AIGP Practice Exams v4 · Review deck · 10 scenarios · 50 questions
Replay the Scenarios
The v4 scenario bank re-presented as a review deck: scenario, then each question with the correct answer highlighted, the rationale, and the BoK tag. Read the scenario, answer in your head, then check.
10 scenarios50 questionsAnswers highlightedRationales + BoK tags
How to use: read the scenario, answer each question mentally, then reveal the highlighted answer and read the rationale. The per-domain breakdown below shows where the bank concentrates.
Scroll to begin ↓
1
Scenario 1 · 5 questions · Domain I
Scenario 1: NeuroForge Foundry
NeuroForge Foundry, a Singapore startup, is building a content-moderation API that classifies text as 'safe', 'borderline', or 'unsafe'. The team uses a large language model as a foundation and fine-tunes it on 500,000 labelled social media posts. Their data scientist, Priya, must explain key AI concepts to the non-technical founding team so they can make governance decisions about whether to build, buy, or partner — and how to validate the model before launch.
Q1Domain I · Competency A
Priya tells the team the model 'learns patterns from data'. Which type of learning is this, and how does it differ from rules-based programming?
A. Both are the same: the model is given explicit rules by the programmer
B. Rules-based programming: humans write explicit if-then rules; machine learning: the system infers rules from labelled examples without being explicitly programmed with rules ✓
C. Machine learning means the model memorises every example; rules-based means it generalises
D. Rules-based systems learn from data; machine learning systems use fixed logic
RATIONALEThe core distinction: rules-based systems apply human-written logic, while ML systems infer patterns from data. This is the foundational concept the exam tests under Domain I — understanding what ML is and how it differs from traditional programming. Body of Knowledge Domain I, Competency A
Q2Domain I · Competency B
The team debates whether to use the foundation model as-is or fine-tune it. What is the key difference between a foundation model and a fine-tuned model?
A. A foundation model is trained from scratch on the company's data; a fine-tuned model is pre-trained
B. Fine-tuning makes the model larger; foundation models are always smaller
C. Foundation models cannot be used for text classification
D. A foundation model is pre-trained on broad data; fine-tuning adapts it to a specific task using smaller, task-specific data ✓
RATIONALEFoundation models are pre-trained on massive general datasets; fine-tuning adapts them to specific tasks with smaller, domain-specific data. The adaptation spectrum (prompt engineering → fine-tuning → continued pre-training) is a core BoK v2.1 topic under Domain I. Body of Knowledge Domain I, Competency B
Q3Domain I · Competency B
After training, the model achieves 99% accuracy on the training data but only 72% on new, unseen data. What is this phenomenon called, and what does it indicate?
A. Underfitting — the model is too simple for the task
B. Overfitting — the model has memorised the training data but does not generalise to new data ✓
C. Data drift — the new data has changed distribution
D. Concept drift — the definition of 'unsafe' has changed
RATIONALEThe gap between high training accuracy and lower test accuracy is the classic overfitting signal — the model memorised the training set instead of learning generalisable patterns. The exam tests this distinction and the mitigation (regularisation, more data, simpler models, cross-validation). Body of Knowledge Domain I, Competency B
Q4Domain I · Competency C
The team wants to understand why the model flags certain posts. The model is a deep neural network whose internal decision process is not transparent. What concept describes this, and what is the practical governance implication?
A. The model is 'interpretable' — its decisions are transparent by design
B. The model is 'deterministic' — every input always produces the same output
C. The model is a 'black box' — its internal reasoning is opaque, creating a governance challenge for accountability and contestability ✓
D. The model has 'concept drift' — its logic changes over time
RATIONALEDeep learning models are black boxes: their internal representations are not human-interpretable. This creates governance challenges — you cannot easily explain why a decision was made, which affects accountability, contestability, and the ability to audit. The exam tests the black-box problem and interpretability vs explainability. Body of Knowledge Domain I, Competency C
Q5Domain I · Competency C
Priya recommends using SHAP values to explain individual moderation decisions. What is SHAP, and what governance principle does it support?
A. SHAP is a training method that makes models more accurate
B. SHAP is a privacy-enhancing technology that anonymises training data
C. SHAP is a post-hoc explainability technique that attributes each prediction to its contributing features, supporting transparency and contestability ✓
D. SHAP is a regulatory framework required by the EU AI Act
RATIONALESHAP (SHapley Additive exPlanations) is a post-hoc explainability tool that shows which features drove a specific prediction. It supports the transparency and contestability principles by making black-box decisions auditable. The exam tests post-hoc explainability techniques under Domain I. Body of Knowledge Domain I, Competency C
Exam tip for this scenarioFoundations tell: the question tests whether you can name the ML paradigm from the scenario facts. Match the mechanism, not the buzzword.
2
Scenario 2 · 5 questions · Domain I & III
Scenario 2: DataTrust Labelling
DataTrust, a German AI company, is building a medical imaging model that detects diabetic retinopathy from eye scans. The model must be trained on labelled images, but the company's data scientists disagree on how to source, label, and validate the training data. The company's medical director, Dr. Weber, insists on clinical accuracy, while the data team is under pressure to launch quickly. They must decide on data sourcing, labelling protocols, bias testing, and validation before the model can be deployed in clinics.
