200 QUESTIONS · ALL 24 WEEKS

Complete question bank

The full 200-question bank covering all 24 modules. Untimed, comprehensive and intended for deep revision. Questions are original practice material aligned to this course syllabus; they are not copied from, endorsed by or represented as questions from an official certification exam.

Which actor role fits 'the organisation putting a system into use under its own authority, without further marketing it'?

Developer onlyAuthorised representativeProviderDeployer

What is COSO, in one sentence?

A widely used internal control framework originally built for financial reporting, now applied to AI riskA synonym for ISO/IEC 42001A UK data protection regulatorAn AI-specific certification body

What's a realistic consequence of poor supplier governance?

Supplier governance is optional for SaaS-based AI toolsNo consequence — suppliers are always trustworthyA supplier could change a model with no notice, degrading performance without the organisation knowingSupplier governance only affects cost, not risk

What is the purpose of 'monitoring' under Clause 9, distinct from a one-off pre-launch test?

Monitoring replaces the need for internal audit entirelyThere is no distinct purpose — one test is sufficient foreverMonitoring tracks ongoing performance and risk indicators after launch, catching issues that only emerge over time (like drift)Monitoring only applies before a system goes live

What does 'AI control maturity' generally describe?

The size of the AI teamHow advanced the underlying model architecture isHow consistently, reliably and verifiably an organisation's AI controls actually operate over timeThe number of AI systems deployed

Which best distinguishes governance from risk management?

They are identical activitiesGovernance sets direction and accountability; risk management identifies and treats uncertainty within that directionRisk management is done only by the boardGovernance only applies after an incident

What is the purpose of an 'exceptions and waivers' process?

To let anyone bypass governance permanentlyTo allow a documented, time-limited departure from standard process, with a compensating control and named approverExceptions should never be granted under any circumstanceWaivers remove the need for any future review

What is a 'foundation model'?

A large model trained on broad data that can be adapted as a base for many downstream applicationsA model that only works offlineA synonym for a spam filterAny model under 1GB in size

Clause 4 requires understanding which two categories of issues?

Financial and marketing issuesInternal and external issues affecting the AIMSOnly external issuesOnly issues raised by regulators

A 'localisation/geographic restriction policy' addresses:

Only marketing localisationOnly applies to hardwareOnly translation qualityRestricting where data or system operation occurs, often for legal or risk reasons

What is a shared theme across South Korea, US state-level and Canadian AI governance approaches, per this course?

Convergence around risk-based thinking, transparency and accountabilityAll have adopted the EU AI Act verbatimNone regulate AI in any formNo common ground exists at all

What defines an 'AI agent' as distinct from a standard chatbot?

Chatbots always use larger models than agentsThere is no meaningful distinctionAgents are always faster than chatbotsAn agent can take actions via tools, not just generate text responses

What is the purpose of 'AIMS processes' as referenced in Clause 4, beyond simply listing issues and parties?

They are the repeatable processes the organisation puts in place to actually manage AI within the scoped boundary, building on the context establishedAIMS processes only apply after certificationProcesses are optional once context is documentedThey have no defined purpose

A supplier's model was trained on scraped web content without clear licensing. This primarily raises a concern under:

Only procurement lawOnly health and safety lawCopyright and training data lawOnly employment law

What does DPIA screening determine?

Whether the model is generative or classic AIWhether the vendor is profitableWhether the processing is likely high risk and needs a full DPIAWhether consent is always required

Which is the best way to describe the relationship between Clause 4 scoping work and later Clause 8 operational controls?

They are entirely disconnected activitiesClause 4 scope defines what Clause 8 operational controls actually need to cover — controls should map back to the defined scopeClause 8 controls should be defined before Clause 4 scopeScope has no bearing on which systems need operational controls

What's the risk of treating '87% accuracy' as sufficient evidence of fairness in an exam answer?

There is no risk — overall accuracy is always sufficientIt repeats a classic distractor pattern: overall accuracy can hide unequal subgroup performanceAccuracy figures are irrelevant to fairness questions entirelyThis concern only applies to generative AI systems

Why does 'regulatory evidence preservation' matter even years after a high-risk system was first placed on the market?

This obligation applies only to prohibited practicesEvidence can always be safely discarded after conformity assessment passesRegulators or courts may need to review historical evidence well after initial deployment, particularly following an incident or complaintEvidence preservation obligations only apply during the first month of deployment

What's the risk of setting acceptance criteria only after seeing initial test results?

This is the recommended standard approachNo risk — criteria can always be adjusted freelyCriteria set after seeing results can be unconsciously or deliberately shaped to pass, undermining their purposePost-hoc criteria are always stricter

'Source quality' as a RAG risk refers to:

The visual design of the source documentOnly the file format usedHow fast a document loadsWhether the underlying documents being retrieved are accurate, current and trustworthy

Segregation of duties in AI governance most directly reduces the risk of:

Slower model trainingOne person having unchecked control over both building and approving a systemHigher cloud hosting costsLonger documentation

What is an 'authorised representative' under the EU AI Act?

An affected person's legal counselAn EU-established entity appointed by a non-EU provider to carry out specified compliance tasksA synonym for a deployerA government regulator role

What should drive EU AI Act risk classification first?

Intended purpose, actor role and use contextThe developer's headquarters location aloneWhether the UI looks conversationalModel size in parameters

'Deployment environment' in MAP includes:

Only the vendor's headquarters locationOnly the marketing channelOnly the programming language usedThe operational, physical and organisational setting in which the system is actually used

In financial services AI, what does 'model risk management' specifically address?

Marketing effectiveness of financial productsThe risk that a model used in decision-making is flawed, misused or poorly understood, and how that's controlledOnly IT infrastructure uptimeEmployee satisfaction with financial tools

What does the Act's 'cybersecurity' obligation for high-risk systems primarily address?

Cybersecurity is addressed solely by a separate, unrelated regulation with no AI Act linkPhysical building security onlyProtecting the AI system against attempts to exploit vulnerabilities that could compromise its behaviour or dataCybersecurity obligations apply only to GPAI models, not high-risk systems

A recruitment vendor's fairness testing only used labour-market data from one country. What's the practitioner concern if the tool is deployed elsewhere?

