Free AIGP Practice Questions

Question 1

After a successful launch, a CTO tells the board that the company's demand-forecasting model has 'completed its life cycle' now that it is in production, so the governance team can be reassigned. As the AI governance lead, what is the most accurate correction?

Accepted Answer

The life cycle is iterative and continues through monitoring, retraining and eventual decommissioning, so governance obligations persist after deployment - The AI development life cycle runs through seven stages from planning and design to decommissioning, and it is iterative rather than linear - building AI is never a one-time process. Deployment is a stage, not the finish line: monitoring, retraining and eventual retirement all carry governance hooks. Option B invents an artificial split; post-deployment activity is part of the same life cycle, not a separate cycle. Option C fails because passing pre-deployment tests says nothing about drift, new risks or changing use cases that emerge in production. Option D is wrong because continuing the life cycle is an inherent operational reality, not something contingent on a regulator demanding it.

Answer

The life cycle ends at deployment, but a separate compliance cycle begins that requires a new team

Answer

The CTO is correct provided the model passed its pre-deployment testing

Answer

The life cycle only continues if regulators specifically require post-market monitoring for this system

Question 2

A vendor pitching a radiology image-analysis product tells a hospital procurement panel that its tool represents 'artificial general intelligence for diagnostics'. The hospital's AI governance adviser is asked to assess the claim. Which response is most accurate?

Accepted Answer

The tool is artificial narrow intelligence, because deployed real-world systems today are ANI and AGI does not currently exist - On the intelligence ladder, only artificial narrow intelligence (ANI) is fully realised today; AGI and ASI do not currently exist. A deployed, task-specific system - however impressive - is ANI, so the vendor's 'AGI' claim is marketing, not fact. Option A fails because no AGI systems exist in any market. Option B doubles the error: ASI would exceed human intelligence across essentially all domains, not just one task, and it does not exist either. Option D fails because processing multiple image types within one narrow diagnostic function does not make a system broad AI; the tool is still confined to a specific task domain.

Answer

The claim is plausible because AGI systems have recently entered regulated healthcare markets

Answer

The tool is artificial superintelligence, since it can outperform radiologists at the task

Answer

The tool is broad AI, because it processes multiple image types

Question 3

A Copenhagen grocery chain runs an unsupervised clustering project to segment its loyalty-card customers. Two data scientists working independently on the same dataset deliver different segmentations - one finds five customer groups, the other eight - and each defends their result as valid. The marketing director asks the AI governance lead how both can possibly be right. Which explanation is best?

Accepted Answer

Unsupervised learning searches for hidden patterns in unlabelled data with no predefined targets, so interpreting the resulting clusters is inherently subjective - Unsupervised learning trains on unlabelled data with no predefined targets: its goal is to surface hidden patterns, structures and relationships, and clustering - automatically grouping data points that share similar attributes - is one of its two classic sub-types. Because there is no ground truth to score against, the trade-off is that results are less predictable and their interpretation is subjective: two competent analysts can carve the same customers into different, equally defensible groups. Option A fails because unsupervised learning uses no labels at all - inconsistent labelling is a supervised-learning risk. Option B is doubly wrong: clustering sits inside unsupervised learning, not supervised, and reward functions belong to reinforcement learning, where an agent learns by trial and error. Option D invents a rule - semi-supervised learning combines a small labelled set with a large unlabelled one to cut labelling costs; it says nothing about producing more clusters.

Answer

The dataset's labels were applied inconsistently, so each scientist worked from a different ground truth

Answer

Clustering is a sub-type of supervised learning, and the two scientists used different reward functions

Answer

One scientist must have used semi-supervised learning, which always produces a larger number of clusters

Question 4

An internal auditor at a Lisbon fintech reviews the risk register for the firm's credit-scoring tool and finds a single scoped entry: 'the model'. She insists the register be re-scoped around the 'AI system' instead. What does that change of term actually add to the register's coverage?

Accepted Answer

It widens the register to the full operational environment - data, algorithms, models, interfaces and infrastructure - The four building-block terms nest inside each other, and the exam tests the boundaries precisely. Data is the raw information used for training; an algorithm is a computational procedure - a set of instructions and rules; a model is a program that applies algorithms to data so it can predict or decide from learned patterns; and the system is the full operational environment that wraps everything, including interfaces and infrastructure. Re-scoping the register from 'model' to 'system' therefore pulls in exactly the components where many real-world failures occur - data pipelines, integration points, and the interfaces through which users and other software touch the tool. Option A fails because a model is only one nested component of a system, never a synonym for it. Option B mistakes a single ingredient (infrastructure) for the whole. Option C is arbitrarily partial - interfaces are explicitly part of the system definition, and excluding them would leave a governance blind spot precisely where customers experience the tool.