Q6Domain III · Competency B
DataTrust collects eye scans from three hospitals in Munich, all serving predominantly white European populations. What is the primary data-governance risk?
A. The training data is not representative of the broader population, risking poor performance on patients of other ethnicities ✓
B. The scans are too high-resolution for the model to process
C. The hospitals charge too much for the data
D. The scans are not personal data under the GDPR
RATIONALERepresentativeness is a core data-quality dimension. Training on a single demographic group creates a model that may not generalise to others — a classic source of bias in medical AI. Data governance for AI requires assessing representativeness, not just volume. Body of Knowledge Domain III, Competency B
Q7Domain III · Competency B
The data team wants to label scans quickly using a crowdsourcing platform with non-expert labelers. Dr. Weber disagrees. Who is correct, and why?
A. Dr. Weber — medical image labelling requires domain expertise because incorrect labels directly affect model accuracy and patient safety ✓
B. The data team — crowdsourcing is faster and cheaper, which is always better
C. Neither — labels do not affect model performance
D. Both — the model can learn regardless of label quality
RATIONALELabel quality is critical for supervised learning. Non-expert labels for medical images introduce noise that degrades the model, and in a safety-critical domain, incorrect labels can cause patient harm. Domain expertise in labelling is a data-governance requirement for high-risk AI. Body of Knowledge Domain III, Competency B
Q8Domain III · Competency C
The model performs well on images from the training hospitals but poorly on images from a hospital in Berlin that serves a more diverse population. What is this phenomenon, and what is the appropriate response?
A. This is a hardware problem with the Berlin hospital's scanner
B. This is acceptable — the model only needs to work in Munich
C. This is overfitting to the Munich population — the model should be retrained on more representative data before deployment in Berlin ✓
D. This is data drift — the Berlin hospital's patients are different
RATIONALEPoor performance on a different population is a generalisation failure caused by non-representative training data. The correct response is to retrain with more representative data before deployment, not to restrict the model to the original hospitals or blame hardware. Body of Knowledge Domain III, Competency C
Q9Domain III · Competency C
Before deployment, the team must validate the model. Which validation approach is most appropriate for a medical AI system?
A. Test on a representative, independent dataset that includes diverse populations and different scanner types, with pre-defined performance thresholds ✓
B. Test on a small holdout set from the same Munich hospitals
C. Skip validation and deploy, relying on post-deployment monitoring
D. Test only on the training data to confirm it memorised correctly
RATIONALEValidation for high-stakes AI requires an independent, representative test set that reflects real deployment conditions — different populations, scanners, and sites — with pre-defined thresholds for acceptable performance. Testing on same-source data or skipping validation are governance failures. Body of Knowledge Domain III, Competency C
Q10Domain II · Competency A
The eye scans are health data. Under the GDPR, what is the correct classification of this data, and what does it trigger?
A. Ordinary personal data — standard consent is sufficient
B. Special category data (health) under Art. 9 — requires a specific Art. 9 condition and a DPIA under Art. 35 ✓
C. Anonymous data — no GDPR obligations apply
D. Non-personal data — because the scans are images, not text
RATIONALEMedical images are special category health data under GDPR Art. 9, which requires a specific processing condition (e.g., explicit consent or healthcare provision) and triggers a mandatory DPIA under Art. 35 for high-risk processing. Treating them as ordinary data is a legal error. Body of Knowledge Domain II, Competency A
Exam tip for this scenarioData governance tell: lineage, quality dimensions, and bias types are the exam's favourite data-stage questions. Ask 'where did the data come from and what is wrong with it'.
3
Scenario 3 · 5 questions · Domain II
Scenario 3: LexAI Compliance
LexAI, a UK legal-tech startup, builds an AI contract-analysis tool that reviews commercial contracts and flags clauses that deviate from market norms. The tool is powered by a large language model trained on publicly available contracts. LexAI plans to sell the tool to law firms in the UK and EU. The company's general counsel must determine how the EU AI Act, the GDPR, and existing legal-sector regulations apply to the product.
Q11Domain II · Competency C
Under the EU AI Act, how is LexAI's contract-analysis tool most likely classified?
A. High risk under Annex III, because it is used in the legal sector
B. Unacceptable risk, because it analyses legal documents
C. Limited risk, because it is not listed as high risk in Annex III and does not interact with individuals in real time ✓
D. Minimal risk, because lawyers make the final decision
RATIONALEContract analysis tools are not listed in Annex III's high-risk categories (which cover employment, education, essential services, law enforcement, etc.). If the tool does not interact with individuals, it does not trigger Art. 50 transparency. It is most likely limited or minimal risk — the human lawyer's final decision does not make it high risk. Body of Knowledge Domain II, Competency C
Q12Domain II · Competency B
LexAI trained its model on publicly available contracts. What is the primary intellectual-property concern?