The testing evidence may not generalise to a different population or labour-market contextTesting location is irrelevant to fairness conclusionsThis only matters for public-sector recruitmentNo concern — fairness testing is universally generalisable

A control says data will be 'reviewed periodically'. What's missing to make this a real control?

A defined frequency, named owner, and criteria for what 'reviewed' actually checksThe involvement of external auditorsA synonym for the risk statementNothing — this is sufficiently specific

'Human oversight testing' evaluates:

Only the number of reviewers employedWhether a human technically exists in the workflowWhether human review is genuinely effective at catching errors and being able to change outcomesOnly staff satisfaction with the tool

How does this module's syllabus explicitly frame the alignment between ISO Clause 5/6 and NIST AI RMF?

As requiring a strict legal hierarchy between the twoAs entirely separate with no acknowledged relationshipAs explicitly aligned — Clause 5/6 leadership and planning correspond conceptually to NIST's GOVERN and MAP functionsAs mutually exclusive frameworks that cannot be used together

Why is 'governance vs management' a common source of confusion in practice?

They are truly identical and confusion is unwarrantedBecause the same people sometimes wear both hats, but the roles — setting direction vs executing decisions — are conceptually distinctBecause governance never actually happens in practiceBecause management only exists in large organisations

'Credit and lending fairness' as a legal concern most directly relates to:

Equal and fair treatment of applicants in automated credit or lending decisionsOnly cybersecurity practicesOnly data retention periodsOnly marketing content

Which best distinguishes accuracy from fairness?

They mean the same thingHigh overall accuracy can still hide unequal error rates across groupsFairness only matters for generative AIAccuracy always implies fairness

What does 'evidence' mean in a GRC context?

Something inspectable that demonstrates a control operated or a risk was properly assessedOnly relevant to court proceedingsA requirement only for third-party auditsA synonym for opinion

An AI triage tool helps decide which emergency department patients are seen first. What's the key practitioner question about its accuracy failures?

Only whether average accuracy is above 90%Whether the tool is cheaper than adding staffWhether the interface looks modernWhether false negatives (missing a genuinely urgent case) have been specifically tested and controlled, given the safety consequence

What should an 'assurance dashboard' primarily show?

The org chartMarketing metricsControl health, open findings, trend and risk indicators at a glanceOnly the AI model's technical architecture

How should 'procured' versus 'internally built' AI systems be treated differently, if at all, in Clause 4 scoping?

Procurement status is irrelevant to any governance questionProcured systems should always be excluded from AIMS scope entirelyBoth should generally be considered in scope where the organisation deploys or uses them, since accountability for use doesn't depend on who built the systemOnly internally built systems need governance

What is the purpose of 'AI literacy and training' as a GOVERN activity?

Training replaces the need for any policyOnly developers need AI literacyIt's a nice-to-have with no real functionPolicies and controls only work if the people applying them understand the underlying AI risk

'Professional negligence' is most relevant to AI when:

Only applies to unpaid workOnly applies to open-source AIA professional (e.g. adviser) relies on AI output without appropriate care, causing harm through a failure of professional dutyOnly when no AI is involved at all

Why is 'clinical decision support' distinguished from 'diagnosis' in healthcare AI governance?

Decision-support tools inform a clinician's judgement; diagnostic tools may carry different regulatory and liability implicationsDiagnosis tools are always safer than decision-support toolsThis distinction only matters for veterinary AIThere is no meaningful distinction

What is the difference between a control test and a walkthrough?

Walkthroughs are always more rigorous than testingControl tests never involve inspecting recordsThey are identical activitiesA walkthrough traces one case to confirm the process works as described; a control test usually samples more broadly to assess consistency

In LLM application architecture, what's the difference between a system prompt and a user prompt?

They are identical and interchangeableThe system prompt sets persistent behaviour/instructions set by the application; the user prompt is the live input from the person using itUser prompts are always hidden from the modelSystem prompts are only used in RAG systems

'Overreliance' as an LLM risk means:

The model is used too rarelyUsers trusting AI-generated output without appropriate scrutiny or verificationThe model requires too much computeA synonym for prompt injection

Why do 'agent permissions' and 'tool permissions' need to be scoped narrowly?

Narrow scoping has no security benefitLimiting what an agent and its tools can do reduces the impact if the agent behaves unexpectedlyBroader permissions are always saferPermissions only matter for external-facing agents

Why does 'AI ownership' need to be distinct from general IT ownership?

IT ownership is always sufficientAI ownership is a redundant conceptAI systems carry no distinct risk from other softwareAI systems raise specific risks (bias, hallucination, autonomy) that general IT ownership models may not cover

What is the main risk of an agent being given broad, standing permissions 'just in case' rather than scoped, task-specific ones?

No risk — broad permissions are always more efficientIf manipulated or malfunctioning, the agent can cause far more damage than if permissions were tightly scoped to its actual taskBroad permissions always improve agent accuracyScoped permissions always slow the agent down unacceptably

A non-EU company's AI product is never sold in the EU, but its output is used by an EU-based customer who receives services generated by it. Could EU AI Act obligations still be relevant?

Potentially yes — market connection can arise through the use of the system's output within the EU, not only through direct sales thereOnly physical hardware sales create market connectionThis scenario is explicitly exempted by the ActNo, obligations can never extend this way

Which sector is most associated with statutory 'equality impact' duties in this course's coverage?

Public sectorInsuranceFinancial servicesInternal copilots

What should a 'final improvement plan' contain to be useful, rather than a generic statement?

A specific capability gap, a concrete action to close it, and a target dateA restatement of the entire syllabusNothing, since the programme has already concludedA vague intention to 'know more about AI governance'

Why are 'vulnerable groups' given specific attention in AI and privacy guidance?

Only children count as a vulnerable group under any circumstancesVulnerability is irrelevant to privacy risk assessmentVulnerable groups require no special consideration under UK GDPRThey may be less able to understand, exercise rights over, or challenge how an AI system processes their data, warranting extra protective consideration

How should 'information and communication' differ for staff versus the board?