Answer

Nothing - in AI governance, 'system' and 'model' are interchangeable terms

Answer

It narrows the register to the hardware infrastructure on which the model runs

Answer

It adds the training data to the register but deliberately excludes user-facing interfaces

Question 5

A ministry official in Wellington is building a national inventory of AI systems used across government agencies and proposes structuring it with the OECD Framework for the Classification of AI Systems. A colleague objects: 'That framework is a binding risk-tier law like the EU AI Act - adopting it would import foreign legal obligations.' How should the official respond?

Accepted Answer

Correct the colleague - it is a user-friendly, non-binding classification tool whose stated purposes include informing registries and inventories and supporting risk assessment, which is precisely this use case - The OECD Framework for the Classification of AI Systems is a user-friendly classification framework, not legislation. Its stated purposes are to promote a common understanding of AI, inform registries or inventories, support sector-specific frameworks, and support risk assessment and risk management - so a national AI inventory is a textbook application. It examines systems across five dimensions (mnemonic PEDMT): People and planet, Economic context, Data and input, AI model, and Tasks and output. Option A fails because the framework carries no mandatory tiers or legal obligations - that is the EU AI Act's territory, and confusing the two is the exact error the colleague made. Option C fails because economic context is only one of five dimensions, not the whole scheme. Option D fabricates a scope limit - the framework classifies AI systems generally and predates the generative AI boom; nothing restricts it to one model family.

Answer

Concede the point - the framework imposes mandatory risk tiers on any organisation that adopts it

Answer

Abandon the framework, because it classifies AI systems along a single dimension: the economic sector they operate in

Answer

Abandon the framework, because it applies only to generative AI models and the government inventory covers all system types

Question 6

A Munich insurer wants a tool that outputs, for each policyholder, the likelihood that they will cancel their policy at renewal - a simple stay-or-cancel event - so the retention team can rank whom to call first. The data science lead asks the AI governance adviser to sanity-check the proposed algorithm choice. Which choice is best matched to the job?

Accepted Answer

Logistic regression, because it is a probabilistic technique for estimating the likelihood of an event with a binary outcome, such as whether a customer stays or churns - Governance professionals are expected to recognise the core algorithms well enough to match them to jobs. Logistic regression is the probabilistic technique for estimating the likelihood of an event, usually a binary outcome - and customer stay-or-churn is the canonical example, which is exactly the insurer's renewal question. Option B is the tempting near-miss: linear regression predicts numeric outcomes from continuous input variables (think damage costs or prices), but the underlying event here is binary, and the tool purpose-built for event likelihoods is logistic, not linear, regression. Option C fails on its justification - neural networks handle highly complex tasks such as natural language processing and facial recognition, but nothing about probability estimation requires one, and reaching for a complex, harder-to-explain model without need creates avoidable governance cost. Option D belongs to a different family altogether: association rule learning is an unsupervised technique for finding relationships between data points (people who buy X also buy Y), not for scoring the probability of a future event.

Answer

Linear regression, because a cancellation likelihood is a continuous number and linear regression predicts continuous values

Answer

A neural network, because probability estimates can only be produced by architectures that mimic the human brain

Answer

Association rule learning, because it reveals which customers tend to buy which products together

Question 7

A dermatology AI trained almost exclusively on images of lighter skin tones performs poorly when diagnosing conditions on darker skin. Patients from underrepresented groups receive less accurate results. Which type of bias best explains the root cause?

Accepted Answer

Sampling bias, because the training data did not represent the population the system serves - Sampling bias arises when the training data fails to represent the real population the system will serve - here, the near-absence of darker skin tones in the dataset is the direct cause of the performance gap. Option A fails because temporal bias concerns data or relationships that change over time, and nothing in the scenario involves outdated data. Option B fails twice over: implicit bias is by definition unconscious, so 'consciously excluded' contradicts the term, and the scenario points to dataset composition rather than developer attitudes. Option D describes a real phenomenon - humans over-relying on automated outputs - but it concerns how users treat the system's results, not why the model itself underperforms.

Answer

Temporal bias, because the training images became outdated

Answer

Implicit bias, because the developers consciously excluded certain patients

Answer

Automation bias, because clinicians over-trusted the system's outputs

Question 8

A city deploys an AI-driven content recommendation system that gradually pushes residents into isolated echo chambers, distorting local political debate ahead of an election. A harms assessment must classify this impact. Which classification is most accurate?