A. There is no IP concern because the contracts are publicly available
B. Publicly available contracts may still be protected by copyright or contain confidential terms; using them for training may require licensing or permission ✓
C. Training on public data is always fair use in all jurisdictions
D. The model's output is never copyrighted
RATIONALEPublic availability does not mean public domain. Contracts may have copyright protection, and using them for training may raise IP infringement risks depending on jurisdiction. IP governance for AI training data is a key BoK v2.1 topic — the provider must assess the legal basis for using training data. Body of Knowledge Domain II, Competency B
Q13Domain II · Competency A
The GDPR applies to LexAI's processing of contract data that may contain personal information (e.g., party names, signatures). Which principle is most relevant when contracts are used for a new purpose (model training)?
A. Data minimisation — the model should use as few contracts as possible
B. Accuracy — the contracts must be free of typos
C. Purpose limitation — using data collected for legal review for a new purpose (model training) requires assessing whether the new purpose is compatible ✓
D. Storage limitation — contracts must be deleted after one year
RATIONALEThe purpose-limitation principle requires that data collected for one purpose (legal contract review) is not repurposed for a materially different purpose (model training) without assessing compatibility or obtaining a new legal basis. This is a core GDPR principle tested under Domain II. Body of Knowledge Domain II, Competency A
Q14Domain I · Competency B
LexAI wants to ensure its model does not reproduce confidential contract clauses verbatim. Which technical approach is most appropriate?
A. Implement output filtering and deduplication to prevent verbatim reproduction of training data, and document the limitation in the model card ✓
B. Delete all training data after each use
C. Make the model larger so it generates more original text
D. There is no risk of verbatim reproduction in LLMs
RATIONALELLMs can memorise and reproduce training data verbatim. Output filtering, deduplication, and documentation in the model card are the standard mitigations. Making the model larger does not prevent memorisation, and denying the risk is a governance failure. Body of Knowledge Domain I, Competency B
Q15Domain II · Competency C
A law firm client asks LexAI to explain how the tool arrived at its assessment of a specific contract. Under the EU AI Act's transparency provisions, what is LexAI's obligation if the tool is classified as limited risk?
A. LexAI must provide a full technical audit of the model to every client
B. LexAI must disclose its source code to the public
C. No transparency obligation applies because the tool is limited risk and does not interact with individuals ✓
D. LexAI must notify every person named in every contract
RATIONALEFor a limited-risk system that does not interact with individuals, the Art. 50 transparency obligations do not apply. However, providing clear documentation (model card, instructions for use) is still good practice and supports the deployer's own accountability — even if it is not a legal requirement under the AI Act. Body of Knowledge Domain II, Competency C
Exam tip for this scenarioLaws tell: identify the jurisdiction and the statute before answering. EU = AI Act + GDPR; US = sectoral; SG = PDPA + frameworks.
4
Scenario 4 · 5 questions · Domain II & III
Scenario 4: Vitalis Health AI
Vitalis Health, a hospital network in France, deploys an AI triage system that ranks emergency-room patients by predicted severity. The system was built by a vendor, MediPredict, and is classified as high risk under the EU AI Act because it is a medical device. Vitalis must implement human oversight, monitor the system, and respond to incidents. A patient was triaged as 'low priority' by the AI but later found to have a life-threatening condition.
Q16Domain II · Competency C
As the deployer of a high-risk AI system, which of the following is Vitalis's primary obligation regarding human oversight?
A. Remove all human oversight to let the AI make decisions faster
B. Ensure human oversight by qualified staff who can override the system's output, with oversight measures adapted to the risk level and autonomy of the system ✓
C. Delegate oversight to the vendor, MediPredict
D. Implement oversight only after an incident occurs
RATIONALEDeployers of high-risk AI systems must ensure human oversight by qualified staff who can override the system. Oversight must be proportionate to the risk level and the system's autonomy, and must be in place before deployment — not only after an incident. The deployer cannot delegate this to the provider. Body of Knowledge Domain II, Competency C
Q17Domain IV · Competency C
The patient was triaged as 'low priority' but had a life-threatening condition. What is the most appropriate immediate governance response?
A. Ignore the incident because the AI is a medical device with CE marking
B. Fire the triage nurse who relied on the AI output
C. Trigger the incident-response plan: log the incident, investigate the root cause, determine whether the system should be suspended, and report to the provider and authorities if serious ✓
D. Immediately remove the AI system permanently without investigation
RATIONALEA near-miss or harm event triggers the incident-response lifecycle: detect, log, investigate, contain (suspend if necessary), report to the provider and the relevant authority if it is a serious incident. Ignoring, blaming individuals, or permanent removal without investigation are all inappropriate. Body of Knowledge Domain IV, Competency C
Q18Domain III · Competency A
MediPredict must maintain a risk-management system for this high-risk AI. Which of the following is a core element of that system?
A. A continuous, iterative process that identifies, estimates, evaluates, and mitigates risks throughout the system's lifecycle ✓
B. A one-time risk assessment at launch that is never revisited
C. A marketing plan for the product
D. A list of hospitals that bought the system
RATIONALEThe EU AI Act requires a continuous, iterative risk-management process (identify → estimate → evaluate → mitigate) throughout the lifecycle, not a one-time assessment. This is one of the core obligations for providers of high-risk systems and a key exam topic. Body of Knowledge Domain III, Competency A
Q19Domain II · Competency C
Vitalis must conduct a fundamental rights impact assessment (FRIA) before deploying this system. Which of the following is the FRIA's primary purpose?
