AIGP domain - 21% of the exam

Understanding the foundations of AI governance

Understanding the foundations of AI governance is 21% of the AIGP exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

72
Practice questions
6
Objectives
12
Easy
48
Medium
12
Hard

Where people go wrong

  • Swapping the named characteristics of AI

    The wrong answers name a neighbouring characteristic instead of the one the scenario shows. A bank that cannot trace which inputs drove a decision has opacity, not speed or autonomy. An agent that re-prices without approval shows autonomy, and one classifier error reaching millions shows speed and scale. Each characteristic carries its own control: explainability and audit for opacity, human oversight for autonomy, staged rollout and rollback for speed and scale.

    See it in a sample question
  • Assigning a control to the wrong responsible AI principle

    The wrong answers credit a control to a neighbouring principle. Encryption serves privacy and security, a model card serves transparency, and scheduled retraining serves safety and reliability, so none of them delivers human-centricity. Human-centricity keeps a clinician able to review and override, or gives a refused applicant a route to contest. Drift in a forecasting model is a safety and reliability failure, and discarding transcripts is a privacy measure.

    See it in a sample question
  • Treating the vendor as carrying the accountability

    The wrong answers hand accountability to the supplier through indemnification, a warranty, published benchmarks, a certification at purchase or a one-time model card. The organisation that puts an AI system into use stays accountable for it. The correct answers keep that accountability with the deployer: its own due diligence, contractual monitoring and audit rights, notice of material changes, and testing of outputs against its own use case, because behaviour can shift after a vendor update.

    See it in a sample question
  • Misplacing duties across three lines of defence and RACI

    The wrong answers give one line or role the duty of another. The first line builds and operates the model and manages its day-to-day risks, so it cannot assure its own work. The second line sets policy and challenges the first line from within management, so it is not independent either. Internal audit, the third line, gives independent assurance. In RACI, the accountable party is the single party answerable for the outcome, and the responsible party does the work.

    See it in a sample question
  • Letting risk tolerance or size lower the governance floor

    The wrong answers let a firm's appetite or scale decide how little governance it needs. A founder's high risk tolerance is a preference, and it cannot reduce controls below the obligations external law imposes on a high-risk system such as credit scoring under the EU AI Act. Tolerance can add rigour above that floor. Objective drivers are the stakes for individuals and binding sector law. A small company gets a lightweight policy with a named owner, not deferred governance.

    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 the foundations of AI governancemedium

A bank deploys a deep neural network to score loan applications. When a customer is declined and asks why, the risk team finds they cannot trace which input features drove the individual decision, even though they hold the full model. Which characteristic of AI does this difficulty most directly illustrate, and why does it demand a dedicated governance response?

  • ASpeed and scale, because the model processes far more applications per hour than human underwriters and that throughput is what blocked the explanation.
  • BAutonomy, because the model issues credit decisions and adjusts its own approval thresholds without any human in the loop.
  • COpacity, because the internal reasoning of the model is not readily interpretable, which obstructs explanation and audit of individual outcomes. Correct
  • DProbabilistic output, because the model returns a likelihood of default rather than a guaranteed repayment outcome for the applicant.
Recognise opacity as the AI characteristic that prevents tracing how a model reached an individual decision, demanding explainability and audit controls. Opacity refers to the limited interpretability of a model's internal workings; because the path from inputs to a specific output is not transparent, holding the model is insufficient to explain a decision, which is why governance frameworks require explainability, record-keeping and the ability to contest outcomes.

Why A is wrong: Throughput is a genuine AI characteristic and the model is indeed fast, which makes this tempting, but the team's problem is the inability to explain a single decision, not the volume of decisions, so speed and scale is the wrong lens here.

Why B is wrong: Autonomy sounds relevant because the model decides outcomes, but the described failure is an inability to interpret the reasoning, not the model acting without oversight, and nothing here shows it changing its own thresholds.

Why C is correct: This is correct: opacity is the property that a model's internal decision logic resists human inspection, so even with full access the team cannot explain one outcome, and that is precisely why governance must mandate explainability, documentation and audit.

Why D is wrong: It is true the model emits a probability, which makes this plausible, but the obstacle described is tracing the drivers of the decision, not the statistical nature of the score, so probabilistic output does not name the core difficulty.

Free sampleUnderstanding the foundations of AI governancemedium

A hospital is procuring an AI triage tool that ranks incoming patients by clinical urgency. The governance team wants the deployment to reflect the responsible AI principle of human-centricity. Which design choice most directly expresses that principle?

  • AEnsuring a qualified clinician reviews each ranking and can override the tool before any patient is acted upon Correct
  • BEncrypting the patient ranking scores both at rest and in transit so that the data cannot be intercepted by an attacker
  • CPublishing a model card that records the training population, evaluation metrics and known limitations of the triage tool
  • DRetraining the model every quarter on the newest patient records so that its accuracy keeps improving over time
Recognise that human-centricity is expressed by keeping a competent human meaningfully in control of consequential AI decisions, including the ability to override. Human-centricity treats people, their rights and their wellbeing as the purpose of the system rather than an afterthought. In a consequential setting such as clinical triage, this principle is realised when a qualified human can review the AI recommendation and override it before action is taken, ensuring the technology augments rather than displaces human judgement.

