You are here: Practice › Question bank
Exam-Level Question Bank
Fifty original practice questions across the four domains, written at exam difficulty. Try each one before you open the reading.
How to use this bank. Read the stem, pick your answer, then open the reading. Note the answer you chose and why. Our readings are starting points, not official answers, so check each one against the Body of Knowledge and the current text of the source it names. These questions are original and written in a similar style to the exam. They are not IAPP items.
Domain I: Foundations of AI governance
11 questions
Q1 · HardA credit model scores 97% accuracy on its training data but 71% on new applications. The governance lead wants the control most likely to catch this gap before deployment. Which control fits best?
- A. Hold out data never used in training and test against defined acceptance thresholds.
- B. Publish the training accuracy in the model card so reviewers can see the strong results.
- C. Add more features so the model can capture the patterns in the data more precisely.
- D. Ask the data science team to confirm in writing that the model is well designed.
- E. Retrain the same model on the full dataset and report the new training accuracy.
Our reading (answer A): The gap points to poor generalisation, often overfitting. Testing on held-out data against thresholds set in advance is the control that checks generalisation before release. Reporting training accuracy describes the problem without detecting it, and extra features can make overfitting worse.
Check: Domain I notes on model performance and validation; your course's section on overfitting.
Q2 · Very hardA company says it has strong AI ethics because it published a set of AI principles. Which statement best identifies the gap an AI governance programme should close?
- A. The principles have not yet been turned into roles, decision rights, controls and evidence.
- B. The principles need to be shorter and simpler so that all staff can recall them without notes.
- C. Ethics should sit in a separate team from governance so that the two functions do not overlap.
- D. The principles only become relevant once a system has been retired from service entirely.
Our reading (answer A): Principles state intent. Governance turns intent into accountable owners, controls and records that can be checked. The other options treat wording, team structure or timing as the main issue, which in our reading they are not.
Check: Domain I material on the purpose of AI governance versus AI ethics.
Q3 · HardA safety team lowers the decision threshold of a defect-detection model so that it flags more potential defects. Which trade-off is most likely?
- A. Precision and recall both rise together.
- B. Precision falls and recall rises.
- C. Recall falls and precision rises.
- D. Accuracy necessarily rises as well.
- E. Both measures stay exactly the same.
Our reading (answer B): A lower threshold flags more items, so more real defects are caught (recall rises) but more flags are wrong (precision falls). Accuracy may move either way and does not necessarily rise, especially where defects are rare.
Check: Your confusion-matrix and metrics notes; check how the course defines each term.
Q4 · HardA hiring model is trained on ten years of past hiring decisions. Those decisions reflected discriminatory practices the company has since stopped. Which type of bias is most directly present in the labels?
- A. Measurement bias
- B. Historical bias
- C. Sampling bias
- D. Automation bias
Our reading (answer B): The labels encode past decisions that were themselves biased, which is historical bias. Measurement bias concerns how a variable is recorded, sampling bias concerns who is in the data, and automation bias concerns over-reliance by people using the output.
Check: The bias taxonomy in your Domain I notes.
Q5 · Very hardA vendor says its deep neural network is interpretable because its full set of weights is in the documentation. Which response best reflects the difference between interpretability and explainability?
- A. Published weights establish explainability in full, so a reviewer can rely on them and no further evidence about individual outputs is needed to approve the system for use.
- B. Interpretability and explainability mean exactly the same thing in every regulation and framework, so the vendor's documentation is complete and correct for review.
- C. Interpretability concerns how the model works internally; explainability concerns the reasons for specific outputs. Raw weights alone may give neither in a usable form.
- D. Explainability only applies to rule-based systems, so a deep neural network cannot be asked to explain any of its individual outputs at all.
- E. Interpretability is a legal term for documentation of weights, while explainability is a technical term for the vendor's own internal testing reports.
Our reading (answer C): In common usage, interpretability is about how a model works and explainability is about why a particular output was produced. Usage varies between sources, so note the definition your course uses. Publishing weights does not, by itself, explain a specific decision.
Check: Your Domain I glossary; compare the definitions in two sources.
Q6 · HardWho should be accountable for the decision to retire a customer-facing AI system that is generating repeated complaints?