Accepted Answer

Societal harm, because the damage is to the democratic process and participation - Societal harm is defined as damage to the democratic process and civic participation - and echo chambers that distort political debate are a textbook example, alongside mass surveillance and disinformation. Option A fails because group harm means discrimination against a definable population subgroup, and no subgroup is singled out here; the injury lands on democratic society as a whole. Option B fails because although feeds are individually personalised, the harm assessed is collective - the mechanism being individual does not make the harm individual. Option C fails because organisational harm concerns damage to a deploying organisation (reputational, financial, legal); the city's reputation is incidental to the core democratic injury described.

Answer

Group harm, because a specific population subgroup is being discriminated against

Answer

Individual harm, because each resident's feed is personalised

Answer

Organisational harm, because the city's reputation may suffer

Question 9

A Warsaw insurtech ships a claims-triage AI assembled from an open-source base model, an outsourced data-labelling vendor and a cloud provider's API. After a wave of wrongly denied claims, the CEO tells the regulator that responsibility is 'distributed across the supply chain', so no single company can fairly be held responsible. Which trustworthy-AI attribute does this position violate?

Accepted Answer

Accountable - the organisation delivering the AI is ultimately responsible for it, irrespective of how many contributors were involved - Trustworthy AI is characterised by three traits: it is human-centric (amplifying human agency and improving the human condition), accountable, and transparent, operating in an expected, legal and fair manner. The accountability trait is defined pointedly for exactly this situation: organisations are ultimately responsible for the AI they deliver, irrespective of the number of contributors. A supply chain of base-model authors, labelling vendors and cloud APIs does not dilute the delivering insurer's responsibility for the outcomes its product creates. Option A misapplies transparency, which is about making the system understandable to its intended audience - technical detail for engineers, plain language for users - not about publishing subcontractor lists. Option B is a wordplay trap: human-centric describes the system's impact on people it affects, not the staffing of vendors. Option D states the exact opposite of the accountability principle - the multi-contributor excuse is precisely what the attribute exists to shut down.

Answer

Transparent - because the company failed to publish its list of subcontractors

Answer

Human-centric - because the subcontractors are staffed by human employees

Answer

None - trustworthy-AI frameworks permit responsibility to be apportioned across the supply chain when multiple vendors contribute components

Question 10

A Dublin energy utility quietly replaces its hardship-support helpline with an AI chatbot. Customers are never told they are dealing with a machine, and applicants whose support requests are rejected are given no way to question the outcome. Which of the OECD AI Principles is most directly engaged by these two failures?

Accepted Answer

Transparency and explainability - people must know when they are engaging with AI and must be able to challenge its outcomes - The OECD's transparency and explainability principle calls for transparency and responsible disclosure so that people know when they are engaging with AI and can challenge outcomes. The utility's two failures - concealing that customers are talking to a machine, and providing no route to contest rejections - map one-for-one onto the two limbs of that principle. Option A fails because inclusive growth concerns trustworthy AI advancing broad development objectives and responsible stewardship; the essential-service setting raises the stakes but does not change which principle the specific conduct breaches. Option C is speculative - robustness, security and safety addresses systems functioning safely with risks continually assessed, and no security or reliability failure is described here. Option D is the near-miss: accountability makes those who develop, deploy or operate AI answerable for its proper functioning, but the conduct described is a disclosure-and-contestability failure, which is transparency and explainability's home ground - and using an external vendor does not itself engage any principle.

Answer

Inclusive growth, sustainable development and well-being - because energy is an essential service

Answer

Robustness, security and safety - because a public-facing chatbot could be compromised by attackers

Answer

Accountability - because the chatbot was built by an external vendor rather than in-house

Question 11

A Helsinki edtech start-up planning an AI tutoring app for teenagers has scheduled exactly one ethics activity: a sign-off against the company's generic corporate ethics policy, timed for one week before launch, after development is complete. The newly hired head of governance objects that this misunderstands ethics by design. What does the approach actually require?

Accepted Answer

Identifying and resolving ethical issues from the start of the project, customising ethical principles to the specific risks of this technology and its audience, and building them into design and development rather than bolting them on at the end - Ethics by design - the sibling of privacy by design - rests on the idea that values can be integrated into the design of technology. Ethical issues are identified and resolved at the start of a project, never as an afterthought; principles are customised to the specific risks of the technology (here, an AI tutor influencing teenagers), built into design and development, and then reassessed as risks evolve. Six core principles guide the exercise: respect for human agency, privacy and data governance, fairness, individual, social and environmental well-being, transparency, and accountability and oversight. Option A describes the exact anti-pattern the approach forbids - a generic policy stamped on a finished product is ethics as afterthought. Option B invents a delegation requirement; ethics by design is an internal design discipline, not an outsourced certification. Option C confuses ethics by design with data protection triggers - it addresses the technology's ethical risks broadly, whether or not special-category data is involved.