A. To assess the system's commercial viability
B. To replace the DPIA required under the GDPR
C. To certify the system as compliant with the AI Act
D. To identify and mitigate risks to fundamental rights (e.g., health, non-discrimination) arising from the use of the system, and to describe the human-oversight measures ✓
RATIONALEThe FRIA (Art. 27) identifies and assesses risks to fundamental rights from deploying a high-risk AI system and describes the human-oversight measures to mitigate them. It is separate from (and additional to) the DPIA required under GDPR Art. 35. Body of Knowledge Domain II, Competency C
Q20Domain III · Competency C
After the incident, Vitalis considers retraining the model on its own patient data to improve local performance. What is the most important governance consideration?
A. Retraining is always better, so no governance is needed
B. Vitalis must ensure the retraining data is representative, labelled correctly, and that the retrained model is re-validated before deployment, as this constitutes a substantial modification under the AI Act ✓
C. Vitalis should retrain the model without telling MediPredict
D. Retraining by a deployer is prohibited under the EU AI Act
RATIONALERetraining a high-risk system on new data may constitute a 'substantial modification' under the AI Act, which can re-trigger the provider's conformity-assessment obligations. The retrained model must be re-validated, and the data must be representative and correctly labelled. Silent retraining without governance is a compliance failure. Body of Knowledge Domain III, Competency C
Exam tip for this scenarioHealth-AI tell: health data at scale = DPIA; high-risk AI deployment = FRIA. The provider does conformity, the deployer does oversight.
5
Scenario 5 · 5 questions · Domain III & IV
Scenario 5: LogiTrans Routing
LogiTrans, a European logistics company, deploys an AI routing system that optimises delivery routes for its 2,000-truck fleet. The system was trained on five years of route and delivery data and makes real-time route recommendations to drivers. After six months, the model begins recommending longer routes more frequently. The operations team also notices the model performs poorly during winter weather, which was not well-represented in the training data.
Q21Domain IV · Competency B
The model's increasing recommendation of longer routes is best investigated as which type of issue?
A. A hardware failure in the GPS system
B. Concept drift or a feedback-loop effect: the model may be optimising for a metric that diverges from the business goal, or its training data has become stale ✓
C. A deliberate attack by a competitor
D. Normal behaviour that requires no investigation
RATIONALEWhen a model's outputs systematically shift from expected behaviour, the first investigation should check for concept drift (the relationship between inputs and the optimal output has changed) or feedback loops (the model is optimising for a proxy metric). Both are core monitoring topics under Domain IV. Body of Knowledge Domain IV, Competency B
Q22Domain I · Competency B
The model performs poorly in winter weather because the training data under-represented winter conditions. What is the most appropriate remediation?
A. Collect and incorporate winter-condition data into the training set and retrain the model, then re-validate on winter data before deployment ✓
B. Deploy the model only in summer
C. Ignore it — drivers can compensate for winter conditions
D. Switch to a rules-based system for winter routing
RATIONALENon-representative training data is a data-governance failure. The correct remediation is to collect the missing data, retrain, and re-validate on the conditions that were under-represented. Restricting deployment seasonally or ignoring the gap are not governance-appropriate responses. Body of Knowledge Domain I, Competency B
Q23Domain IV · Competency B
LogiTrans wants to establish a monitoring plan for the routing system. Which metric is most appropriate for ongoing monitoring?
A. The number of trucks in the fleet
B. The company's stock price
C. The number of employees in the IT department
D. The model's prediction accuracy against actual delivery outcomes, tracked over time with pre-defined thresholds that trigger review ✓
RATIONALEProduction monitoring for AI systems must track the model's real-world performance against ground truth (actual delivery outcomes) over time, with pre-defined thresholds that trigger investigation or retraining. Fleet size, stock price, and headcount are not model-performance metrics. Body of Knowledge Domain IV, Competency B
Q24Domain III · Competency A
LogiTrans's routing AI does not make final decisions — drivers can ignore recommendations. What level of human oversight does this represent?
A. Human-on-the-loop (HOTL), because the system acts autonomously while humans can supervise and intervene ✓
B. Human-in-the-loop (HITL), because every decision requires human approval before execution
C. Human-out-of-the-loop (HOOTL), because humans have no role
D. No oversight, because drivers are not AI experts
RATIONALEWhen a system operates autonomously and humans can supervise and intervene but do not approve every action, it is human-on-the-loop (HOTL). HITL requires human approval before every action; HOOTL means no human role. The HITL/HOTL/HOOTL spectrum is a core BoK topic tested under Domain III. Body of Knowledge Domain III, Competency A
Q25Domain IV · Competency A
LogiTrans wants to build an AI inventory. Which of the following is the minimum information each entry should contain?