- A. The data scientist who built the model and understands its technical behaviour best
- B. The vendor, because it owns the model weights and controls the hosting infrastructure
- C. The business owner named in the governance charter, with risk and legal input
- D. The customer-support team that receives the complaints and handles them every day
Our reading (answer C): Accountability for a business-impact decision should sit with a named owner who has the authority to act, with specialist input. Builders and support staff may advise, and a vendor may not control the deployment decision.
Check: RACI or decision-rights material in Domain I.
Q7 · HardAn organisation wants to begin governing AI but does not know which tools staff use. What is the most proportionate first step?
- A. Ban all external AI tools immediately, so that no unapproved system of any kind can be used by staff at all.
- B. Ask each team to write its own AI policy, so that local needs are reflected properly.
- C. Wait for a regulator to publish guidance, so the organisation can follow it precisely.
- D. Build an inventory of AI systems and use cases, including tools staff have adopted informally.
- E. Buy a single enterprise AI platform and require every team to use only that product for all of its AI work.
Our reading (answer D): You cannot govern what you have not identified, so an inventory is usually the first control. A ban can push use underground, and waiting leaves the risk unmanaged.
Check: Domain I material on inventories and shadow AI; Domain IV on procurement.
Q8 · Very hardA generative AI assistant confidently cites a court case that does not exist. Which risk category best describes the issue?
- A. Fairness, because the assistant treated some users unequally in its responses.
- B. Privacy, because personal data about a named individual was revealed.
- C. Security, because an outside attacker must have deliberately caused the output.
- D. Reliability and accuracy of outputs, which can include fabricated content.
Our reading (answer D): In our reading this is primarily an accuracy and reliability issue, since the output is confidently wrong. Depending on context it could also raise other concerns, so the best answer depends on what the question asks you to focus on.
Check: Your framework's list of trustworthiness characteristics (for example, the NIST AI RMF).
Q9 · HardA team must show where a training dataset came from, who owns it, and under what licence it may be used. Which record best serves this purpose?
- A. Data lineage
- B. A model card
- C. An audit log
- D. A data dictionary
- E. Data provenance
Our reading (answer E): Provenance covers origin, ownership, custody and licence. Lineage traces transformations and flows through a pipeline over time. A model card and an audit log record other things.
Check: The provenance and lineage definitions in Domain I; check the course's wording.
Q10 · HardA loan model's inputs look the same as last year, but defaults now occur in different customer profiles because economic conditions have changed. Which term best describes this?
- A. Concept drift
- B. Data drift
- C. Sampling bias
- D. Model collapse
Our reading (answer A): The input data have not shifted, but the relationship between inputs and outcomes has. That is concept drift. Data drift would mean the distribution of the inputs has changed.
Check: Drift definitions in Domain I; compare with your course's examples.
Q11 · HardWhy might an AI governance body be cross-functional rather than owned by a single department?
- A. AI risks span legal, privacy, security, business and technical functions, so decisions benefit from several perspectives.
- B. It avoids the need for executive sponsorship, so that decisions can be taken quickly by staff without senior involvement at all.
- C. Regulators require equal numbers of members from every function, so the committee must be balanced in size by law.
- D. It removes the need to document decisions, since the group's discussions are recorded informally in its minutes.
- E. It keeps AI decisions away from the legal department, so that lawyers cannot slow down product launches.
Our reading (answer A): In our reading, the case for a cross-functional body rests on the breadth of AI risk. Executive sponsorship is still needed, and documentation is still required. Check any claim about regulatory membership rules against the source.
Check: Domain I material on governance structures and roles.
Domain II: Laws, standards and frameworks
12 questions
Q12 · HardUnder the EU AI Act, which statement best describes the relationship between prohibited practices and high-risk systems?
- A. Prohibited practices are a subset of high-risk systems that need extra documentation and a separate sign-off before they go live.
- B. Prohibited practices are banned under Article 5, with limited exceptions. High-risk systems are a separate category with their own duties.
- C. Both are regulated in exactly the same way under the Act, and the only real difference between the two is the level of the fine.
- D. High-risk systems are banned outright, and prohibited practices are only discouraged through voluntary codes of conduct that firms may choose to adopt for their own use.
Our reading (answer B): Article 5 lists practices that are not permitted, subject to narrow exceptions. Article 6 and Annex III cover high-risk systems, which are permitted subject to requirements. Treating the two as the same category is a common error.