Answer

Nothing more - a board-approved ethics policy applied at launch fully satisfies ethics by design

Answer

Handing the ethics review to an external certification body, since internal teams can never assess their own products

Answer

Applying ethics by design only if the app processes special-category personal data

Question 12

A Toronto fitness-app company collected years of workout and heart-rate data under a notice stating the data would be used 'to personalise your training plan'. It now feeds that archive into a new model that scores life-insurance leads for a commercial partner - without asking users and without identifying any legal exception. Which two Fair Information Practices are most directly breached?

Accepted Answer

Purpose specification and use limitation - Purpose specification requires disclosing specific purposes up front and then using the data only for those or compatible purposes; use limitation requires using data only as specified unless the data subject consents or a legal exception applies. Repurposing training-plan data to price insurance leads for a third party - with neither consent nor exception - breaches both limbs squarely, and it is a classic AI-era failure mode: an existing archive quietly becomes model fuel for an incompatible purpose. Option B fails because nothing in the scenario describes weak protection (security safeguards) or inaccurate, stale data (data quality). Option C fails because the original collection was proportionate to the stated purpose - the sin is downstream reuse, not over-collection - and openness concerns clarity about practices generally rather than this specific repurposing. Option D names principles that are only indirectly implicated: individual participation covers access, correction and challenge rights, and accountability makes companies answerable for all the FIPs, but the specific conduct here is textbook purpose-and-use violation.

Answer

Security safeguards and data quality

Answer

Collection limitation and openness

Answer

Individual participation and accountability

Question 13

A global manufacturer wants central consistency in AI policy but knows its business units in different regions face very different AI use cases and regulations. It decides that a central AI office will set policy and standards while regional units run their own day-to-day AI risk decisions within that framework. Which governance structure model has it adopted?

Accepted Answer

Hybrid governance, combining central policy-setting with devolved local decision-making - Hybrid governance combines a central body that sets policy, standards and consistency with decentralised business units that make operational decisions locally - exactly the split described. Option A fails because pure centralised governance concentrates both policy and operational decision-making in one function, whereas here the regions retain real decision authority. Option B fails for the mirror-image reason: pure decentralised governance lacks the central policy spine this company has deliberately built. Option D is a fabricated term - it is not one of the three classic governance structure models (centralised, decentralised, hybrid).

Answer

Centralised governance, because a central office exists

Answer

Decentralised governance, because regions make operational decisions

Answer

Federated certification, an ISO-recognised structure for multinationals

Question 14

A US company that builds and sells a hiring-recommendation model is mapping its obligations under both the Colorado AI Act and the EU AI Act. Its counsel warns of a terminology trap between the two statutes. What is it?

Accepted Answer

Colorado calls this actor a 'developer' while the EU AI Act calls a similar actor a 'provider' - The terminology trap is that the Colorado AI Act says 'developer' where the EU AI Act says 'provider' for essentially the same actor - the organisation that builds the system and places it on the market. Recognising which label belongs to which statute is a recurring exam point. Option B simply reverses the two terms, which is exactly the error the trap invites. Option C fails because neither law collapses all actors into a single 'operator' role; both distinguish the builder from the user-side actor. Option D is doubly wrong: 'business operator' is South Korea's AI Basic Act terminology, and 'deployer' in Colorado refers to the entity using the system, not building it.

Answer

Colorado calls this actor a 'provider' while the EU AI Act uses 'developer'

Answer

Both laws use 'operator' for any organisation touching the system

Answer

The EU AI Act uses 'business operator' while Colorado uses 'deployer' for the builder

Question 15

A Lyon health-tech company is comparing two candidate systems. System One mis-books physiotherapy slots several times a week, causing minor rescheduling annoyance. System Two errs perhaps once a year - but a single error could suppress an urgent oncology referral. The CFO argues System Two is 'obviously the lower-risk choice, because its errors are so rare'. Applying the standard risk formula, how should the governance lead respond?

Accepted Answer

Disagree - risk is the probability of a harm multiplied by its severity, so a rare but potentially catastrophic harm can score higher than a frequent trivial one; System Two cannot be ranked lower on frequency alone - The working risk formula multiplies the probability of a harm by its potential severity - the two factors are co-equal, and neither dominates. A suppressed oncology referral is a false negative with potentially fatal consequences, and the risk process explicitly requires enumerating potential harms including false positives and false negatives before applying the formula; multiplying a tiny probability by a catastrophic severity can easily outscore frequent, trivial booking errors. The scoring then drives action: a high result means avoid or change the AI, a medium result means explore and mitigate. Option A inverts the formula's logic by making probability decisive on its own. Option C fabricates a threshold - severity is always in the multiplication, at every probability level. Option D confuses two different instruments: the probability-times-severity calculus is an internal assessment tool applied when evaluating any AI system, whereas the EU risk pyramid is a separate regulatory tiering; the formula is not gated on a legal classification.