A. The vendor's name and the purchase price
B. The system's purpose, the data it uses, its risk tier, the named owner, and the date of the last review ✓
C. The model's source code
D. The names of all drivers who used the system
RATIONALEAn AI inventory must capture, at minimum: purpose, data used, risk tier, named owner, and last review date. This is the foundation of AI governance — you cannot govern what you have not inventoried. Source code and driver names are not inventory-level information. Body of Knowledge Domain IV, Competency A
Exam tip for this scenarioDeployment tell: drift type questions hinge on WHAT changed — inputs (data), rule (concept), or answer key (label).
6
Scenario 6 · 5 questions · Domain III & IV
Scenario 6: ScholarAI University
ScholarAI, an EU-based edtech company, sells an AI tutoring system to universities. The system generates personalised study plans and predicts student dropout risk. A university deployer notices the dropout-prediction model flags students from lower-income postcodes at higher rates. The university's data protection officer must also respond to a student's GDPR subject-access request about the dropout-prediction system.
Q26Domain III · Competency C
The dropout model flags lower-income-postcode students at higher rates. What is the most appropriate first step?
A. Conduct a disparate-impact analysis: measure flag rates across postcode groups, examine the features driving the disparity, and determine whether the model uses a proxy for a protected characteristic ✓
B. Accept the result — the model is probably correct because low income correlates with dropout
C. Remove the postcode feature and retrain immediately
D. Stop using the model permanently
RATIONALEA disparate-impact analysis is the correct first step: measure the disparity, investigate the features causing it, and determine whether a feature is a proxy for a protected characteristic. Accepting biased results, blindly removing features, or abandoning the model without investigation are all governance failures. Body of Knowledge Domain III, Competency C
Q27Domain II · Competency C
Under the EU AI Act, how is the dropout-prediction system classified?
A. Minimal risk, because the university makes the final intervention decision
B. Limited risk, because it only generates predictions
C. High risk under Annex III area 3, because it is used to determine access to education and evaluate learning outcomes ✓
D. Unacceptable risk, because it uses postcode data
RATIONALEAI systems used to determine access to or evaluate learning outcomes in education are high risk under Annex III area 3. The university's final decision does not change the classification — the system itself is high risk. Body of Knowledge Domain II, Competency C
Q28Domain IV · Competency A
A student submits a GDPR subject-access request asking about the dropout-prediction system's use of their data. Which of the following must the university provide?
A. The student's raw test scores only
B. The model's source code and training data
C. Information about whether their data is being processed, the purposes of processing, the logic involved in automated decisions, and the significance of the envisaged consequences ✓
D. Nothing — the system is proprietary
RATIONALEUnder GDPR Art. 15 (subject access) and Art. 22 (automated decisions), the data subject has the right to know whether their data is processed, the purposes, the logic involved, and the consequences. Source code and training data are not part of this right, and proprietary status does not override data-subject rights. Body of Knowledge Domain IV, Competency A
Q29Domain III · Competency D
ScholarAI, as the provider, must maintain technical documentation. Which of the following is NOT a required element of that documentation?
A. A description of the system's intended purpose and the categories of data used
B. The personal names and addresses of all students who used the system ✓
C. The way the system achieves its output, including the model architecture and training process
D. Pre-deployment testing results and performance metrics
RATIONALETechnical documentation for high-risk AI must include intended purpose, data categories, model architecture, training process, and testing results. Personal names and addresses of all students are not technical documentation — they are personal data that must be protected, not published in a technical file. Body of Knowledge Domain III, Competency D
Q30Domain IV · Competency C
The university wants to decommission the dropout-prediction model. Which of the following is the most appropriate decommissioning consideration?
A. Plan the decommissioning: archive or securely delete personal data per the retention policy, document the decision and rationale, and ensure ongoing systems that depended on the model are not disrupted ✓
B. Delete the model and all associated data immediately without review
C. Keep the model running indefinitely because it is too hard to decommission
D. Sell the model and data to another university
RATIONALEDecommissioning an AI system requires planning: personal data must be archived or deleted per the retention policy, the decision must be documented, and dependent systems must be managed. Immediate deletion without review, keeping it running forever, and selling data are all inappropriate. Body of Knowledge Domain IV, Competency C
Exam tip for this scenarioEdTech tell: purpose creep and subgroup accuracy gaps are the governance answers. Document, restrict, and re-validate.
7
Scenario 7 · 5 questions · Domain IV
Scenario 7: RetailMind Marketing
RetailMind, a Singapore e-commerce company, uses an AI system that personalises product recommendations and marketing emails based on customer browsing and purchase history. A customer complains that the system only shows them discount products, limiting their choices. The company's marketing director wants to expand the system to use customer location data and facial-recognition cameras in physical stores to track in-store behaviour. The PDPC has issued guidance on AI in retail.
Q31Domain IV · Competency A
The customer complains the recommendation system limits their choices by only showing discount products. Which governance concept is most relevant?