Check: Article 5 and Article 6 of Regulation (EU) 2024/1689, in the current consolidated text.
Q13 · Very hardA hiring tool automatically rejects applicants without any human review. Which GDPR provision is most directly engaged?
- A. Article 6, on lawful bases for processing, which applies to every decision that uses personal data in any form
- B. Article 22, on decisions based solely on automated processing that significantly affect the person
- C. Article 33, on notifying the authority of a personal data breach within the required time
- D. Article 12, on the format and clarity of information provided to the people concerned
- E. Article 5, on the principles of processing, including fairness, transparency, purpose limitation and accuracy of the data
Our reading (answer B): In our reading, Article 22 is the most directly relevant provision because the decision is solely automated and has a significant effect. A data protection impact assessment under Article 35 may also be needed. Check how the provision applies in your jurisdiction.
Check: GDPR Articles 22 and 35; your privacy notes.
Q14 · Very hardA company wants to process health information to train a model. Which analysis is most complete?
- A. An Article 6 lawful basis is enough on its own, provided the purpose is clearly stated.
- B. An Article 9 condition is enough on its own, because health data is already special.
- C. An Article 6 lawful basis and an Article 9 condition must both be identified.
- D. Anonymised data needs no lawful basis at all, so no further analysis is needed here.
Our reading (answer C): Special category data under GDPR Article 9 generally needs both an Article 6 lawful basis and an Article 9 condition. The exact conditions and their scope should be checked against the text and local guidance.
Check: GDPR Articles 6 and 9; the Article 9(2) conditions.
Q15 · HardA medical AI product receives an update that introduces a defect causing patient harm. Which legal area is most likely to be relevant to the claim?
- A. Copyright infringement
- B. Trademark registration
- C. Product liability
- D. Competition law
- E. Contract law
Our reading (answer C): Product liability is the most relevant area for physical harm from a defective product. How software and updates are treated varies by jurisdiction, so check the regime that applies to the product.
Check: Product liability rules that apply where the product is sold; your legal notes.
Q16 · Very hardA company trains a model on publicly available web text that includes copyrighted material. Which governance control is most directly relevant?
- A. Rename the model so that it does not infringe anything, since the name is what matters most.
- B. Publish the full training data on the company's public website so that owners and the public can see what was used.
- C. Delete all copyrighted items once training has finished, so no trace of them remains anywhere.
- D. Document data sources, licence and opt-out status, and obtain legal review of the basis for use.
Our reading (answer D): Documentation of sources and rights, plus legal review of the basis for use, is the most direct control. The legal position on training data differs by jurisdiction and is unsettled in some places, so treat this as a governance question for counsel.
Check: Your notes on training data and intellectual property; current guidance in your jurisdiction.
Q17 · HardA deployer buys an AI tool from a vendor. Which contract term best supports the deployer's own compliance obligations?
- A. A disclaimer stating that the vendor is not liable for anything that happens to the users.
- B. A promise that the tool will never change in any way during the whole life of the contract.
- C. The vendor's marketing claims listed in the contract as binding warranties on performance.
- D. Rights to receive technical documentation, notice of material changes, and incident information.
- E. A right to audit the vendor's offices once a year, without giving the vendor any notice at all first.
Our reading (answer D): A deployer needs information from the provider to meet its own duties and to monitor the system. Disclaimers and unchanging promises do not provide that. Check the specific information duties in the Act.
Check: Provider and deployer duties in the EU AI Act; your procurement notes.
Q18 · HardA customer-facing chatbot powered by AI is deployed in the EU. Which transparency duty is most directly relevant?
- A. Informing people that they are interacting with an AI system, unless this is obvious from the circumstances
- B. Publishing the full source code of the chatbot publicly so that anyone can inspect how it works.
- C. Obtaining a conformity certificate for every single conversation the chatbot has with its users, filed before each reply.
- D. Registering every conversation with the national authority in the EU database before answering.
Our reading (answer A): In our reading, the most relevant transparency duty is the duty to inform people they are interacting with AI, with limited exceptions. Check the Article 50 text in the current consolidated version before relying on this.
Check: Article 50 of Regulation (EU) 2024/1689.
Q19 · Very hardWhich statement about a fundamental rights impact assessment (FRIA) under the EU AI Act is most accurate?