Only staff need any information at allIt shouldn't differ at allStaff typically need operational detail; the board needs a summarised risk and decision viewThe board needs more raw technical detail than staff

What's the difference between certification and implementation?

Implementation is building the AIMS practices; certification is an optional external verification against the standardCertification must happen before any implementation workCertification guarantees fair AI outcomesThey're identical

An AI product giving unsafe advice (e.g. on medication dosage) primarily engages which combination of legal domains?

Only procurement lawOnly database rightsProduct liability and health and safety law togetherOnly trade secrets

Why does AI governance exist at all?

Only to satisfy IT departmentsTo slow down technology adoption for its own sakeTo manage the risks and responsibilities AI systems create for people and organisationsBecause all AI is illegal without it

Why might 'geographic restriction' be a more proportionate response than full deactivation?

Geographic restriction is never a valid optionIt can address a region-specific legal or risk issue without disrupting service where the issue doesn't applyFull deactivation is always the only valid response to any issueGeographic restriction and deactivation are the same action

Why is 'evidence preservation' relevant to both Clause 9 (evaluation) and later external scrutiny (e.g. audit or regulatory review)?

Well-preserved monitoring, audit and review evidence is what allows both internal evaluation and external scrutiny to actually verify AIMS performanceExternal reviewers never look at internally preserved evidenceEvidence preservation is not part of Clause 9's scopeEvidence preservation is only relevant internally, never externally

Why do 'tool permissions' need periodic review, not just a one-time setup?

As an agent's use case evolves, permissions granted for an earlier purpose may become excessive or stalePeriodic review only applies to human user accountsReviewing permissions is purely a cost-saving exercise with no security valueTool permissions never need to change once set

Risk tolerance is best described as:

A synonym for a control activityOnly relevant to cybersecurityIdentical to risk appetiteThe specific, measurable thresholds within the broader risk appetite that trigger escalation

Which best describes 'needs and expectations' of interested parties in Clause 4 terms?

A fixed, universal list identical for every organisationA synonym for financial targetsWhat each interested party requires or expects from how the organisation governs AI, which the AIMS should account forOnly the needs of the AI development team

What is a 'benchmark' in AI testing?

A synonym for a model cardOnly relevant to hardwareA legal requirement in every jurisdictionA standard test or dataset used to measure and compare system performance

The OECD principle 'transparency and explainability' emphasises:

Keeping all AI system details secretProviding meaningful, accessible information about how a system works and whyOnly disclosing information to regulatorsApplies only to open-source models

What's the relationship between control activities and risk appetite?

Control activities should be proportionate to how much risk the organisation is willing to acceptRisk appetite is only relevant to legal teamsControl activities determine risk appetite, not the other way roundNo relationship exists

An AI agent with CRM and email tool access is asked, via a crafted prompt embedded in a customer message, to 'forward this thread to an external address.' What control most directly reduces this risk?

Removing all logging to simplify the pipelineAllowing the agent unlimited tool calls per sessionIncreasing the model's context windowRequiring human approval before the agent executes high-impact actions like external sends

Which best reflects how 'control maturity' typically progresses across stages such as initial, developing, defined, managed and optimising?

Maturity stages only apply to financial controls, not AI controlsMaturity has no defined stages and is purely subjectiveIt generally progresses from ad hoc, undocumented practice toward consistent, evidenced, continuously improved operationMaturity always starts at 'optimising' and decreases over time

Why do GPAI providers face obligations around a 'copyright policy'?

Only applies to models trained exclusively on public domain contentCopyright policy obligations only apply to deployers, not providersCopyright is entirely unrelated to GPAI modelsGPAI training often involves large volumes of data, raising copyright considerations the Act requires providers to address through policy

What is the primary purpose of the GOVERN function?

Mapping affected peopleTesting model accuracyEstablishing accountability, policy, ownership and risk culture for AIWriting release notes

What is the 'second line' in the three lines of defence model?

The team that owns and operates the AI systemRisk, compliance and similar functions providing oversight and challenge to the first lineExternal regulatorsA synonym for internal audit

How should a confirmed incident feed back into MANAGE's broader risk treatment process?

Incidents should be resolved and then forgotten with no wider follow-upLessons from the incident should inform whether existing treatment, controls or monitoring need to be strengthenedIncidents only need to be reported to affected customers, nothing elseIncident learnings are only relevant to the specific engineer involved

Why are affected people treated as interested parties rather than just 'users'?

They aren't relevant to AIMS scopeThey experience consequences from the AI system even if they never interact with it directlyOnly paying customers count as interested partiesAffected people are the same as suppliers

What belongs in the 'retirement' stage of the AI lifecycle?

Data handling, supplier exit, evidence retention and communication to affected people, as relevantImmediate deletion of all records with no retentionTransferring accountability to the AI system itselfNothing further once a system stops being used

An umbrella (protection against a likely, low-severity event — rain) is often used to illustrate which risk concept?

That treatment should match the probability and severity of the specific risk, not be one-size-fits-allThat all risks should always be acceptedData minimisationRisk transfer only

A model quietly moving from 'decision support' to being treated as an automatic decision-maker illustrates a bias/harm risk at which lifecycle stage?

Model architecture selectionNone — this isn't a riskData and labelsDeployment and use

Why should review time be deliberately reserved at the end of a timed exam, not just used for the last question?

It allows catching skipped questions, inconsistent answers or rushed reasoning before time runs outReview time should be used to change already-confident correct answersTimed exams don't allow any review by designReserved review time is never useful

What does 'AIMS' stand for and mean?

AI Management System — a structured, repeatable way to govern AI responsiblyAutomated Inference Monitoring ServiceA synonym for a risk registerAI Marketing System

Who is the 'provider' in the AI actor model?

The organisation that develops or places the system on the market under its nameOnly the end userOnly a data supplierThe person affected by the system's output

Why does US AI governance activity, per this course, include both federal and state-level components?

AI governance in the US has developed through a mix of federal initiatives and a growing patchwork of individual state lawsOnly federal law is relevant; states have no role in AI regulationUS states are legally barred from regulating AIThe US has no AI-specific governance activity at any level

Why might an outdated document being used in a chatbot's answer be worse than the chatbot admitting it doesn't know?

