AIGP domain - 27% of the exam

Understanding how to govern AI development

Understanding how to govern AI development is 27% of the AIGP exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

92
Practice questions
6
Objectives
31
Medium
61
Hard

Where people go wrong

  • Swapping data provenance and data lineage

    The wrong answers swap the two terms, or reach for a retention schedule, consent management or a model registry. Provenance records where a dataset originated and the legal and consent conditions of its acquisition. Lineage records the transformations data passes through and traces each source forward into the features and model versions derived from it. Lineage is therefore what scopes remediation when a dataset is found to be unlawful or when an erasure request reaches training data.

    See it in a sample question
  • Ranking warnings and insurance above elimination

    The wrong answers lead with a disclaimer, insurance, an output filter or acceptance of the risk, because each is quick and sounds responsible. The mitigation hierarchy orders approaches by how decisively they remove the source of harm. Eliminating the hazard by design comes first, such as dropping a proxy feature or keeping personal data out of the model's accessible context. Reduction controls follow. Warnings, risk transfer and acceptance sit lower because they leave the hazard in place.

    See it in a sample question
  • Confusing data drift with concept drift

    The wrong answers name concept drift when only the applicant population has moved, or answer a changed relationship by restoring the input distribution or adjusting a threshold. Data drift is a change in input distribution while the input-to-target relationship holds. Concept drift is a change in that relationship itself. Input monitoring shows that inputs moved but not whether outcomes still match predictions. Ground-truth labels show that, and confirmed concept drift calls for retraining on recent labelled data.

    See it in a sample question
  • Giving deployers and model providers the wrong EU AI Act duties

    The wrong answers move an obligation onto a party that does not owe it: a deployer re-running conformity assessment, a general-purpose model provider publishing a fundamental rights impact assessment, or a GDPR transfer impact assessment or resident consent standing in for an AI Act step. Conformity assessment and CE marking belong to the provider of a high-risk system. A public-authority deployer carries out a fundamental rights impact assessment and registers its use in the EU database. A general-purpose model provider keeps technical documentation and publishes a training content summary.

    See it in a sample question

Objectives in this domain

What to study

The official documents this domain's practice questions cite most.

Sample questions from this domain

Free sampleUnderstanding how to govern AI developmentmedium

Months after deploying a recommendation model, an organisation learns that one of the third-party datasets used in training was collected without a valid lawful basis and must be removed. Which data governance capability most directly enables the team to identify every downstream model and feature that was derived from that dataset?

  • AA data retention schedule that specifies how long each category of training data is kept before deletion
  • BA consent management platform that records the permissions each data subject has granted for processing
  • CData lineage records that trace each dataset forward through transformations into the features and model versions derived from it Correct
  • DA model registry that stores the performance metrics and approval status of each deployed model version
Understand that data lineage enables forward tracing from a defective source dataset to every derived feature and model version. Data lineage documents how data moves and is transformed across a pipeline, linking sources to the features, training runs and model versions produced from them. When a source dataset is later found to be unlawfully obtained or otherwise defective, lineage allows the team to trace that dataset forward and enumerate every downstream artefact that must be retrained, retired, or re-examined. Without lineage the organisation cannot reliably scope the impact, which is why it is a core capability for governing AI development.

Why A is wrong: A retention schedule is tempting because it governs deletion timelines, but it states when data should be removed rather than tracing which models and features were built from a specific dataset, so it cannot identify the affected downstream artefacts.

Why B is wrong: Consent management is relevant to lawful basis and is tempting given the problem arose from collection, but it tracks individual permissions rather than the flow of data into derived models, so it does not reveal what was trained on the dataset.

Why C is correct: Lineage records the path data takes through the pipeline, so following the affected dataset forward reveals every feature, training run and model version derived from it, which is exactly what the team needs to scope the remediation.

Why D is wrong: A model registry is tempting because it inventories deployed models, but it catalogues models and their metrics without linking them to the specific upstream datasets used, so by itself it cannot trace which models depend on the tainted source.

Free sampleUnderstanding how to govern AI developmenthard

Design-stage risk work for AI systems commonly applies a risk mitigation hierarchy. As a concept, what does this hierarchy actually prescribe about the order in which mitigation approaches should be considered?

  • AConsider eliminating or designing out the hazard first, then reducing the residual risk with controls, and only then transferring or accepting what genuinely remains. Correct
  • BConsider transferring the risk through insurance or contractual indemnity first, because financial cover is the fastest mitigation to put in place before any redesign.
  • CConsider warning affected users with disclaimers first, since informing people of a hazard discharges the organisation's duty before any engineering change is attempted.
  • DConsider accepting the risk first whenever its estimated likelihood is low, then revisit stronger controls only if monitoring later shows the likelihood was underestimated.
The AI risk mitigation hierarchy ranks approaches by decisiveness: eliminate or design out first, then reduce, then transfer or accept residual risk. The hierarchy is ordered by how fully each approach removes the source of harm rather than by cost or speed, so eliminating the hazard outranks engineered reduction, which outranks warnings, transfer and acceptance; the lower tiers leave the hazard present and are reserved for risk that cannot be designed out.