- A. All providers must complete one for every AI system they place on the market in the EU, before it is first used.
- B. It replaces the data protection impact assessment entirely, so only one assessment is needed.
- C. It is required only for general-purpose AI models that are used across several sectors.
- D. It is a voluntary good practice with no legal role, so no organisation has any duty to complete it.
- E. It applies to specified deployers of certain high-risk systems, so it is not a universal requirement.
Our reading (answer E): In our reading, the FRIA duty is targeted at certain deployers and certain high-risk uses, and it sits alongside other assessments rather than replacing them. Check Article 27 and the current text, as the scope has been discussed in amendments.
Check: Article 27 of Regulation (EU) 2024/1689; the current consolidated version.
Q20 · HardA company fine-tunes a general-purpose AI model for internal use. Which question should it answer first?
- A. What colour the user interface should be, so that it matches the company's brand guidelines
- B. Whether the modification and use make it a provider under the Act's criteria
- C. Whether the model's name is a registered trademark in the company's home jurisdiction
- D. How many users the system will have once it is rolled out to the whole organisation
Our reading (answer B): Whether the company becomes a provider depends on the Act's criteria for placing a system on the market or substantially modifying it. The answer determines which obligations apply, so it comes first. Check the definitions and the value-chain provisions.
Check: Article 3 definitions and Article 25 in the current text.
Q21 · Very hardA US company sells a predictive tool only to non-EU businesses, but EU-based decision-makers use its reports. Which statement best describes the position?
- A. EU rules may still apply depending on how the outputs are used in the EU, so the company should check its scope.
- B. The company is always outside EU rules because it sells only to businesses located outside the European Union, wherever they are.
- C. EU rules apply only to companies whose headquarters are located inside the European Union.
- D. Only the EU customers are regulated, so the company's own obligations are unaffected overall.
- E. Because the reports are free to download, no regulatory duties arise for the company at all.
Our reading (answer A): The Act's scope can reach some providers and deployers outside the EU where outputs are used in the Union. Whether it applies depends on the scope provisions and the facts, so a scope assessment is the right first step.
Check: The scope article (Article 2) in the current consolidated text.
Q22 · HardWhich pairing best matches a framework to its typical character?
- A. NIST AI RMF: a certification scheme for AI vendors that is run by a national government body
- B. EU AI Act: a voluntary code of conduct that companies may adopt on a purely optional basis
- C. ISO/IEC 42001: an AI management system standard that organisations can be certified against
- D. OECD AI Principles: binding national legislation that applies directly in all member states
Our reading (answer C): In our reading, ISO/IEC 42001 is a management system standard that organisations can be certified against. The NIST AI RMF is voluntary guidance, the EU AI Act is binding regulation, and the OECD AI Principles are non-binding. Confirm each against its official text.
Check: The official pages for each framework (see the Resources page).
Q23 · Very hardUnder the EU AI Act, which party generally carries the duty of post-market monitoring for a high-risk system?
- A. The deployer, in every case, without any help from the provider
- B. The national regulator, which carries the duty in practice for each product
- C. The provider, under its own quality and monitoring system
- D. The end user, who must report every problem they notice themselves
- E. The importer, which carries the whole duty for the product after sale
Our reading (answer C): In our reading, post-market monitoring is primarily a provider duty, with deployers having their own monitoring and reporting duties. Check the monitoring article and the deployer article in the current consolidated text.
Check: Article 72 and the deployer obligations in Article 26 of Regulation (EU) 2024/1689.
Domain III: AI development
13 questions
Q24 · HardA team needs a standard document that describes how a dataset was collected, how it was labelled, and which gaps are known. Which artifact best fits?
- A. A penetration test report for the system
- B. A datasheet that documents the dataset
- C. A marketing brief for the product launch
- D. A deployment checklist for the release
- E. A staff training record
Our reading (answer B): A datasheet records collection, composition, labelling and known limitations of a dataset. The other artifacts serve security, marketing or deployment purposes.
Check: Domain III notes on data documentation.
Q25 · HardA model card for a face-recognition tool should prioritise which information for governance reviewers?
- A. The model's name and version number, together with the date on which it was first released
- B. The team's favourite training techniques and the tools they used during development work
- C. The hardware the model runs on, including the server models, the cloud region used and the storage it needs
- D. Intended and out-of-scope uses, performance across relevant subgroups, and known limitations
Our reading (answer D): Reviewers need to know what the system is for, where it should not be used, how it performs for different groups and what it cannot do. Names and hardware matter less for this decision.