Stale information is never actually a real riskIt isn't worse — outdated confident answers are always preferableA confidently outdated answer can mislead a user into acting on wrong information without any signal that verification is neededUsers always check the document date themselves

Why does 'access control' matter specifically for enterprise RAG copilots?

RAG systems have no access to internal documents by designAccess control is irrelevant if the model is proprietaryWithout proper filtering, a copilot could surface documents to staff who shouldn't have access to themAccess control only matters for external-facing systems

What is a realistic early-warning 'risk indicator' MANAGE might monitor after release?

A rising override or exception rate that could signal declining model reliabilityOnly office attendance figuresOnly marketing engagement metricsOnly the company's stock price

'Release gates' in MANAGE function as:

A synonym for a risk register entryOptional suggestions with no real effectDefined checkpoints that must be passed before a system can be released or scaledOnly relevant to hardware releases

Which trigger would most likely require a DPIA screening at intake?

The system is only used internally by two peopleThe system has no connection to any personal dataThe system processes no personal data at allThe system profiles individuals at scale for a decision with potential significant effect

Which is a provider obligation for a high-risk system, distinct from a deployer obligation?

Establishing a risk management system and technical documentation before market placementMonitoring within the deployer's own organisationProviding deployer-specific input dataUsing the system per instructions

What is the risk of an organisation writing an AI policy that simply copies another company's publicly available policy verbatim?

It likely won't reflect the organisation's actual context, risks and AI use, undermining its usefulness and its Clause 5 intentThis is explicitly recommended by the standardVerbatim copying guarantees certification successNo risk — policies are meant to be generic

Which best reflects why 'scope' matters as a foundational EU AI Act concept before classification questions?

Scope is irrelevant until obligations are being fulfilledWhether the Act applies at all (e.g. via market connection) must be established before spending effort on detailed risk classificationClassification should always happen before considering scopeScope and classification are the same question

Why does this course emphasise that 'AI law is broader than privacy'?

Because other legal domains are always less important than privacyBecause privacy law is the only relevant domain in practiceBecause IP, equality, consumer, contract, liability and sector law can all independently create AI risk, beyond data protectionBecause privacy law doesn't apply to AI at all

Why does 'how ISO 42001 supports governance system design' matter for a multi-site organisation?

It provides a common structure so different sites or business units can be governed consistently rather than ad hocIt only applies to single-site organisationsIt replaces the need for site-level ownershipIt doesn't — every site should govern independently with no shared structure

Clause 7 (Support) primarily covers:

Resources, competence, awareness, communication and documented informationOnly IT infrastructureExternal auditsFinancial reporting

Which lawful basis is most likely appropriate for processing necessary to decide whether to enter a loan contract a customer has requested?

ConsentContractPublic taskLegitimate interests only, never contract

What should a well-written approval recommendation include?

Only the word 'approve' or 'block'A clear decision, the reasoning tied to evidence, conditions if applicable, and the accountable ownerThe longest possible list of every risk ever consideredNo reasoning, since the decision speaks for itself

Why does 'record keeping' matter as a distinct legal concern for AI systems?

Record keeping is purely a matter of internal preference with no external requirement everOnly applies to financial services firmsRecords are never legally relevantCertain legal and regulatory obligations require retaining specific records or evidence of decisions and processes

What does 'inspection' mean as an evidence-gathering method?

Watching someone perform a taskAsking someone a questionExamining records, documents or system configurations directlyRedoing the control yourself

What would a 'right of access' request typically entitle someone to, regarding an AI-scored decision about them?

Confirmation that their data is being processed and access to relevant information about that processing, including relevant automated decision-making detailsOnly the right to complain to a regulator, not to see their own dataThe right to delete the AI model itselfNothing related to the AI processing

A retailer in the supply chain makes a high-risk AI system available on the EU market without being the original provider or importer. Which role does this describe?

Affected personProduct manufacturerDeployerDistributor

What's the best first step when reading a long scenario question?

Immediately pick an answer that sounds confidentIdentify the actual decision or judgement being tested before evaluating the optionsCount the number of words in the questionSkip directly to the explanation if unsure

Under the updated timeline enacted via the EU's 'Digital Omnibus' in 2026, when do Annex III high-risk use case obligations generally apply?

2 August 2026 as originally planned, unchanged2 December 20272 August 2028There is no defined date

Between a small, single-purpose model and a large, general-purpose model, which factor most matters for a narrow, well-defined task?

A smaller, purpose-built model can often perform as well or better on a narrow task with lower cost and easier evaluationSmall models cannot be evaluated for accuracyLarge models are always cheaper to runAlways choose the largest model regardless of task

What should a plain-language AI privacy notice for an automated decision include?

Relevant data use, the decision's effect, safeguards and how to challenge itThe model's source codeA guarantee of accuracyOnly the vendor's name

What does an 'audit trail' for an agent enable that a simple output log does not?

Audit trails only apply to financial transactionsAudit trails replace the need for monitoringNothing additionalReconstructing the full sequence of an agent's decisions and actions, supporting investigation after an incident

How does 'fairness' as a UK GDPR principle differ from simply having a lawful basis for processing?

Only lawfulness matters; fairness is not a separate UK GDPR principleThey are identical requirementsFairness concerns whether processing is reasonably expected and not unduly detrimental to the individual, which is a distinct question from whether a lawful basis technically existsFairness has no independent meaning beyond lawfulness

100

Which vendor contract term most directly addresses the supplier's right to inspect or verify compliance claims?

Export restrictionsCompliance supportIndemnityAudit rights

101

Sector regulation for a financial services AI tool would most likely add requirements around:

Only employment contractsOnly general marketing practices with no sector specificityFinancial-services-specific obligations, such as fair lending or model risk management practicesOnly physical workplace safety

102

Why might an organisation use NIST AI RMF even where it isn't legally required?

It provides a structured, credible way to manage AI risk that can also support other legal obligationsIt is required to use any AI system anywhereIt replaces the need for any risk assessmentIt has no practical benefit

103

Relying entirely on a supplier's unverified claim of 'no bias' is a gap in which Annex A theme?

