Free AIGP practice questions

A bank collected customer transaction records under a privacy notice that stated the data would be used to operate accounts and detect fraud. The data science team now wants to reuse those same records to train a marketing propensity model. Under the GDPR principle of purpose limitation, what must the team establish before proceeding on this basis?

• AThat the marketing model will, as a secondary benefit, improve the accuracy of the existing fraud detection model for the same customers
• BThat the records have been pseudonymised so that direct identifiers are replaced with tokens before training begins
• CThat the new marketing purpose is compatible with the original purposes, or otherwise obtain a fresh lawful basis such as consent for the reusecheck_circle Correct
• DThat the resulting model will be evaluated for demographic bias before any marketing campaign is launched

Why A is wrong: This is tempting because linking the new use to the original fraud purpose sounds like a compatibility argument, but a marketing propensity model is a distinct commercial objective, so an incidental fraud benefit does not make the marketing use lawful under the original notice.

Why B is wrong: Pseudonymisation is a useful safeguard and can support a compatibility assessment, but on its own it does not authorise a new incompatible purpose because pseudonymised data remains personal data subject to purpose limitation.

Why C is correct: Purpose limitation permits further processing only where it is compatible with the purposes for which data was collected, and where it is not compatible the controller must secure a separate lawful basis such as fresh consent before reusing the records.

Why D is wrong: Bias evaluation is good governance and may be required for fairness, but it addresses model outputs rather than the lawfulness of reusing the data, so it does not resolve the purpose limitation question about whether the reuse is permitted at all.

An AI governance lead is distinguishing two data documentation practices for a training pipeline. One practice records the original sources and the legal and consent conditions under which each dataset was acquired, while the other records the sequence of transformations the data passed through inside the pipeline. Which option correctly names the first practice and explains why it matters for governing AI development?

• AData minimisation, because limiting the volume of collected data reduces the legal exposure of the training set
• BData lineage, because tracing data backwards through each pipeline stage demonstrates the integrity of the final dataset
• CData anonymisation, because removing identifiers from source records establishes a defensible origin for the training set
• DData provenance, because recording each dataset's origin and the rights under which it was obtained shows the organisation may lawfully train on itcheck_circle Correct

Why A is wrong: Data minimisation is tempting because it is a recognised data governance principle, but it concerns restricting how much data is collected rather than documenting where data came from and on what legal terms, so it does not describe the first practice.

Why B is wrong: Data lineage is a close and tempting answer, but lineage describes the flow and transformation of data through the pipeline, which is the second practice in the scenario rather than the recording of original sources and acquisition conditions.

Why C is wrong: Anonymisation is tempting because it touches on lawful use, but it is a technique for stripping identifying information rather than a documentation practice, and it neither records the source of data nor proves the rights under which it was collected.

Why D is correct: Provenance captures the origin and chain of custody of data, including the legal and consent basis for acquisition, which is exactly what evidences a lawful right to use the data and supports accountability when sources are later challenged.

A team is assembling a training dataset for a model that will screen job applicants across an organisation operating in several regions. Their internal hiring history is accurate and well labelled, but it reflects past recruitment that drew almost entirely from one region. When assessing data quality for this use, which dimension should the governance team treat as the priority gap to address?

• ARepresentativeness, because the data over-represents one region and is unlikely to reflect the applicant population the model will actually screencheck_circle Correct
• BCompleteness, because the team should confirm that no individual applicant records contain missing fields before training
• CTimeliness, because older hiring records may no longer reflect current job requirements and should be refreshed first
• DAccuracy, because the labels in the hiring history must be verified against authoritative records before the data is used

Why A is correct: Representativeness assesses whether the training data reflects the population the system will be applied to, and a dataset skewed to one region will misrepresent applicants elsewhere, producing biased screening, so closing this gap is the governance priority.

Why B is wrong: Completeness is tempting because missing fields are a common data quality issue, but the described problem is that the records over-represent one region, not that individual records have gaps, so completeness is not the priority gap here.

Why C is wrong: Timeliness is a genuine quality dimension and sounds relevant for hiring data, but the scenario identifies a skew toward one region rather than data being out of date, so addressing recency would not close the stated gap.