Check: Domain III notes on documentation and model cards.
Q26 · Very hardA model performs well overall but poorly for a small subgroup. What is the best next step?
- A. Report the overall accuracy only, because the subgroup is too small to matter for the decision.
- B. Remove the subgroup from the test set so that the overall results look more consistent.
- C. Evaluate the subgroup separately, and assess sample size and uncertainty before drawing any conclusions.
- D. Raise the overall accuracy target so that the subgroup's weaker results are no longer a concern.
- E. Tell the subgroup that the model is accurate enough for everyone, and continue with the rollout.
Our reading (answer C): Overall metrics can hide poor performance for a subgroup. The next step is to measure that subgroup properly and check whether the sample is large enough to support a conclusion, since a small sample can produce unreliable figures.
Check: Domain III notes on testing across subpopulations.
Q27 · HardAn LLM agent reads web pages and can send emails. A page contains hidden instructions telling the agent to forward customer data. Which control is most directly relevant?
- A. Restricting the agent's tools, filtering untrusted inputs, and requiring approval for external messages
- B. Differential privacy during training, so that the model cannot reveal details about any one record
- C. Increasing the model's size so that it becomes better at spotting hidden instructions in pages
- D. Translating the page into another language before the agent reads it, which would remove any hidden text from the page
Our reading (answer A): This is an injection risk from untrusted content. Limiting what the agent can do, treating web content as untrusted, and requiring approval for consequential actions reduce the impact. Training-time privacy controls do not address this path.
Check: Domain III notes on security and agentic architectures; your course's section on prompt injection.
Q28 · Very hardWhich statement about differential privacy is most accurate?
- A. It guarantees that the model is fair to every group of people whose data is used in training.
- B. It removes the need for a lawful basis, because the data is protected by the mathematics of the method itself and by the model.
- C. It works only on image data, so it cannot be used for text, tabular or other structured data.
- D. It provides a formal, quantifiable limit on how much the output can reveal about any one individual's data.
- E. It is a legal certificate issued by regulators that confirms the training data has been anonymised.
Our reading (answer D): In its formal definition, differential privacy bounds how much any one record can change the output, through a privacy parameter. It says nothing about fairness, and it does not replace legal bases.
Check: Your Domain III notes on privacy-enhancing technologies.
Q29 · Very hardA company trains models with federated learning so that raw data stays on local devices. Which statement is most accurate?
- A. Keeping data local means that all privacy risks are eliminated completely, so no further checks are needed.
- B. Keeping data local reduces some risks, but shared updates can still leak information, so safeguards may be needed.
- C. Federated learning cannot be used in the European Union, so the company must choose a different method for its training.
- D. Federated learning removes the need for data governance, because the raw data never leaves the device.
Our reading (answer B): Federated learning reduces data centralisation but does not remove privacy risk, since updates can reveal information about the local data. Additional safeguards are often considered, and the right design depends on the threat model.
Check: Your Domain III notes on federated learning and privacy.
Q30 · HardA retrained model is ready to deploy the same day. Which governance step is most likely missing?
- A. A new logo on the product screen, so that users can see clearly that the system has been updated properly and safely.
- B. Notifying the press of the release on the same day, so that the change is publicly visible early.
- C. Deleting the old versions immediately, so that the team does not confuse the two models later on.
- D. Publishing the new model's weights openly, so that outside researchers can check the change quickly.
- E. Re-validation against defined acceptance criteria, with the change documented and impact reassessed as needed
Our reading (answer E): A retrained model is a change to the system, so it needs validation against agreed criteria, documentation of what changed, and a check on whether the impact assessment still holds. Deleting old versions removes the audit trail.
Check: Domain III notes on change management and re-validation.
Q31 · HardWhich pairing of TEVV terms is most accurate in common usage?
- A. Verification asks whether the system is the right one for its purpose; validation asks whether it was built to specification.
- B. The two terms mean the same thing, so organisations can use either one interchangeably in their documents.
- C. Verification asks whether the system was built to specification; validation asks whether it fits its intended purpose.
- D. Validation is done only after the system has been retired from service, as a closing step in its life.