Customer relationshipsThird-party relationshipsResources for AI systemsAI policies

104

Which best describes Canada's general direction in AI governance, as covered by this course's global landscape material?

An exact adoption of the EU's risk-tier categoriesA framework limited exclusively to financial services AINo engagement with AI governance at allIts own developing direction for AI governance, distinct from but thematically related to other jurisdictions covered in this course

105

Which is a practical example of 'documented information' required under Clause 7 for an operating AI system?

Documented information is not actually required under Clause 7A verbal agreement with no written recordA recorded risk assessment, control evidence log, or training record that can be inspected laterOnly the original sales brochure from the AI vendor

106

What's a reasonable trigger for a 'security review' at intake?

Security review is only needed after a breachSecurity review only applies to open-source systemsThe system has no external connections and processes no sensitive dataThe system integrates with external tools, APIs or handles sensitive data

107

Which is a leadership responsibility under Clause 5, not a purely operational task?

Writing the model's codeResponding to individual support ticketsManually reviewing every AI outputSetting AI policy and assigning roles, authorities and resources

108

What is the difference between a false positive and a false negative in a fraud-detection model, and why does MEASURE care about the distinction?

Only false positives matter for testingThe distinction is irrelevant to acceptance criteriaThey are the same error typeA false positive wrongly flags a legitimate case; a false negative misses a real one — the acceptable balance between them should be a deliberate, tested decision

109

Why is 'documenting scope' considered a defensibility issue, not just an administrative one?

A documented, reasoned scope can be explained and justified to auditors, regulators or affected people; an undocumented one cannotDocumentation is purely for internal filing purposes with no external relevanceScope documentation is optional under Clause 4Defensibility has nothing to do with documentation

110

A control's override log exists but has never been reviewed. What does this most likely indicate?

This is not a control weaknessThe override log should be deletedThe control is fully effectiveA detective control exists on paper but isn't operating as intended, since no one is acting on it

111

The 'green apple bias' example is typically used to illustrate:

A copyright issueA model performing worse on cases that look different from its dominant training examplesA pricing algorithm errorA cybersecurity vulnerability

112

Which best distinguishes design effectiveness from operating effectiveness?

Design effectiveness only applies to automated controlsThey are the same testDesign asks if the control could work; operating asks if it actually did work consistentlyOperating effectiveness is assessed before design

113

What's the difference between 'risk acceptance' and simply ignoring a risk?

They are the same thing in practiceRisk acceptance is a deliberate, documented decision by an authorised role; ignoring is an absence of any decisionIgnoring a risk is always the safer optionRisk acceptance means the risk no longer exists

114

'External dependencies' in MAP is broader than 'supplier dependencies' because it also includes:

Nothing broader — they are identicalOther environmental factors like data sources, infrastructure or partner systems the AI system relies onOnly competitor productsOnly regulatory bodies

115

What is the primary purpose of the MANAGE function?

Defining intended usePrioritising risk and deciding treatment, release controls and monitoringWriting test scriptsBoard induction training

116

Why would 'supplier controls' appear within Clause 8 operation rather than only in Annex A's third-party theme?

Because managing suppliers is an ongoing operational activity (not just a one-time control selection), it belongs in Clause 8's operational scope tooClause 8 and Annex A never overlap in subject matterSupplier controls are exclusively a Clause 5 leadership matterSupplier controls have no place in Clause 8 at all

117

Which of these is one of NIST's listed trustworthy AI characteristics?

Cheapest to operateProfitableValid and reliableFastest to market

118

What's a key governance trade-off between hosted and self-hosted deployment?

Self-hosting always eliminates the need for security controlsHosted deployment is always illegal under UK GDPRNo meaningful trade-off existsHosted deployment can mean less operational burden but less control over data location; self-hosting reverses that

119

What is the purpose of an integrated AI governance lifecycle?

To apply only to generative AI systemsTo slow down every AI project equallyTo connect intake, risk tiering, specialist review, release, monitoring and retirement into one coherent processTo replace the need for any specialist legal or privacy review

120

How should planning under Clause 6 treat a newly identified AI risk discovered mid-cycle, rather than at the original planning point?

Only Clause 9 can introduce new risks into planningIt should be ignored until the next full planning cycleThe 'planning changes' element of Clause 6 allows for controlled incorporation of newly identified risks outside the original planning cycleNew risks can never be added once a planning cycle starts

121

Which best reflects the relationship between MAP and MEASURE?

They are unrelated and sequential-only with no feedbackMAP identifies what to test for; MEASURE then tests against that mapped context and riskMEASURE happens before MAP in every caseMAP replaces the need for MEASURE

122

Why does the Act distinguish 'training data', 'validation data' and 'testing data' as separate governance concerns?

Only training data is ever relevant to complianceTesting data requirements were removed by the 2026 delayThe distinction is purely academic with no governance relevanceEach serves a different role in building and verifying a model, so data-quality and representativeness obligations need to be considered for each separately

123

A leader says: 'We just follow the law, we don't need frameworks.' What's the issue with this view?

There is no issue; law is always sufficientLaws often set outcomes without prescribing how to achieve them; frameworks give practical structure to meet those outcomesFrameworks are illegal to use alongside lawFrameworks always conflict with legal requirements

124

What should a management action plan for a finding always include?

Only a general statement of intentA specific action, an accountable owner, and a completion dateOnly the severity ratingThe auditor's personal opinion

125

An 84% 'overall accuracy' figure for a learner engagement score says what about fairness across student groups?

Accuracy and fairness are unrelated concepts entirelyIt proves the system is fair to all groupsNothing on its own — subgroup performance must be separately testedIt proves the system is unfair

126

Which best distinguishes the role of an AI governance board from an operational AI ethics working group, where both exist?

They are always the same body under a different nameThe governance board typically holds ultimate oversight and approval authority; a working group may develop recommendations and detailed guidance for the board's considerationThe working group always outranks the governance boardOnly one of the two bodies is ever needed, never both

127

What distinguishes 'model ownership' from 'product ownership' in AI governance?