Answer

Agree - probability dominates the calculation, so rarer errors always mean lower overall risk

Answer

Agree, provided System Two's error probability stays below fifty per cent, since severity only enters the formula above that threshold

Answer

Explain that the formula is inapplicable, because it only covers systems already classified as high-risk under the EU AI Act

Question 16

The governance board of a Madrid smart-home vendor is populating the privacy-risk section of its AI risk register and asks which incident type belongs there. Which of the following is best classified as a privacy risk?

Accepted Answer

The voice assistant retains household recordings long after the purpose they were collected for has ended, and routinely captures the conversations of visiting guests who never agreed to any processing - The four AI risk categories are operational, legal, security and privacy. Option A combines the two signature privacy-risk phenomena: data persistence (data outliving the purpose for which it was collected) and data spillover (collecting data about unintended individuals - here, the house guests). These are countered with data minimisation, transparency and compliance with regimes like the GDPR, which is exactly what a privacy-risk register tracks. Option B is a security risk: manipulated inputs that alter outputs are an adversarial attack, sitting alongside data poisoning, hallucinations and deepfakes on the security card, defended by regular audits, human oversight and continuous updates. Option C is a legal risk - intellectual property disputes and liability for harm are core entries in that category, managed through legal reviews and collaboration with legal experts. Option D is an operational risk: high compute costs from specialised processors, large training sets and fast networking are the operational card's classic contents.

Answer

Researchers demonstrate that subtly manipulated audio commands can alter the assistant's outputs in ways its developers never intended

Answer

A rights-holder sues the company after discovering copyrighted audiobooks in the model's training corpus

Answer

Training runs demand specialised GPUs and network upgrades that blow through the annual infrastructure budget

Question 17

The chief product officer of a São Paulo hiring platform reads about ISO/IEC 42005:2025 and instructs the AI governance manager: 'Get us certified under it so we can market the candidate-screening tool as officially safe.' What should the manager explain?

Accepted Answer

ISO/IEC 42005 is a guidance standard for conducting AI system impact assessments - helping organisations identify, evaluate and document potential impacts across the AI life cycle - not a product-safety certification to market against - ISO/IEC 42005:2025 is an international standard giving structured guidance for conducting AI system impact assessments: it supports transparency, accountability and trust by helping organisations identify, evaluate and document potential impacts across the AI life cycle. It is guidance for running a disciplined assessment - and for a hiring platform, that assessment matters: AI impact assessments identify risks early in design, protect fundamental rights particularly for vulnerable or underrepresented groups, may be mandated for high-risk applications, and can guide the go or no-go decision. Option A overstates what any ISO standard can do - following guidance does not certify statutory compliance with the EU AI Act. Option C confuses the voluntary standard with the EU Act's conformity assessment, a distinct legal mechanism for high-risk systems. Option D gets the relationship backwards: the standard structures and strengthens the organisation's own assessment work; it does not perform or replace it.

Answer

ISO/IEC 42005 certifies conformity with the EU AI Act, so certification would also cover European sales

Answer

ISO/IEC 42005 is the mandatory conformity assessment that every high-risk AI system must pass before entering any market

Answer

ISO/IEC 42005 automates the impact-assessment process, removing the need for the platform to conduct its own

Question 18

A regional public housing authority in the EU is about to start using a third-party high-risk AI system to help prioritise housing applications. The authority builds nothing itself. Which pair of obligations applies to it as a deployer?

Accepted Answer

Conducting a fundamental rights impact assessment and retaining system logs for at least six months - Deployers' obligations centre on transparency and monitoring: they must retain automatically generated logs for a minimum of six months, and public bodies deploying high-risk AI - like this housing authority - must carry out a fundamental rights impact assessment (FRIA) before use. Option A lists provider duties: technical documentation and CE marking belong to the actor who builds and places the system on the market. Option C likewise assigns provider tasks - the provider runs the conformity assessment and draws up the EU declaration of conformity; the deployer's role as a public authority is limited to verifying the system is registered in the EU database before use. Option D invents an obligation: no per-decision notification to the Commission exists anywhere in the Act.

Answer

Drawing up the technical documentation and affixing the CE marking

Answer

Running the conformity assessment and drawing up the EU declaration of conformity

Answer

Notifying the European Commission before each individual housing decision

Question 19

An Amsterdam HR-tech vendor pitches a 'workforce engagement' module to a logistics company: warehouse webcams feed an AI that infers each employee's emotional state during shifts so that supervisors can 'intervene early'. The logistics company's counsel is asked whether the EU AI Act permits deployment. What is the best advice?