Our reading (answer C): In common usage, verification checks the system against its specification and validation checks fitness for the intended use. Some sources define the terms differently, so use the definitions in your course.
Check: Your TEVV glossary entries.
Q32 · Very hardAn agent can book purchases and access a company's shared drive. Which design control is most important before launch?
- A. Scope its permissions to the minimum needed, require approval for consequential actions, and log its actions.
- B. Give it full administrator rights so that it can work smoothly without interruptions or delays.
- C. Remove logging entirely to save storage space, since the agent's actions are already visible to everyone in the business.
- D. Allow it to change its own permissions whenever it decides that the task needs more access.
- E. Let it act first and report to a manager once a month, so that no task is delayed at all.
Our reading (answer A): Least privilege, approval for high-impact actions and logging are the core controls for an agent that can act. Full rights or self-granted permissions widen the blast radius of an error.
Check: Domain III notes on agentic architectures and least privilege.
Q33 · HardA team reweights training examples to reduce disparity before training begins. Which bias mitigation stage does this represent?
- A. In-processing
- B. Post-processing
- C. Monitoring
- D. Pre-processing
Our reading (answer D): Reweighting changes the data before the model is trained, which is pre-processing. In-processing changes the training objective or algorithm, and post-processing adjusts outputs after training.
Check: Domain III notes on bias mitigation approaches.
Q34 · Very hardAn attacker can infer whether a particular person's record was used to train a model by querying it. Which attack is this?
- A. Prompt injection attack
- B. Membership inference attack
- C. Label flipping attack
- D. Denial of service attack
- E. Training data poisoning attack
Our reading (answer B): Determining whether a specific record was in the training data is membership inference. Prompt injection manipulates instructions, label flipping corrupts training labels, and denial of service disrupts availability.
Check: Your Domain III notes on adversarial attacks on models.
Q35 · HardFor a rare-event fraud model in which 0.5% of transactions are fraudulent, why is accuracy a weak headline metric?
- A. A model that labels every transaction as legitimate scores about 99.5% accuracy while catching no fraud.
- B. Accuracy cannot be calculated for classifiers, so the metric is simply unavailable for this model.
- C. Accuracy is always higher for rare events, so a high score on a rare-event model shows that it is working well overall.
- D. Accuracy measures fairness across all customer groups, so it is the right headline metric to report to the board here.
Our reading (answer A): If 0.5% of transactions are fraud, labelling everything as legitimate gives 99.5% accuracy with no fraud detected. Precision, recall and related measures show what the model actually catches and misses.
Check: Your confusion-matrix notes; work through the arithmetic yourself.
Q36 · HardFor a high-risk system, which record is most directly associated with traceability of the system's operation over its lifetime?
- A. The marketing brochure for customers
- B. The vendor's invoice for the year
- C. Automatically generated logs
- D. The staff rota for the month
- E. The press release for the launch
Our reading (answer C): Logging over the system's lifetime supports traceability and investigation. Other records may be useful, but they do not record the system's operation. Check the record-keeping article for the exact requirements.
Check: Article 12 of Regulation (EU) 2024/1689 (record-keeping), current consolidated text.
Domain IV: Governing AI deployment and use
14 questions
Q37 · HardA reviewer approves 99% of AI recommendations within seconds and rarely overrides them. Which concern is most relevant?
- A. The system is accurate, so the reviewer's approvals are a sign that the model works well.
- B. Human oversight may be nominal, and automation bias may be present.
- C. The reviewer is saving time, which shows that the workflow is working as the team intended.
- D. The model needs more layers, so that it produces recommendations that are harder to doubt.
Our reading (answer B): Near-total approval within seconds suggests the review may not be meaningful. Oversight needs a reviewer who can understand, question and override the output, and override rates and review time are worth monitoring.
Check: Domain IV notes on human oversight and automation bias.
Q38 · HardA monitoring dashboard shows accuracy dipping 3% in one week and recovering the next. Which response is most proportionate?
- A. Pause the system immediately, because any drop in accuracy is a sign that something is badly wrong.
- B. Retrain the model immediately with fresh data, so the dip cannot happen again on the next day of operation.
- C. Delete the logs for that week to keep the dashboard clean and easy for managers to read during the monthly review.
- D. Record it, track the trend against agreed thresholds, and escalate if the pattern persists or breaches a limit.