A. Filter bubble and choice restriction — the system should be designed to present a diverse set of options and allow users to adjust or opt out of personalisation ✓
B. Data minimisation — the system should collect less browsing data
C. Concept drift — the recommendation logic has changed
D. Overfitting — the model memorised the customer's history
RATIONALEFilter bubbles and choice restriction are governance concerns for recommendation systems. The system should present diverse options and allow opt-out or adjustment. Data minimisation and drift do not address the user-choice problem directly. Body of Knowledge Domain IV, Competency A
Q32Domain II · Competency A
RetailMind wants to use facial-recognition cameras in stores. Under Singapore's PDPA, what is the most significant compliance consideration?
A. Facial recognition is not personal data under the PDPA
B. Retail stores are exempt from the PDPA
C. Facial recognition can be deployed without any compliance steps
D. Biometric data is personal data; collecting it requires consent, notification, and a lawful purpose, and the PDPC has issued specific guidance on facial recognition in retail ✓
RATIONALEFacial-recognition data is biometric personal data under the PDPA. Collection requires consent, notification, and a clear lawful purpose. The PDPC has issued guidance specifically addressing facial recognition in commercial settings — it is not exempt and cannot be deployed without compliance steps. Body of Knowledge Domain II, Competency A
Q33Domain II · Competency A
The marketing director wants to use location data for personalisation. Which of the following is the most important safeguard?
A. Provide clear notice, obtain appropriate consent for the new purpose, and allow customers to opt out of location-based personalisation ✓
B. Collect location data continuously without telling customers
C. Sell the location data to third parties
D. Use location data only for marketing, never for security
RATIONALEUsing location data for a new purpose (personalisation) requires notification and consent under the PDPA. Customers must be able to opt out. Continuous covert collection, selling data, and restricting use to marketing only are all non-compliant. Body of Knowledge Domain II, Competency A
Q34Domain IV · Competency C
RetailMind must establish an incident-response plan for its AI systems. Which of the following is the most critical element?
A. A plan to double marketing spend after every incident
B. A plan to delete all customer data after every incident
C. A plan to blame the vendor for every incident
D. A defined process for detecting, triaging, and responding to AI-related incidents, with a named owner who can pause or stop the system and a notification process for affected customers ✓
RATIONALEAn AI incident-response plan must define detection, triage, response (including the ability to pause or stop the system), a named owner for the kill switch, and a notification process for affected individuals. Marketing spend, data deletion, and vendor blame are not incident-response elements. Body of Knowledge Domain IV, Competency C
Q35Domain IV · Competency B
The company wants to evaluate whether its recommendation system is performing well. Which evaluation approach is most appropriate for a deployed recommendation system?
A. Evaluate only the model's offline accuracy on historical data
B. Evaluate only the model's training loss
C. Monitor online metrics (click-through, conversion, diversity of recommendations) alongside user feedback, with thresholds that trigger review when performance degrades ✓
D. No evaluation is needed — recommendation systems do not degrade
RATIONALEDeployed recommendation systems require online monitoring of business-relevant metrics (click-through, conversion, diversity) plus user feedback, not just offline accuracy. Training loss is a development metric, not a production one. The exam tests the difference between offline and online evaluation. Body of Knowledge Domain IV, Competency B
Exam tip for this scenarioRetail tell: emotion recognition in retail is limited-risk, NOT prohibited. Transparency and consent are the levers.
8
Scenario 8 · 5 questions · Domain II & IV
Scenario 8: GovServe TaxAI
GovServe, a national tax authority, deploys an AI system that flags tax returns for audit based on risk scoring. The system was trained on ten years of historical audit data. A journalist discovers the system flags returns from self-employed workers at three times the rate of those from large corporations. The tax authority must respond to the public concern and demonstrate its AI governance to parliament.
Q36Domain II · Competency C
Under the EU AI Act, how is the tax-authority's risk-scoring system classified?
A. Minimal risk, because tax audits are administrative
B. High risk under Annex III, because it is used by a public authority to evaluate eligibility for benefits or services and may affect individuals' rights ✓
C. Limited risk, because it only flags returns
D. Unacceptable risk, because it processes financial data
RATIONALEAI systems used by public authorities to evaluate eligibility for public services and benefits, or to evaluate tax returns, are high risk under Annex III area 5. The administrative nature of tax audits does not reduce the classification. Body of Knowledge Domain II, Competency C
Q37Domain I · Competency C
The system flags self-employed workers at three times the rate of large corporations. What is the most likely root cause?
A. Self-employed workers commit more tax fraud
B. The historical training data reflects past audit priorities that concentrated on self-employed workers, creating a biased training set that perpetuates the pattern ✓
C. The model is broken and should be replaced
D. The large corporations bribed the tax authority
RATIONALEHistorical audit data reflects past enforcement priorities. If past audits concentrated on self-employed workers, the model learns to flag them, creating a feedback loop that perpetuates the bias. This is a classic data-bias problem, not evidence that self-employed workers commit more fraud. Body of Knowledge Domain I, Competency C
Q38Domain IV · Competency A
The tax authority must respond to parliament. Which of the following is the most appropriate governance demonstration?