Why A is correct: The hierarchy ranks approaches by how decisively they remove the source of harm, so designing the hazard out ranks above engineered reduction, which in turn ranks above transferring or accepting the leftover residual risk.

Why B is wrong: Buying cover is a real mitigation and is genuinely quick to arrange, which makes it tempting, but transfer sits near the bottom of the hierarchy because it pays for harm rather than preventing it, so it cannot come first.

Why C is wrong: Disclaimers feel like a responsible early step and are cheap to add, but warnings are among the weakest controls because they leave the hazard intact and shift the burden onto users, so they rank low rather than first.

Why D is wrong: Accepting low-likelihood risk and watching it sounds pragmatic and resource-efficient, but acceptance is the last resort in the hierarchy, applied only after stronger options are exhausted, so leading with it inverts the order.

Free sampleUnderstanding how to govern AI developmenthard

A bank operates a deployed loan-approval model. Over three months, the live applicant population shifts towards younger borrowers with thinner credit histories, yet the statistical relationship between the input features and repayment behaviour stays the same. The monitoring team must classify what they are observing so they can choose the right response. How should this phenomenon be classified?

  • AData drift, because the distribution of the model's input features has changed while the relationship between inputs and the target has remained stable. Correct
  • BConcept drift, because the underlying relationship between the input features and the repayment outcome has fundamentally changed.
  • CLabel leakage, because outcome information has contaminated the features the model receives at inference time.
  • DAlgorithmic bias, because the model now produces systematically different decisions for a protected demographic group.
Distinguish data drift from concept drift by whether the input distribution or the input-to-target relationship has changed. Data drift and concept drift are different failure modes in monitoring. Data drift is a shift in the input feature distribution while the learned input-to-output mapping stays valid, whereas concept drift is a change in that mapping itself. Correctly separating the two is what lets a team pick the right corrective action, because a population change calls for refreshed training data while a relationship change calls for model redevelopment.

Why A is correct: Data drift (covariate shift) is a change in the distribution of input features over time without a change in the underlying input-to-target relationship, which is exactly the younger, thinner-file population described, so this classification correctly drives a data-representativeness response such as retraining on current data.

Why B is wrong: Concept drift is tempting because both phenomena degrade live performance, but the scenario explicitly states the input-to-outcome relationship is unchanged, so labelling it concept drift would misdiagnose the cause and point the team at the wrong remedy.

Why C is wrong: Label leakage is a plausible-sounding data-quality fault, but it is a training-design defect where target information bleeds into features, not a post-deployment shift in the live population, so it does not fit an input distribution that moves while relationships hold.

Why D is wrong: Bias is a credible concern given a demographic shift, but the scenario describes a change in who applies rather than evidence of unfair treatment of a group, so naming it bias asserts an unfairness finding the facts do not yet support.

Free sampleUnderstanding how to govern AI developmenthard

A bank is designing a credit-scoring model that meets the EU AI Act definition of a high-risk AI system. The data protection team has already begun the data protection impact assessment required under the GDPR. At the design stage, what additional impact assessment does the AI Act specifically oblige a deployer that is a body governed by public law or a private entity providing public services to carry out before putting this system into use?

  • AA fundamental rights impact assessment that identifies the specific risks of harm to the rights of the people likely to be affected by the high-risk system. Correct
  • BA conformity assessment performed by the deployer to issue the system its own CE marking before any internal use begins.
  • CA transfer impact assessment confirming that any personal data leaving the European Economic Area reaches a jurisdiction with an adequacy decision.
  • DA data protection impact assessment repeated by the provider so that the supervisory authority receives two independent copies of the analysis.
Recognise that the EU AI Act adds a fundamental rights impact assessment for designated deployers of high-risk systems, separate from the GDPR data protection impact assessment. The two assessments protect different interests: the data protection impact assessment under the GDPR focuses on processing of personal data, while the fundamental rights impact assessment under Article 27 of the AI Act examines the broader rights impacts of deploying a high-risk system, so completing one does not discharge the other.

Why A is correct: Article 27 of the EU AI Act requires certain deployers of high-risk systems to perform a fundamental rights impact assessment, which complements rather than replaces the GDPR data protection impact assessment.

Why B is wrong: Conformity assessment and CE marking are obligations placed on the provider before the system is placed on the market, not a design-stage duty the deployer discharges, so this confuses the provider and deployer roles.

Why C is wrong: A transfer impact assessment is a GDPR international-transfer mechanism and is tempting because data flows are involved, but it is not the assessment the AI Act adds for high-risk deployment.

Why D is wrong: Duplicating the data protection impact assessment adds no new safeguard and is not required, so this option mistakes administrative repetition for the distinct fundamental rights analysis the AI Act mandates.

Other domains in this exam

See also the AIGP cert hub, the study guide, and the cheat sheet.

Examworthy is not affiliated with or endorsed by IAPP. Original, blueprint-aligned practice material only.