- E. Ignore the dip completely, because a single week of lower accuracy is never worth recording at all.
Our reading (answer D): A single short fluctuation may be normal variation. A proportionate response records it and watches the trend against thresholds set in advance. Overreacting wastes effort and underreacting misses real drift.
Check: Domain IV notes on monitoring thresholds and escalation.
Q39 · Very hardA deployed assistant gave users harmful medical guidance. What should happen first?
- A. Retrain the model overnight with new medical content and fresh safety data, so that the harmful guidance is replaced before users see it again.
- B. Publish a public apology and continue operating the assistant as normal while the team investigates the matter.
- C. Contain the harm, for example by restricting or pausing the affected function, preserve logs, then run the incident process.
- D. Ask the vendor to delete the logs of the conversations, so that the harmful guidance is no longer visible.
Our reading (answer C): Containment comes first to stop further harm, followed by preserving evidence and running the incident process. Retraining before understanding the cause can hide what went wrong.
Check: Domain IV notes on incident response; your organisation's incident policy.
Q40 · Very hardA deployer moves an internal drafting tool into a new use: scoring customers for credit decisions. Which statement is most accurate?
- A. Nothing changes, because the tool itself is exactly the same product that it was before the change.
- B. The vendor takes on all responsibility for the new use, because the tool is the same product it always was.
- C. It only needs a new name on the internal list, so the team can record the change quickly.
- D. It needs no assessment, provided that the new users have accepted the updated terms of use.
- E. The new use may change the risk profile and obligations, so it should be reassessed before deployment.
Our reading (answer E): A new intended purpose can change the risk level and the duties that apply, so the change should trigger a fresh assessment. Responsibility does not transfer to the vendor just because the tool is the same product.
Check: Your Domain IV notes on change of use; the Act's provisions on intended purpose.
Q41 · HardWhich step is most important when procuring a third-party AI system for a sensitive use?
- A. Compare prices only, because the cheapest tool usually performs the same as the most expensive one in most real settings.
- B. Accept the vendor's marketing summary as the full evidence, since it is written by the people who know the product best and is updated regularly.
- C. Choose the newest model available, because newer systems are usually safer and more accurate for every use.
- D. Obtain documentation on intended use, data, evaluation and limits, and contract for change notices and incident reports.
Our reading (answer D): Due diligence needs evidence the buyer can review, plus contractual rights to be told about changes and incidents. Price, marketing and novelty are not substitutes.
Check: Domain IV notes on procurement and vendor management.
Q42 · Very hardA loan applicant is refused after an AI-supported assessment. Depending on the applicable regime, which information is most relevant to provide?
- A. Notice that automated processing was used, meaningful information about the logic involved, and a route to challenge it
- B. The model's full source code, so that the applicant can check every line of the calculation for errors personally
- C. The names of the engineers who built the model, so the applicant knows who to contact about the decision
- D. The vendor's pricing schedule, so that the applicant can compare the cost of the product with alternatives on the market
- E. A copy of the training dataset with all personal records included, so the applicant can verify the outcome
Our reading (answer A): Regimes such as GDPR give individuals rights to know that automated decisions were involved and to contest them in some cases. Source code and staff names are not usually what the applicant needs. The exact duties depend on jurisdiction.
Check: GDPR Articles 13 to 15 and 22; the rules in your jurisdiction.
Q43 · HardWhy might an organisation keep operational logs for a high-risk system?
- A. To support traceability, incident investigation and post-market monitoring, within data minimisation rules
- B. To sell them to advertisers, who can use the records to target the people in the logs more precisely
- C. Logs are never needed for AI systems, so they can be deleted as soon as the system goes live into production use
- D. To avoid all regulation, since keeping records is the only obligation that the Act places on the operator
Our reading (answer A): Logs serve traceability, investigation and monitoring. Retention must still respect data minimisation and any other applicable rules, so the purpose should be defined and limited.
Check: Article 12 and the data protection principles in your notes.
Q44 · HardStaff are pasting customer records into an unapproved public AI tool. Which response is most effective?
- A. Ban the whole category of tools with no approved alternative, so that staff have nothing at all to use for their work.
- B. Provide an approved alternative, set clear data rules, add technical data-loss controls, and train staff.
- C. Fine each employee individually for every single use of the tool, so the message is clearly understood.