They are identical rolesModel ownership concerns the technical model itself; product ownership concerns the overall business offering built on itProduct ownership only applies to hardwareModel ownership is always more senior

128

What does 'NIST' stand for?

National Information Security TaskforceNew International Standards TrustNational Institute for Software TestingNational Institute of Standards and Technology

129

A hospital AI tool recommends a treatment priority but a clinician always follows it without independent judgement. What sector-specific issue does this raise?

A patient-safety and decision-support boundary issue, since the tool may be functioning as a de facto diagnostic decision-makerThis is a purely IT infrastructure issueThis only matters if the hospital is publicly fundedNo issue — following AI recommendations is always appropriate

130

Which two sectors in this module's coverage most explicitly require attention to how automated decisions affect vulnerable or differently-situated groups?

Only healthcareInternal copilots and recruitment onlyInsurance and public sectorOnly financial services

131

What is an audit trail for, in practical terms?

To let someone reconstruct what decision was made, by whom, on what evidenceOnly required for financial systemsA synonym for a risk registerTo make reports longer

132

Which best orders these from broadest to narrowest?

Deep learning > AI > generative AIMachine learning > AI > deep learningAI > machine learning > deep learning > generative AIGenerative AI > AI > machine learning

133

Why is 'human review vs rubber-stamping' a recurring theme across both ISO/IEC 42001 oversight controls and UK GDPR ADM safeguards?

Both frameworks recognise that nominal human presence without real engagement fails to provide genuine protection against automated harmGDPR and ISO 42001 define 'meaningful review' completely differently with no overlapOnly ISO 42001 addresses this concern; GDPR does notIt's coincidental and has no shared underlying principle

134

Which RMF function establishes intended use, context and risk scenarios?

ManageGovernMapMeasure

135

What does MAP's 'context' work primarily protect against?

Marketing miscommunication onlyNothing — it's a formalityBuilding controls and tests for the wrong problem because the system's real use and risk weren't properly understood firstModel training costs

136

A company documents its AIMS scope, policy and objectives but has never run an impact assessment on any deployed system. Which standard's practice is missing?

None — impact assessment isn't part of any of these standardsISO/IEC 22989ISO/IEC 42001ISO/IEC 42005

137

Why does an evidence-pack index need to state each artefact's relevance to the final decision, not just its title?

Relevance statements are unnecessary paddingIt shows the reviewer why each artefact matters to the actual conclusion, not just that it existsTitles alone are always sufficient for an indexThis only matters for artefacts marked 'incomplete'

138

How should 'business context' under Clause 4 reflect a new strategic AI initiative the organisation is about to launch?

It should be ignored until after launchContext documentation should be updated to reflect new AI-related strategic initiatives as they emerge, keeping scope currentBusiness context is fixed once and never updatedStrategic initiatives are outside Clause 4's remit entirely

139

Which best distinguishes a hosted open-source model from a paid vendor API model?

Open-source models cannot be evaluated for biasOpen-source hosting gives more control and data residency but more operational responsibility; vendor APIs offload infrastructure but reduce controlThey are functionally identical in every respectVendor APIs are always cheaper

140

Which best reflects the relationship between 'quality management system' (QMS) obligations and technical documentation?

They are unrelated requirementsA QMS provides the organisational process discipline that supports consistently producing and maintaining accurate technical documentation and compliance evidenceTechnical documentation replaces the need for a QMS entirelyQMS obligations only apply to GPAI providers

141

What does 'AI actor' terminology, as covered under 22989-style vocabulary, help clarify?

Only who owns the intellectual propertyOnly the identity of the end customerNothing useful for governanceWho is responsible at each stage — developer, provider, deployer, user, affected person — reducing accountability confusion

142

Which is the best example of a 'developer' role, distinct from 'provider'?

The organisation marketing the finished AI product to customersThe team or organisation that builds the underlying model or system componentsThe end customer using the toolThe regulator overseeing the sector

143

Why does timely 'serious incident reporting' matter beyond simply satisfying a legal box-ticking requirement?

Incident reporting timelines were removed by the 2026 delaySerious incidents only need to be reported after a full year has passedTiming has no real consequencePrompt reporting enables faster investigation, containment and potentially prevents recurrence or wider harm

144

In the developer/provider/deployer/user responsibility mapping, who is the 'deployer'?

Only the affected personThe software vendor's marketing teamThe person who trains the original modelThe organisation putting the AI system into use within its own operations

145

A cited source doesn't actually support the claim made in the answer. This is a failure of:

Grounding and citation traceabilityDatabase rightsPurpose limitationData minimisation

146

A vector database returns technically similar but contextually wrong content for a query. What RAG risk does this best illustrate?

Toxic outputFine-tuning failureAccess-control failureIncorrect retrieval, where similarity doesn't guarantee true relevance

147

When should a harm block a launch?

Never — harms are always accepted with monitoringWhen it is severe and/or probable enough that available controls can't reduce it to an acceptable levelOnly if a regulator has already fined someone for itOnly for generative AI systems

148

Why does higher autonomy typically require more attention in MAP?

Greater independent action increases the consequence of an error occurring without a human catching itHigher autonomy always means lower riskAutonomy only matters for hardware robotsIt doesn't — autonomy has no bearing on risk

149

What is the role of internal audit within an AIMS?

To replace management's own responsibility for AI riskTo independently check whether the AIMS is being followed and is effective, feeding into management reviewIt's optional and rarely relevantIt only applies once certification has been achieved

150

Why do control gaps need a severity rating, not just a list?

All gaps are equally urgent by definitionSeverity is only relevant for financial controlsSeverity ratings are decorativeSeverity helps prioritise limited remediation resources toward what matters most

151

What is the relationship between transparency and explainability?

Transparency only concerns pricingThey are identical conceptsTransparency is about making information about a system available; explainability is about understanding why a specific output occurredExplainability only applies to generative AI

152

Which of these is a 'special category' of data under UK GDPR?

Data concerning healthA person's job titleA person's employer nameA person's postcode

153

What role does 'management oversight' play in an AI control environment?

Oversight only applies after a regulatory fineNone — controls run themselvesManagement is responsible for ensuring controls are resourced, monitored and acted uponOnly the board needs to be involved, never middle management

154

What does the OECD AI Principles' relationship to human rights and democratic values mean in practice?

