model immediately without analysis
  2. Reassess the risk using post-change data and update controls and thresholds accordingly Correct answer.
  3. Wait for the next annual review cycle
  4. Keep the original assessment because the model code did not change

Explanation

Risk is a function of the system in its operating context, so a change in user behavior warrants reassessment even when the code is unchanged. Freezing the old assessment or waiting a year leaves the exposure unmanaged.

Question 48AI Risk Management

Two business units want to deploy the same foundation model for different purposes with different harm profiles. How should the risk manager structure the assessment?

  1. Assess each use case separately while sharing common model-level evidence Correct answer.
  2. Apply one assessment because the underlying model is identical
  3. Defer assessment until both systems are in production
  4. Assess only the use case with the larger budget

Explanation

Risk depends on use context and affected populations, so each use case needs its own assessment even when the model is shared. Common model-level evidence can be reused to avoid duplication.

Question 49AI Risk Management

A risk manager is designing key risk indicators for an AI system. Which indicator best provides early warning of emerging model risk rather than lagging harm?

  1. The number of complaints resolved last year
  2. The count of incidents already reported to regulators
  3. The total number of model deployments completed
  4. A rising trend in the rate of low-confidence outputs flagged for human review Correct answer.

Explanation

Early-warning indicators track precursor signals, such as growing uncertainty in model outputs, before harm materializes. Incident counts, resolved complaints, and deployment totals are lagging or unrelated measures.

Question 50AI Risk Management

An AI risk manager must present residual risk to an executive committee that includes non-technical members. Which presentation approach is most effective?

  1. Present raw model logs and hyperparameters without interpretation
  2. State that all risks are technical and require no decision
  3. Report only the number of open tickets
  4. Express risks in business impact terms with clear likelihood, severity, and decision options Correct answer.

Explanation

Executive audiences need risks translated into business impact with clear likelihood, severity, and options for decision. Raw technical artifacts or ticket counts do not support informed governance decisions.