- D. Do nothing for now, because the tool is probably safe and the company has no evidence of any harm.
- E. Ask staff to delete their chat history each week, so that no record of customer data remains anywhere.
Our reading (answer B): Bans without alternatives tend to push use out of sight. Combining an approved option, clear rules, technical controls and training addresses both the need and the risk.
Check: Domain IV notes on shadow AI and data protection controls.
Q45 · HardA chatbot offers information to people who may be in a mental health crisis. Which governance step is most relevant?
- A. Add a footer saying the assistant is not a doctor, so that users understand what it cannot do for them.
- B. Assess risks with relevant expertise, design escalation to human support, and test responses to crisis-related inputs.
- C. Reduce response quality across the board to limit liability, since shorter answers carry less risk overall.
- D. Remove the chatbot's name so that users cannot form an attachment to it or trust it too much in a difficult moment, whatever it says.
Our reading (answer B): A footer does not manage a foreseeable risk. Expert-informed risk assessment, escalation paths and testing with crisis-related inputs are the controls that address it.
Check: Domain IV notes on vulnerable users and risk assessment.
Q46 · HardA system's fairness was checked at launch. Its user population has since changed. What is the most appropriate step?
- A. Assume fairness continues unchanged, since the launch testing already covered the users at that time.
- B. Remove demographic data from the logs, so that fairness can no longer be measured in any way.
- C. Monitor outcomes for the actual deployed population and compare them with the launch baseline.
- D. Wait for a complaint before acting, because complaints show exactly where the real problems are.
Our reading (answer C): Launch testing describes the population at launch. A change in users can change outcomes, so outcomes should be monitored for the population actually served.
Check: Domain IV notes on fairness in deployment.
Q47 · HardA deployer wants to detect whether a system drifts in a way that affects one group more than others. Which signal is most useful?
- A. The overall average score only, which is simple to report to the board
- B. The number of staff who logged in that week, which shows how busy the team has been
- C. The model's version name, which shows when the most recent update was released
- D. Outcome rates for that group over time, compared with a baseline
- E. The number of complaints from all groups, counted once a year at the year end
Our reading (answer D): Group-level outcome rates over time, against a baseline, show whether the effect is changing. Overall averages can hide a group-specific shift, and version names and login counts say nothing about outcomes.
Check: Domain IV notes on monitoring for group-level effects.
Q48 · Very hardA vendor silently updates its model, changing outputs for your users. Which contractual and governance measures best address this?
- A. Nothing at all, because updates from a vendor are always beneficial to the users in the long run
- B. Require the vendor never to update the model, even when a security fix is urgently needed
- C. Ask for a bigger logo on the screen, so that users know which company is responsible for the output
- D. Change-notification terms, a defined trigger for re-validation, and rollback options.
Our reading (answer D): Silent changes can alter risk without anyone reviewing them. Notice terms, re-validation triggers and rollback give the deployer the information and control to respond. Blanket freezes may be impractical.
Check: Domain IV notes on vendor change management.
Q49 · Very hardAn agentic system can carry out multi-step tasks across several systems. Which oversight question matters most at deployment?
- A. How fast it responds to each request, measured in seconds from the moment the request arrives
- B. What colour its interface is, so that the screens match the company's brand guidelines and the design system used by every team.
- C. How many tokens it uses for each task, which the finance team can then use to budget costs
- D. How many users can log in at the same time without the service slowing down noticeably
- E. Which actions it can take on its own, which need approval, and how its actions can be stopped and reviewed
Our reading (answer E): For agents, the critical questions concern autonomy boundaries, approval points and the ability to stop and review actions. Speed, appearance and token use are operational details.
Check: Domain III and IV notes on agentic architectures; your oversight design.
Q50 · HardA deployment team asks who should evaluate whether affected users were properly informed that AI was used. Which answer is most appropriate?
- A. The deployer, with input from its legal team and from the provider of the system
- B. Nobody, because users should check for themselves whether the system is being used at all
- C. Only the regulator, since it alone has the authority to decide how users are informed
- D. The model itself, since the system can be asked to confirm what it tells people about its use
Our reading (answer A): Informing users is typically a deployer responsibility, supported by legal review and by information from the provider. Leaving it to users, or to the model, does not meet a governance standard.
Check: Deployer obligations in your notes and the transparency article.