Accepted Answer

No - emotion recognition in the workplace is explicitly on the EU's prohibited list, so the deployment cannot lawfully proceed regardless of safeguards - Emotion recognition in sensitive settings - explicitly, workplaces and schools - is one of the recurring prohibited-AI categories, and the EU AI Act lists it expressly. Prohibited practices are treated as inherently harmful: they sit above the high-risk tier, and no amount of procedural compliance converts them into permissible deployments. Option A offers the high-risk pathway - conformity assessment plus CE marking - but that route only exists for high-risk systems; a prohibited practice cannot be assessed into legality. Option B similarly misfiles the system into the limited-risk tier, whose transparency-only obligations apply to systems like ordinary chatbots, not banned practices. Option D fails because consent cannot cure a prohibition - and the employment context makes the suggestion doubly hollow, since the power imbalance between employer and employee is part of why the workplace was singled out as a sensitive setting in the first place. The other recurring prohibited categories worth knowing: social scoring, manipulation exploiting vulnerabilities, certain biometric applications, and predictive policing and mass surveillance without safeguards.

Answer

Yes - it is a high-risk workplace system, deployable once a conformity assessment is completed and CE marking affixed

Answer

Yes - it is a limited-risk system, so a transparency notice in the staff handbook suffices

Answer

Yes - provided each employee signs a consent clause in their employment contract

Question 20

A forty-person Paris lab releases a mid-sized general-purpose language model that sits well below any computing threshold associated with systemic risk. The founder assures investors that the EU AI Act's GPAI chapter 'only bites the frontier giants, so we have nothing to do'. Is that correct?

Accepted Answer

No - baseline duties apply to every GPAI provider, including maintaining technical documentation, publishing training-data summaries while respecting copyright, and giving downstream providers transparency through model cards, usage conditions and limitations - Chapter V of the EU AI Act creates two tiers of duty for general-purpose AI. All GPAI providers carry baseline obligations: maintaining technical documentation, publishing training-data summaries while respecting intellectual property and copyright, providing transparency to downstream providers via model cards, usage conditions and limitations, and appointing an EU representative if established outside the EU. Systemic-risk GPAI - very large models above computing thresholds - carries those baseline duties plus extras: risk assessment and mitigation, serious-incident reporting, red-teaming and adversarial testing, robust cybersecurity and physical safeguards. Option B erases the baseline tier entirely, which is exactly the founder's error. Option C reverses the Act's architecture - Chapter V regulates the model's provider directly, because modern laws increasingly regulate GPAI as models, not just as downstream systems. Option D assigns the lab the wrong tier's duties: red-teaming and serious-incident reporting are systemic-risk extras, not the baseline that actually applies to this modest model.

Answer

Yes - GPAI obligations only attach once a model is designated as carrying systemic risk

Answer

Yes - under the Act, only downstream deployers of GPAI carry obligations, never the model's provider

Answer

No - but the lab's only actual obligation is to red-team the model and report serious incidents to the AI Office

Question 21

A Seoul-based AI vendor is preparing an enforcement-exposure memo before launching its product in the EU, the United States and China. Which statement should the memo make about who would enforce against it in each market?

Accepted Answer

Overall EU supervision sits with the EU AI Office; US state AI laws such as Colorado's are enforced by state attorneys general; and China's CAC enforces through filings, fines, suspensions and takedowns - The enforcement map pairs each regime with its authority: overall EU supervision sits with the EU AI Office; South Korea's Ministry of Science and ICT supervises domestically; China's Cyberspace Administration (CAC) enforces its filing regime with fines, suspensions and takedowns; US state laws like Colorado's and California's are enforced by state attorneys general; and Japan operates soft law enforced largely through industry practice and reputation. Sectoral regulators (financial, health) enforce within their domains alongside the central bodies. Option A promotes the EU AI Board beyond its station - it is an advisory body providing technical guidance, not the supervising or fining authority. Option B invents a federal AI regulator; the US state wave is precisely characterised by attorney-general enforcement at state level. Option C attributes Japan's soft-law model to China, whose CAC regime - mandatory filings before public release, with corrective powers - is emphatically hard law.

Answer

In the EU, fines are issued directly by the EU AI Board, which holds overall supervisory authority for AI systems

Answer

In the United States, a single federal AI regulator enforces the state AI acts on the states' behalf

Answer

In China, AI enforcement is soft law, resting on industry self-regulation and reputational pressure

Question 22

A Tallinn online lender approves consumer loans through a fully automated system, relying on GDPR Article 22's contract-necessity exception. To handle challenges, it routes every complaint to a junior clerk who has no access to documentation of the algorithm's logic and no authority to alter outcomes; the clerk sends each applicant a templated confirmation of the original decision. What is the correct compliance analysis?