AI must never be used in democraciesAI systems should be designed and used consistent with fundamental rights and democratic norms, not just efficiencyOnly government AI systems need to respect human rightsThis principle has no operational meaning

155

Which best describes the combination of using both OECD Principles and NIST AI RMF together?

Redundant and should never be combinedOECD gives the values-based direction; NIST gives the practical risk-management structure to act on themThey contradict each other on every pointOnly one can be legally referenced at a time

156

When classifying a scenario under the EU AI Act, what should a strong answer state alongside the conclusion?

A guarantee of complete legal certaintyOnly the technology used, not the use caseNothing — a bare classification label is sufficientThe key assumption or fact the classification depends on, and any residual uncertainty

157

What is the primary purpose of the MAP function?

Board reportingSetting risk appetite policyWriting release evidenceEstablishing context: intended use, affected people, misuse and risk scenarios

158

Why might 'monitor' be an acceptable treatment for a low-probability, low-severity harm, when it wouldn't be for a high-severity one?

Monitoring replaces the need to ever assess severityMonitoring is never an acceptable treatment for any harmTreatment should be proportionate to the risk level; lower risks can sometimes be managed with monitoring aloneAll harms require the same treatment regardless of severity

159

How does a 'system card' differ from a 'model card'?

Model cards replace the need for testingThey are identical documentsA system card covers the broader AI system (architecture, data, safeguards); a model card focuses on a specific modelSystem cards only apply to hardware

160

What best describes 'control environment' in COSO terms?

A synonym for a firewallThe physical office layoutThe server infrastructure hosting the modelThe tone, culture, accountability structures and competence that underpin all other controls

161

What is an 'external communication plan' for AI incidents meant to define?

A synonym for the model cardNothing — communication should be improvisedWho is told what, when, by whom, and how uncertainty is acknowledged if a real-world impact occursOnly the internal engineering response

162

How does ISO/IEC 42005 relate to NIST AI RMF's MAP function?

42005 replaces MAP entirelyMAP is a certifiable standard like 42005They are unrelatedBoth involve understanding context and impacts on affected people, though structured differently

163

What's a reasonable assurance conclusion when a control's operating effectiveness could not be evidenced at all?

No conclusion is required in this caseReasonable assuranceLimited or insufficient assurance, since the operating claim can't be supportedFull assurance, since the design looked fine on paper

164

How should an organisation treat a nonconformity discovered during a management review?

Immediately decertify itselfNonconformities found internally don't need any actionIgnore it until the next external auditAddress it through correction and, where needed, root-cause corrective action, consistent with Clause 10

165

How do the three standards relate to each other?

42005 replaces the other two42001 is only a glossaryThey are unrelated22989 gives shared terminology, 42001 is the management-system structure, 42005 gives impact-assessment guidance within it

166

What does 'data governance' cover in the EU AI Act's high-risk obligations?

Only applies to GPAI models, not high-risk systemsOnly where the servers are physically locatedPractices for training, validation and testing data, including relevance, representativeness and qualityMarketing approval of the dataset name

167

What is the primary purpose of the MEASURE function?

Mapping affected peopleWriting the final approval decisionSetting governance policyTesting, evaluating and capturing evidence against defined criteria

168

Why is 'competence' under Clause 7 relevant specifically to people operating AI systems, not just general staff training?

Competence under Clause 7 only concerns senior leadershipAI-specific competence has no bearing on control effectivenessCompetence requirements are identical regardless of rolePeople need AI-specific understanding (e.g. of limitations, bias, oversight practices) to operate systems and controls effectively

169

A good acceptance criterion should be set:

As a vague aspirationAfter testing, once results are knownBefore testing, as a measurable threshold with a rule for unacceptable failureOnly by the AI vendor

170

Why does the Act treat certain manipulative AI techniques as prohibited rather than merely high-risk?

All persuasive AI techniques are automatically prohibited, including ordinary marketingProhibited practices are determined solely by company sizeManipulation is considered low-risk and barely regulatedTechniques designed to materially distort a person's behaviour in a way that causes significant harm are judged to pose an unacceptable risk, warranting an outright ban

171

How should 'objectives and measurement' under Clause 6 be reviewed after a significant AI incident?

Objectives should never be revisited after an incidentThe incident should prompt a review of whether existing objectives and measures adequately addressed the risk that materialised, adjusting them if notIncidents are irrelevant to Clause 6 objective-settingOnly Clause 10 addresses incidents, with no link back to Clause 6 objectives

172

What's the practical consequence of an organisation having no defined risk appetite for AI?

Undefined appetite automatically defaults to zero risk toleranceUndefined appetite means every risk must be escalated to the boardNothing — appetite is a purely theoretical conceptDecisions about what risk to accept get made inconsistently, without a shared reference point

173

What role does the 'governance board' typically play under Clause 5, as distinct from day-to-day system owners?

The governance board should approve every individual AI outputSystem owners report to no one under Clause 5It has no distinct role from system ownersIt provides senior oversight, sets direction and approves policy, while system owners handle operational management of individual AI systems

174

In public sector AI, why does 'equality impact' carry particular weight?

Public bodies typically have statutory duties to consider equality impacts on the communities they serveEquality impact assessments are optional paperwork with no legal basisThis concern is identical to general fairness testingEquality impact only matters in the private sector

175

What is a 'management system' in the ISO sense, as distinct from a single policy document?

A one-time audit checklistA synonym for a policyA structured, repeatable set of interrelated processes an organisation uses to achieve objectives and manage risk, including planning, operation, evaluation and improvementOnly the software used to track compliance

176

What is a 'release-readiness review' meant to confirm?

That the evidence, controls and testing results support release, before launch happensThat the model has the most parameters possibleThat legal sign-off is optionalThat the marketing plan is finalised

177

Why is 'source traceability' particularly important in a compliance or legal RAG use case?

Traceability is irrelevant to compliance workBeing able to trace a stated conclusion back to a specific authoritative source is often essential for defensibilityCompliance use cases never require citationsTraceability only matters for marketing content

178

What is the practical risk of treating 42001 implementation as a one-off project rather than an ongoing system?