A. Refuse to comment, citing national security
B. Publish the full source code of the model
C. Deny that any disparity exists
D. Publish a summary of the system's purpose, data sources, bias-testing results, human-oversight measures, and the steps taken to address the disparity ✓
RATIONALETransparency and accountability for public-sector AI means publishing a summary of the system's purpose, data sources, bias-testing results, and oversight measures. Refusing to comment, publishing source code, or denying the disparity are all inappropriate responses to a legitimate public concern. Body of Knowledge Domain IV, Competency A
Q39Domain IV · Competency C
The tax authority wants to implement a kill switch for the system. What is the primary purpose of a kill switch?
A. To delete all tax returns processed by the system
B. To fire the data science team
C. To immediately stop the system's operation when a significant harm is confirmed, while the root cause is investigated and remediated ✓
D. To switch the system to a different model instantly
RATIONALEA kill switch immediately stops the AI system's operation when significant harm is confirmed, allowing investigation and remediation. It does not delete data, fire staff, or switch models — it contains the impact while the root cause is addressed. Body of Knowledge Domain IV, Competency C
Q40Domain II · Competency C
Which of the following best describes the difference between a DPIA and an AI conformity assessment?
A. They are the same document with different names
B. A DPIA assesses risks to data-protection rights under the GDPR; an AI conformity assessment assesses compliance of the AI system with the EU AI Act's requirements before market placement ✓
C. A DPIA is required only for low-risk systems; a conformity assessment is for all systems
D. A conformity assessment is optional; a DPIA is mandatory for all AI
RATIONALEA DPIA (GDPR Art. 35) assesses risks to data-protection rights; an AI conformity assessment (EU AI Act) assesses the system's compliance with AI Act requirements before the provider places it on the market. They are separate assessments with different scopes and legal bases — not interchangeable. Body of Knowledge Domain II, Competency C
Exam tip for this scenarioPublic-sector tell: bias testing and error-rate analysis are the assessment activities. Same-count comparisons are a distractor.
9
Scenario 9 · 5 questions · Domain III & IV
Scenario 9: AgenticOps Automation
AgenticOps, a US technology company, deploys agentic AI assistants that can autonomously execute multi-step business tasks: drafting and sending emails, updating CRM records, and scheduling meetings. The agents use a large language model and can call external APIs. The company's CISO is concerned about prompt-injection attacks, unauthorised actions, and the lack of a governance framework for autonomous AI. A BoK v2.1 addition, agentic AI, raises new governance questions about oversight, accountability, and control.
Q41Domain I · Competency A
What distinguishes an agentic AI system from a traditional predictive model in terms of governance?
A. Agentic AI is always lower risk because it assists humans
B. Agentic AI does not need governance because it is a tool, not a decision-maker
C. Agentic AI and predictive models are governed identically under all frameworks
D. Agentic AI can take autonomous actions in the world (calling APIs, sending emails), which creates new risks — unauthorised actions, cascading errors, and the need for action-level oversight and guardrails ✓
RATIONALEAgentic AI introduces new governance risks because it can take autonomous real-world actions, not just make predictions. This requires action-level oversight (guardrails, approval gates, action logging) that goes beyond prediction-level oversight. BoK v2.1 added agentic AI as a key topic under Domain III. Body of Knowledge Domain I, Competency A
Q42Domain I · Competency C
The CISO is concerned about prompt-injection attacks. What is a prompt-injection attack, and what is the appropriate governance response?
A. A prompt injection is a hardware attack on the server; the response is to upgrade the servers
B. A prompt injection is a type of phishing email; the response is to train employees on email security
C. A prompt injection is an attack where malicious instructions are embedded in data the model processes, causing it to execute unauthorised actions; the response is to implement input filtering, output validation, and action guardrails ✓
D. Prompt injection is not a real risk for LLMs
RATIONALEPrompt injection embeds malicious instructions in data the model processes (e.g., a website the agent reads), causing it to execute unauthorised actions. The governance response includes input filtering, output validation, and guardrails on the actions the agent can take. This is a key BoK v2.1 topic. Body of Knowledge Domain I, Competency C
Q43Domain III · Competency A
Which human-oversight pattern is most appropriate for an agentic AI that can send emails and update CRM records autonomously?
A. Human-out-of-the-loop (HOOTL) — the agent acts fully autonomously with no human role
B. No oversight is needed because the agent is just a tool
C. Oversight only after an incident has occurred
D. Human-on-the-loop (HOTL) or human-in-the-loop (HITL) for high-risk actions, with approval gates before irreversible actions like sending external emails ✓
RATIONALEFor agentic AI that can take real-world actions, HOTL or HITL for high-risk actions is the appropriate oversight pattern, with approval gates before irreversible actions. HOOTL is inappropriate for systems that can send emails or modify records, and post-incident oversight is too late. Body of Knowledge Domain III, Competency A
Q44Domain IV · Competency C
AgenticOps wants to log the agent's actions for auditability. Which of the following is the most important element of an action log?