Accepted Answer

Deficient - the route to contest an automated decision must lead to a reviewer who is competent with AI, working from documented and understandable algorithmic logic, and genuinely able to review the outcome; a template-issuing clerk honours none of that - Falling within an Article 22 exception is the start of compliance, not the end of it. Individuals subject to automated decision-making need a genuine route to complain and request reviews of automated decisions - and for that right to be honoured, reviewers must be competent with AI, and the algorithm's logic must be documented and understandable. A clerk with no documentation, no comprehension of the logic and no power to change outcomes is a rubber stamp, not a review. Option A treats the exception as extinguishing safeguards, when the safeguards are precisely what make automated decision-making tolerable under the exception. Option C misreads the structure of Article 22: contract necessity, legal authorisation and explicit consent are alternative gateways - a lender validly relying on contract necessity does not also need consent stacked on top. Option D relocates the obligation: the review duty runs against the controller operating the system; data protection authorities handle complaints and enforcement, but they do not conduct the lender's decision reviews for it.

Answer

Compliant - once the contract-necessity exception applies, no further safeguards are owed to applicants

Answer

Deficient only because explicit consent must be collected in addition to contract necessity before any automated approval

Answer

Deficient only because reviews of automated decisions must be carried out by the national data protection authority rather than by the lender

Question 23

A model performs superbly on its training data but its accuracy collapses when the team evaluates it before release. Separately, the CRO asks how human oversight of the build phase should be organised. Which pairing of diagnosis and oversight structure is correct?

Accepted Answer

The model has failed to generalise beyond its training set, and oversight should follow the three-lines-of-defence model (do, watch, check) with effective challenge - Proving a model generalises means testing it on new data beyond the training set - stellar training performance with poor evaluation performance is the classic failure to generalise (overfitting). Oversight during build-and-train follows the three-lines-of-defence (3LOD) model: the first line does the work, the second watches, the third checks - paired with the effective challenge principle so reviewers genuinely probe decisions. Option B misdiagnoses (underfitting means poor performance even on training data) and collapses 3LOD into a single audit line. Option C gets the diagnosis right but invents an embedded-regulator requirement no framework imposes. Option D misdiagnoses without evidence of an attack and confuses the NIST Core functions with the 3LOD structure - they are different frameworks.

Answer

The model is underfitting, and oversight should be delegated entirely to the internal audit team

Answer

The model has failed to generalise, and oversight requires an external regulator embedded in the development team

Answer

The model is data-poisoned, and oversight should apply the four NIST Core functions as lines of defence

Question 24

During a pre-acquisition audit of an AI vendor, the buyer discovers the vendor cannot explain where its training data originally came from, how it was transformed, or which datasets fed which model versions. Which governance concepts does this gap most directly implicate?

Accepted Answer

Data lineage and data provenance - Data lineage (tracking data's journey and transformations through the pipeline) and data provenance (documenting data's origin) are exactly what the vendor cannot demonstrate - and without them, data quality, legal rights to the data and downstream accountability all become unverifiable: garbage in, garbage out. Option A concerns how input variables are crafted and how training is tuned - engineering activities, not origin-tracking. Option C describes techniques for shrinking models for deployment, irrelevant to documentation gaps. Option D lists privacy-enhancing techniques; useful tools, but they address how data is protected during processing, not the missing record of where the data came from and how it moved.

Answer

Feature engineering and hyperparameter tuning

Answer

Model compression and quantisation

Answer

Differential privacy and federated learning

Question 25

A project team preparing a new AI-driven triage system must choose the right assessment for three distinct needs: (1) evaluating the system's overall benefits, risks and limitations across its life cycle; (2) addressing risks to personal data; (3) confirming PII is handled in compliance with requirements. Which mapping is correct?

Accepted Answer

(1) AIA, (2) DPIA, (3) PIA - The division of labour is: the algorithmic impact assessment (AIA) - the AI-specific impact assessment - evaluates the system's benefits, risks and limitations across the life cycle, including data issues; the data protection impact assessment (DPIA) identifies and minimises risks to personal data; and the privacy impact assessment (PIA) confirms that PII handling complies with applicable privacy requirements. Option A swaps the AIA and DPIA and misuses the conformity assessment, which demonstrates regulatory compliance for high-risk systems rather than confirming PII handling. Option B assigns the broad life-cycle role to the PIA, which is narrower than that. Option D leads with the conformity assessment, which is a compliance-demonstration exercise, not the general benefits-and-risks evaluation described in need (1).

Answer

(1) DPIA, (2) AIA, (3) conformity assessment

Answer

(1) PIA, (2) conformity assessment, (3) AIA

Answer

(1) conformity assessment, (2) PIA, (3) DPIA

Question 26

A hospital group adopts two third-party AI products in the same quarter: an AI engine embedded into its core patient-scheduling operations, and an off-the-shelf writing assistant that staff may use for drafting routine correspondence. The procurement lead proposes running the identical light-touch review for both. What is the correct governance response?