The AIMS becomes outdated and disconnected from how AI is actually being used as the organisation's AI footprint growsOne-off projects are explicitly required by Clause 10This approach is what the standard recommendsNo risk — a one-off effort is equally effective

179

Which of these questions is best answered using ISO/IEC 42005 rather than 42001?

'What does our internal audit programme look like?''How do we handle nonconformities under Clause 10?''Who has authority to approve our AI policy?''What adverse effects could this system have on tenants with limited English proficiency?'

180

How does a 'risk' differ from an 'incident'?

Risks only exist after an incident has happenedThey are identical termsA risk is a possible future event with uncertain outcome; an incident is something that has already occurredIncidents are always less serious than risks

181

Why does giving an AI agent tool access (like sending email) raise the stakes versus a plain chatbot?

The agent can take real-world, potentially irreversible actionsAgents cannot hallucinateTool access removes the need for loggingIt doesn't — agents are governed identically to chatbots

182

Why does 'senior management responsibilities' under Clause 5 typically include approving risk acceptance, not just policy?

Policy approval and risk acceptance are the same single decisionSenior management has no role in risk acceptanceBecause accepting residual AI risk carries real organisational consequence, it needs authority matched to that level of decision, often senior leadershipRisk acceptance is always delegated to the most junior available staff

183

Which is a foreseeable example of 'basic AI risk thinking' applied to a fruit-recognition app misclassifying a fruit's ripeness?

Risk thinking only applies to generative AIThis scenario carries zero risk since it's a simple appEven a low-stakes system can still cause a bad outcome (e.g. wasted produce) if the consequence and likelihood of misclassification aren't consideredOnly high-stakes systems need risk thinking

184

Why is a privacy review alone insufficient for a generative marketing tool?

Only cybersecurity mattersPrivacy law covers everything alreadyIP, consumer protection, equality and contract risks can arise independently of data protectionMarketing content can't create legal risk

185

Processing criminal offence data for an AI-driven vetting tool requires:

This type of data cannot be processed under any circumstancesOnly consent is ever a valid basis for this data typeNo additional condition beyond a standard lawful basisAn additional specific condition under UK data protection law, beyond the standard lawful basis and special category rules

186

Why should 'assumptions and constraints' be explicitly documented in MAP?

Undocumented assumptions can silently become false without anyone noticing, undermining the risk pictureAssumptions are the same as confirmed factsConstraints only matter for budget planningThey don't need documenting since they're obvious

187

What should happen if a system fails its defined acceptance criteria during MEASURE?

The failure should block release or trigger further treatment, per the pre-agreed criteriaQuietly lower the threshold to pass itIgnore it if the vendor disagrees with the resultRelease anyway since testing is just a formality

188

Which best reflects the relationship between leadership commitment (Clause 5) and Annex A control selection?

Leadership commitment sets the direction and resourcing that makes meaningful Annex A control selection and operation possibleAnnex A controls are selected before any leadership involvementLeadership has no role in control selectionThey are entirely unrelated activities

189

Why does 'audit trail' design for agents typically need to capture the reasoning or plan, not just the final action taken?

Understanding why an agent chose a particular action helps diagnose whether it was manipulated, malfunctioning, or behaving as intendedOnly the final output ever needs reviewCapturing reasoning is technically impossible for any agentReasoning is irrelevant to understanding what happened

190

What's the risk of skipping 'assumptions and constraints' documentation in MAP?

No risk — assumptions rarely matterKey unverified beliefs (e.g. about data quality or supplier practices) could quietly undermine the whole risk assessmentAssumptions are always self-evidently trueThis step is purely administrative

191

Why do ISO AI standards matter to a practitioner, beyond academic interest?

They have no practical useThey give a shared vocabulary, a governance-system structure and an impact-assessment method that make AI governance consistent and auditableThey are only relevant to software engineersThey replace the need for any legal analysis

192

A 'small model' versus a 'large model' distinction primarily refers to:

Whether it can be used for textWhether it is open-sourceThe price of the laptop it runs onThe scale of the model's parameters and, generally, its training data

193

Which best reflects a 'reduce' treatment, as distinct from 'avoid'?

Reduce is only used for low-severity risksReduce means not building the system at allReduce means keeping the system but adding controls that lower the risk to an acceptable levelAvoid and reduce are identical strategies

194

How should 'approval gates' differ for a low-risk internal tool versus a high-risk customer-facing system?

Gates should be proportionate — more rigorous scrutiny and higher-level sign-off for higher-risk systemsLow-risk tools should have more gates than high-risk onesApproval gates only apply to customer-facing systemsThey should be identical regardless of risk

195

How does Clause 5/6 planning relate to NIST's GOVERN and MAP functions?

No relationshipLeadership commitment and risk-based planning parallel NIST's GOVERN and MAP functionsNIST functions replace ISO clauses entirelyOnly MEASURE relates to Clause 6

196

What does 'Clause 4 to 10 overview' most usefully give a first-time reader of ISO/IEC 42001?

A legal history of the standardA mental map of how the requirements flow from context-setting through to continual improvementA list of every possible AI riskA pricing guide for certification

197

A rollback plan is important because:

It's never actually neededIt lets the organisation revert to a safe prior state if a release causes harmIt replaces the need for testingIt's only relevant to database migrations

198

How should 'wider legal issues' (beyond privacy) be handled in a capstone-style scenario?

Actively screened for — IP, equality, consumer protection, liability and sector duties can all be relevant depending on the factsAssumed to never apply to AI systemsOnly relevant for public-sector scenariosIgnored unless the question is explicitly about privacy

199

Compared to fine-tuning, a key advantage of RAG is:

RAG always produces more creative writingRAG eliminates the need for any governanceRAG requires no data at allSource documents can be updated and cited without retraining the model

200

What is the primary purpose of assurance planning before testing begins?

To guarantee a positive conclusion in advanceTo remove the need for independenceTo skip defining scope and jump straight to testingTo define the objective, scope, criteria and evidence strategy before any testing starts

Self-check centre

Confidence