Why D is wrong: Accuracy is tempting and the question even notes the data is accurate, which is the trap, because verifying already-accurate labels does nothing about the regional skew that threatens performance for applicants from other regions.

A governance lead is categorising the AI systems in use across a bank. One system drafts new marketing copy from a short brief, while another scores loan applications as approve or decline. What distinguishes the copywriting system as generative rather than classic AI?

• AIt produces new content by sampling from patterns it has learned, whereas the classic system assigns inputs to predefined outcome categories.check_circle Correct
• BIt runs on cloud infrastructure, whereas classic systems are always installed and operated entirely on local hardware.
• CIt was trained on labelled examples, whereas classic systems are built only from unlabelled data with no human supervision.
• DIt guarantees factually accurate outputs on every request, whereas the classic system can return scored predictions that are occasionally incorrect.

Why A is correct: Generative AI synthesises novel artefacts such as text by modelling the data distribution, while classic discriminative AI maps an input to one of a fixed set of labels such as approve or decline.

Why B is wrong: Deployment location is tempting because hosting often differs in practice, but where a model runs has no bearing on whether it is generative or classic; both kinds run on cloud or local hardware.

Why C is wrong: This inverts the truth and is tempting because supervision is a real distinction; many classic classifiers use labelled data, and generative models often learn in a self-supervised way, so the contrast is wrong.

Why D is wrong: This is tempting because accuracy matters to governance, but generative models do not guarantee correctness and can hallucinate, so reliability does not define the generative category.

A product team must run a language model directly on a battery-powered handheld device with no reliable network connection. Weighing a small language model against a large one for this setting, which trade-off should most influence the team's choice?

• AA large model removes the need for any prompt design at all, whereas a small model alone demands carefully engineered prompts to function correctly.
• BA small model needs less memory and compute, letting it run on-device with lower latency, whereas a large model offers broader capability.check_circle Correct
• CA small model can be trained without any data, whereas a large model requires a very large labelled training corpus.
• DA large model always produces more harmful outputs, whereas a small model is inherently safe and needs no output review.

Why A is wrong: Prompt design is a tempting axis, but prompting helps models of every size and is not the deciding factor; large models still benefit from good prompts, so this misstates the trade-off.

Why B is correct: Small models fit the tight memory and power budget of a handheld and respond quickly offline, while large models bring wider capability at a much higher resource cost that this device cannot meet.

Why C is wrong: Training data is a real consideration, but no useful model trains on zero data, and both small and large models learn from large corpora, so the contrast as stated is false.

Why D is wrong: Safety is a genuine governance concern, but the claim is an absolute that does not hold; model size does not by itself determine safety, and outputs of any size need review.

A team is documenting the inputs each AI system accepts. They describe one model as multimodal. In governance terms, what does describing a model as multimodal indicate about it?

• AIt has been deployed across several geographic regions at once to keep latency low for users in different locations.
• BIt is governed by more than one regulatory regime because it processes personal data in several jurisdictions.
• CIt can take in or produce more than one type of data, for instance combining text input with images or audio.check_circle Correct
• DIt blends several distinct algorithms into one ensemble so that their predictions can be averaged together.

Why A is wrong: Multi-region deployment sounds related to the word multi, but it concerns hosting topology, not the data types a model handles, so it does not describe modality.

Why B is wrong: Multiple regulatory regimes is a plausible reading of multi, but modality refers to data types the model handles, not the number of laws that apply to it.

Why C is correct: Multimodal means the model works across several data types, for instance accepting an image alongside a text prompt, in contrast to a unimodal model limited to a single type.

Why D is wrong: Ensembling combines multiple models, which is tempting near the idea of multi, but a multimodal model is defined by handling multiple input or output types, not by averaging algorithms.

A product team is starting to build a recommendation model that will process customers' personal data, and the privacy lead wants the design to embody the GDPR principle of data protection by design and by default rather than bolting privacy on at launch. Which design choices genuinely give effect to that principle for this AI system? Select TWO.