Accepted Answer

Disagree - the AI integrated into business operations requires the more comprehensive assessment, and reduced visibility into a vendor's system never reduces the organisation's responsibility - Third-party AI splits into two contexts with different treatment: systems integrated into business operations demand the more comprehensive assessment, while off-the-shelf employee tools warrant a proportionate but lighter review - and the governing maxim is that less visibility never means less responsibility. Risks such as data lineage and traceability issues ride along with vendor AI and must be managed through structured screening. Option A fails because responsibility cannot be outsourced with the software; the deploying organisation remains accountable for outcomes. Option B inverts the logic - depth of assessment tracks operational criticality and potential harm, not headcount of users. Option D fails because vendor certifications inform, but never replace, the organisation's own assessment of how the system behaves in its specific context.

Answer

Agree - third-party tools carry the vendor's responsibility, so a light review suffices for both

Answer

Disagree - the off-the-shelf staff tool needs the deeper assessment because more employees will touch it

Answer

Agree - provided both vendors hold ISO certifications, no internal assessment is needed

Question 27

A Stockholm travel platform replaces its single FAQ chatbot with a network of cooperating AI agents that autonomously rebook flights, issue refunds and email customers. The risk team proposes carrying over the chatbot's control regime unchanged: one static risk assessment, refreshed annually. Why is that inadequate?

Accepted Answer

Agentic systems require dynamic decision-making risk modelling - real-time monitoring, audit trails, explainability and human-in-the-loop override - and must account for emergent behaviours across interacting agents through behavioural simulations and multi-agent frameworks such as MAESTRO - Agentic systems are active participants in digital environments - they engage, interact and influence rather than sit passively - and that demands distinct infrastructure (supporting autonomy, long-term memory and multi-step actions), distinct risk models and dynamic, multi-layered, proactive governance. Risk modelling must be dynamic: real-time monitoring, audit trails, explainability and human-in-the-loop or override mechanisms, with emergent behaviours addressed through behavioural simulations, scenario-based modelling and multi-agent frameworks like MAESTRO. A static annual assessment built for a passive chatbot cannot see any of this - and the stakes are real: an airline was held liable for its chatbot's misleading policy advice, the legal cost of insufficient guardrails. Option A misuses the 'not new' point - simple task-specific agents like antivirus software are indeed old, but networks of autonomous cooperating agents are the newer, riskier form. Option C misses the core problem: individually tested agents can still produce emergent behaviours when they interact, which is precisely why multi-agent simulation exists. Option D borrows an unrelated obligation from the EU's systemic-risk GPAI regime; it is no answer to agentic operational risk.

Answer

It is not inadequate - agent technology is not new, so legacy controls carry over unchanged

Answer

The annual assessment can be dropped altogether, because each individual agent already passed pre-deployment testing

Answer

The only necessary addition is disclosing the system's energy consumption to the regulator

Question 28

A Manchester logistics firm's post-deployment review of its delivery-routing model surfaces the two classic risks at once: the sales team wants to reuse the model to promise precise arrival windows to customers - a purpose it was never modelled for - and a brand-new telematics data feed is about to start flowing into it. Which pairing of mitigations correctly matches each risk?

Accepted Answer

Document the original purpose and any new purposes in a model card, and keep snapshots of the algorithm and its outputs so the system can be rolled back if the new data degrades performance - The two predictable post-deployment risks come with matched mitigations. A new purpose the AI was not modelled for is mitigated by documentation - model cards are standardised records of a system's key features, data used, versions, intended use and performance metrics, and they document the original purpose and any new purposes, exposing when a proposed reuse falls outside what the model was built for. New data entering the algorithm is mitigated by keeping snapshots of the algorithm and its outputs, so the firm can roll back if the telematics feed shifts behaviour. Option B reverses the pairing: a snapshot cannot police purpose creep, and a model card cannot restore a model that new data has degraded. Option C is an operational spend that touches neither risk - capacity does not validate a new use or provide a rollback path. Option D is the complacency trap: these two risks arise precisely because conditions change after launch, which is what pre-deployment testing by definition could not cover - and why monitoring continues for deviations in accuracy through to the human-operated shutdown procedure.

Answer

Keep snapshots to control the new purpose, and rely on the model card to manage the incoming data feed

Answer

Purchase additional server capacity, which addresses both the new purpose and the new data simultaneously

Answer

Take no action - pre-deployment testing already validated the model for all future uses and inputs

Foundations of AI

AI Impacts and Responsible AI

AI Governance and Risk Management

AI Regulation and Existing Laws

Governing AI Development and Deployment