A. For each action: the trigger, the action taken, the data accessed, the timestamp, and whether it was auto-executed or human-approved ✓
B. The name of the LLM vendor
C. The agent's internal token count
D. The colour scheme of the user interface
RATIONALEAn action log for agentic AI must capture the trigger, the action, the data accessed, the timestamp, and the approval status for each action. This supports auditability, incident investigation, and accountability. Vendor name, token count, and UI design are not audit-relevant. Body of Knowledge Domain IV, Competency C
Q45Domain III · Competency A
Which of the following is the most appropriate governance control for an agentic AI system's ability to call external APIs?
A. Give the agent unrestricted access to all APIs for maximum flexibility
B. Restrict API access to a whitelist of approved endpoints, with rate limits and approval gates for high-impact actions ✓
C. Disable all API access — agents should not call APIs
D. Let the agent decide which APIs to call based on its own judgement
RATIONALERestricting API access to a whitelist with rate limits and approval gates for high-impact actions is the standard governance control for agentic AI. Unrestricted access and agent-decided access create unmanageable risk; disabling all APIs removes the system's utility. Body of Knowledge Domain III, Competency A
Exam tip for this scenarioAgentic tell: guardrails, human oversight, and kill switches are the answers for autonomous agents. Contain first, investigate after.
10
Scenario 10 · 5 questions · Domain III & IV
Scenario 10: GreenAI Energy
GreenAI, a renewable-energy company, deploys an AI system that optimises wind-turbine output by predicting wind patterns and adjusting blade angles. The system was trained on meteorological data and sensor readings. After a software update, the model's predictions degrade during storms, and the engineering team must decide whether to roll back the update or retrain. The company's board has asked for an AI governance framework that covers the full lifecycle from development to decommissioning.
Q46Domain IV · Competency C
After the software update, the model's predictions degrade during storms. What is the most appropriate first step?
A. Ignore it — storms are rare and the system works fine in normal conditions
B. Retrain the model from scratch on new data without investigating the cause
C. Keep the update and hope the model self-corrects
D. Roll back to the previous version immediately and investigate the update's impact on storm-period performance before re-deploying ✓
RATIONALEWhen a software update degrades performance in a specific condition, the first step is to roll back to the known-good version and investigate the update's impact. Ignoring storm performance, retraining without diagnosis, or hoping for self-correction are all governance failures for a safety-relevant system. Body of Knowledge Domain IV, Competency C
Q47Domain III · Competency A
Which of the following best describes the AI development lifecycle governance approach that GreenAI should adopt?
A. Governance is only needed at deployment; development and decommissioning do not need governance
B. Governance is only needed for the training phase
C. Governance is optional for renewable-energy AI systems
D. Governance applies across the full lifecycle: planning, data, design, training, validation, deployment, monitoring, and decommissioning — each stage has specific controls and documentation requirements ✓
RATIONALEAI governance applies across the entire lifecycle, not just at one stage. Each phase — planning, data, design, training, validation, deployment, monitoring, and decommissioning — has specific controls, documentation, and accountability. This lifecycle approach is a core BoK v2.1 concept. Body of Knowledge Domain III, Competency A
Q48Domain IV · Competency A
GreenAI's board wants to know which framework to adopt. Which of the following is the most appropriate combination for a company deploying high-risk AI?
A. No framework is needed — the company should rely on its own judgement
B. Only the NIST AI RMF, because it is the only framework that matters
C. Only the EU AI Act, because it covers everything
D. NIST AI RMF for the risk-management process, ISO/IEC 42001 for the AI management system, and jurisdictional regulations (e.g., EU AI Act) for compliance ✓
RATIONALEA mature AI governance programme combines NIST AI RMF (risk-management process: Govern-Map-Measure-Manage), ISO/IEC 42001 (certifiable AI management system with PDCA), and jurisdictional regulations (EU AI Act for compliance). Relying on one alone or on none is insufficient. Body of Knowledge Domain IV, Competency A
Q49Domain IV · Competency C
GreenAI must decommission an old turbine-prediction model. Which of the following is the most important decommissioning step?
A. Delete the model and forget about it
B. Keep the old model running alongside the new one indefinitely
C. Document the decommissioning decision and rationale, archive or securely delete the training data per the retention policy, and ensure the replacement model has been validated and the transition does not disrupt operations ✓
D. Sell the old model to a competitor
RATIONALEDecommissioning requires documentation, data management per retention policy, validation of the replacement, and operational continuity. Deleting without documentation, running both indefinitely, and selling the model are all inappropriate. Body of Knowledge Domain IV, Competency C
Q50Domain III · Competency A
The engineering team must decide whether to use the model's predictions or to add a rules-based safety override for storm conditions. Which governance principle supports the override?
A. Data minimisation — the model should use less data during storms
B. Transparency — the model's storm predictions should be published
C. Accountability — the engineering team should take a vote
D. The precautionary principle and defence-in-depth: for safety-critical conditions, a rules-based override provides a failsafe when the model's predictions are unreliable ✓
RATIONALEDefence-in-depth and the precautionary principle support adding a rules-based override for safety-critical conditions where the model is known to be unreliable. This is a standard engineering governance control — the model is not the sole decision-maker in high-risk conditions. Body of Knowledge Domain III, Competency A
Exam tip for this scenarioLifecycle tell: decommissioning and data handling per retention policy. Retire safely, handle data, document the decision.