• APseudonymise the training records as an integral technical measure built into the data pipeline from the outset, so identifiers are separated before modelling begins.check_circle Correct
• BConfigure the default settings so that only the personal data necessary for the recommendation purpose is collected and processed, with broader fields off unless a user opts in.check_circle Correct
• CCollect every available customer attribute now and run a single privacy review just before launch to decide which fields the model may keep.
• DPresent a broad consent banner covering any present or future use of the data, so the team need not revisit the design as the model evolves.
• ERetain the full raw dataset indefinitely so the model can be retrained later, deferring any minimisation decisions until a concrete need arises.

Why A is correct: Pseudonymisation is named in the GDPR as an example of a technical measure that implements data protection by design when embedded into processing from the start.

Why B is correct: Data protection by default requires that, without user action, only data necessary for each specific purpose is processed, which this minimising default delivers.

Why C is wrong: Tempting because a launch review feels diligent, but by-design means embedding measures throughout development, not a one-off pre-launch check after broad collection.

Why D is wrong: Seems to secure a lawful basis, but blanket future-proofed consent is not specific and does nothing to build protective measures into the system's architecture.

Why E is wrong: Appears prudent for future retraining, but indefinite raw retention contradicts both default minimisation and storage limitation that the principle is meant to enforce.

A retailer wants to train a churn-prediction model on existing customers' purchase histories and intends to rely on legitimate interests as its GDPR lawful basis, since the analytics use was not part of any contract and consent was never sought. To make that basis defensible, what must the retailer establish before the training processing begins? Select TWO.

• AThat a specific, real legitimate interest is being pursued and that processing this personal data is necessary to achieve it, with no less intrusive means reasonably available.check_circle Correct
• BThat a balancing assessment weighs the retailer's interest against the rights and reasonable expectations of the customers, and that those interests do not override the processing.check_circle Correct
• CThat each affected customer has given fresh explicit opt-in consent to the churn analytics use before any of their records are allowed to enter the training set.
• DThat the supervisory authority has been formally notified of and has pre-approved the churn-modelling activity before the training run starts.
• EThat the purchase histories are already in the public domain, which on its own removes the need to assess the affected customers' interests at all.

Why A is correct: The legitimate-interests test requires identifying a genuine interest and showing the processing is necessary for it, which is the first limb the retailer must satisfy.

Why B is correct: Legitimate interests demands a balancing test against data subjects' rights and reasonable expectations; without it the basis cannot lawfully be relied upon.

Why C is wrong: Tempting because consent is also a lawful basis, but requiring it here confuses the two bases; legitimate interests applies precisely when consent has not been obtained.

Why D is wrong: Sounds like due diligence, but the GDPR sets no general duty to pre-notify or seek authority approval for ordinary legitimate-interests processing of this kind.

Why E is wrong: Plausible-sounding shortcut, but data being public does not dispense with the balancing test, and these purchase histories are not public in any event.

A controller has completed a data protection impact assessment for an AI system that profiles applicants, and the assessment shows a residual high risk to individuals that the planned safeguards do not bring down to an acceptable level. Acting as controller under the GDPR, what must the organisation do before this processing begins? Select TWO.

• AConsult the supervisory authority for prior advice about the intended processing before it starts.check_circle Correct
• BProvide the supervisory authority with the completed DPIA as part of that prior consultation.check_circle Correct
• CObtain a binding sign-off from the data protection officer that legally authorises the processing to start.
• DNotify each affected data subject individually of the residual risk before any of their data is processed.
• EAbandon the processing permanently, since a residual high risk after a DPIA legally bars the system from ever going live.

Why A is correct: Where a DPIA shows a residual high risk that safeguards do not reduce, the controller must seek prior consultation with the supervisory authority before processing.

Why B is correct: When prior consultation is triggered, the GDPR requires the controller to give the supervisory authority the DPIA itself, so supplying the assessment is a distinct duty tied to the residual high risk.

Why C is wrong: Tempting because a DPO advises on the DPIA, but the DPO gives advice and cannot grant legal authorisation; that role is not how the GDPR resolves residual high risk.

Why D is wrong: Tempting as it sounds protective, but the GDPR routes unresolved high risk to the regulator, not to mandatory individual pre-processing notices for every subject.

Why E is wrong: Tempting because the risk sounds disqualifying, but a residual high risk triggers prior consultation rather than an automatic permanent ban, and the authority may still advise that the processing can proceed.

A facilities team builds a heating controller whose behaviour is fixed entirely by rules a programmer wrote, such as switching the boiler on whenever a sensor reads below 18 degrees. A governance lead is asked why this controller does not meet the widely used OECD and EU AI Act definition of an AI system. Which characteristic, present in that definition, does the controller lack?

• AIt infers from the inputs it receives how to generate outputs such as predictions or decisions, rather than only executing rules a human wrote.check_circle Correct
• BIt runs continuously rather than only when an operator manually triggers each individual action through the interface.
• CIt connects to the internet so that it can transmit its sensor readings to a remote server for storage.
• DIt stores a historical log of past temperature readings that an engineer can later download and inspect.

Why A is correct: The OECD and EU AI Act definition centres on a system that infers from input how to produce outputs like predictions, content, recommendations or decisions, which a fixed rules-only controller does not do.

Why B is wrong: Running continuously is tempting because people associate AI with always-on services, but autonomous scheduling is not what the definition turns on, and many simple automated devices run continuously without being AI.

Why C is wrong: Network connectivity sounds modern and AI-like, yet the definition says nothing about connectivity, and an offline model is still AI while a connected thermostat is not.

Why D is wrong: Keeping a data log feels relevant because AI uses data, but merely recording readings is passive storage and is not the inference of outputs that the definition requires.

A vendor markets a customer-service chatbot as artificial general intelligence because it answers questions on many topics. A governance reviewer notes that the tool is trained and tuned only for support conversations and fails outside that scope. Which type of AI does this tool actually represent?

• AArtificial general intelligence, because the system can hold a conversation across a wide range of everyday subjects with its users.
• BArtificial narrow intelligence, because it is designed and trained to perform within a single, bounded task domain.check_circle Correct
• CArtificial superintelligence, because the model responds far more quickly than a human support agent can type a reply.
• DSymbolic expert-system AI, because its replies are produced by a hand-curated decision tree of explicit support rules.

Why A is wrong: Conversing across topics looks like generality, but breadth of small talk is not the same as human-level competence across arbitrary tasks, which is what general AI denotes.

Why B is correct: Narrow AI is built for a specific, bounded task and does not generalise beyond it, which exactly describes a chatbot trained only for support and failing outside that scope.

Why C is wrong: Superintelligence is tempting given the speed, yet speed alone is not the criterion, and superintelligence refers to capability surpassing the best humans across virtually all domains.

Why D is wrong: Expert systems are a real AI category, so this is plausible, but a trained conversational model learns from data rather than running a hand-built rule tree, so the label does not fit.

A bank trains a loan-approval model on a decade of its own lending decisions. Independent testing shows the model approves a far lower share of applicants from one ethnic group than from others with comparable finances, mirroring historical patterns in the training data. Which AI risk does this most directly illustrate?

• AA cybersecurity risk, in which an external attacker has poisoned the training data to change the model's lending behaviour.
• BAn intellectual-property risk, because the model may have memorised confidential applicant records that it can later reproduce.
• CA bias and discrimination risk, in which the model reproduces unfair disparities present in the historical training data.check_circle Correct
• DA transparency risk, because applicants are not told that an automated model contributed to the decision on their loan.

Why A is wrong: Data poisoning is a genuine threat, which makes it tempting, but the scenario attributes the skew to historical patterns the bank itself produced, not to an external attacker tampering with the data.

Why B is wrong: Memorisation of training records is a real concern, yet the harm described is unequal approval rates across groups, not the leaking or copying of protected content.

Why C is correct: When a model learns from biased historical decisions it reproduces those disparities, here disadvantaging one group despite comparable finances, which is the defining pattern of AI bias and discrimination risk.

Why D is wrong: Failing to disclose automated decision-making is a transparency concern that often co-occurs, but the measured harm here is the unequal outcome itself, which is discrimination rather than non-disclosure.