Why A is correct: Human-centricity places human wellbeing and agency at the centre of the system, so preserving a clinician's ability to review and override the AI output keeps a competent person meaningfully in control of consequential decisions.

Why B is wrong: Encryption is tempting because protecting patient data clearly matters, but it serves the privacy and security principle rather than human-centricity, which concerns keeping people in control of and benefiting from the system.

Why C is wrong: A model card is a strong responsible AI practice, but it primarily advances transparency and explainability by disclosing how the tool works rather than ensuring humans retain meaningful oversight over its decisions.

Why D is wrong: Scheduled retraining supports the safety and reliability principle by maintaining performance, but on its own it does nothing to keep a human in control of the decision, which is the essence of human-centricity.

Free sampleUnderstanding the foundations of AI governancehard

An insurer integrates a third-party vendor's pricing model into its quoting platform and presents the quotes to consumers under its own brand. The vendor trained and tested the model. Which approach best preserves the insurer's accountability across the AI life cycle?

  • ATreating accountability as transferred to the vendor because the vendor designed, trained, and tested the model under its own controls
  • BRelying on the vendor's published accuracy benchmarks and deferring any independent assessment until a consumer complaint is received
  • CRequiring the vendor to indemnify the insurer for any losses, which removes the need for the insurer to monitor the model in production
  • DSecuring the vendor's documentation, defining contractual monitoring rights, and assessing the model against the insurer's own use case and risk criteria Correct
Understand that deploying a third-party AI system does not transfer accountability, which is preserved through due diligence, contractual controls, and context-specific assessment. Accountability for an AI system rests with the organisation that puts it into use, even when a third party built it. Sound third-party risk management therefore combines pre-contract due diligence, contractual flow-down of obligations and monitoring or audit rights, and an independent assessment of the model against the deployer's own use case and risk appetite. Indemnities and vendor benchmarks address liability or marketing claims after the fact and do not discharge the deployer's life cycle oversight duties.

Why A is wrong: This is tempting because the vendor did the technical work, but an organisation that puts an AI system into use under its own brand remains accountable for that use and cannot contract away its obligations to the consumers it serves.

Why B is wrong: Published benchmarks feel reassuring, but they reflect the vendor's data and context rather than the insurer's, and waiting for a complaint replaces proactive oversight with reactive cleanup, leaving the insurer unable to show it managed the risk.

Why C is wrong: Indemnification is attractive because it appears to shift financial exposure, but it allocates liability after harm occurs and does nothing to prevent harm or satisfy the insurer's own duty to oversee an AI system it has placed into use.

Why D is correct: Effective third-party risk management means the deploying organisation performs its own due diligence, flows obligations and audit or monitoring rights into the contract, and evaluates the model in its specific use context, which keeps accountability with the party putting the system into use.

Free sampleUnderstanding the foundations of AI governancemedium

A bank adopts the three lines of defence model for its AI governance programme. The data science team builds and operates a credit-scoring model, and a risk and compliance function reviews the model against policy before approval. Senior leadership now wants assurance that the controls themselves are designed and operating effectively, provided by a party with no role in building or signing off the model. Which function should provide that assurance?

  • AThe data science team, because the engineers who built the credit-scoring model understand its controls in the most depth and can attest to how well they work.
  • BThe risk and compliance function, because it already reviews the model against policy and is therefore best placed to confirm that the controls operate effectively.
  • CInternal audit, because as the third line it is independent of both model development and the control framework and can objectively assure that controls are designed and operating effectively. Correct
  • DThe board of directors, because ultimate accountability for the AI programme rests with the board and it should therefore perform the assurance review itself.
In the three lines of defence model, independent assurance over AI controls is the role of the third line, typically internal audit. The three lines model separates ownership of risk and controls in the first line, oversight and challenge in the second line, and independent assurance in the third line, so only the third line can objectively evaluate controls it neither built nor monitors.

Why A is wrong: It is tempting because the builders know the model intimately, but the first line owns and operates the controls, so it cannot independently assure its own work without a conflict of interest.

Why B is wrong: This is plausible because compliance does challenge the first line, but as the second line it sets and monitors the control framework it would be assuring, so it is not independent of it.

Why C is correct: Correct: the third line, typically internal audit, is independent of both the first line that builds and the second line that oversees, so it can objectively assure the design and operating effectiveness of controls.

Why D is wrong: It is tempting because the board is ultimately accountable, but the board provides oversight and direction rather than performing operational assurance testing, which is delegated to an independent assurance function.

Free sampleUnderstanding the foundations of AI governancemedium

A fast-growing startup is deploying an AI system that scores consumer creditworthiness for lenders in a tightly regulated market. Its founders, who accept more risk than established rivals, want the lightest governance the situation allows. A governance adviser must identify which factors objectively raise the governance intensity the firm needs, independent of the founders' appetite for risk. Which factors raise that objective requirement? Select TWO.

  • AThe system makes consequential decisions about individuals in a regulated, high-impact domain, raising the objective stakes of an error. Correct
  • BA binding sector law imposes specific obligations on how the system is built, tested and overseen, independent of the firm's own appetite. Correct
  • CThe founders' stated willingness to accept more risk than rivals, which lets the firm scale back governance to match its chosen appetite.
  • DThe firm is small and early-stage, so a lighter, leaner governance process is automatically sufficient for whatever the system does.
  • EThe vendor markets the underlying model as advanced, which signals the deployment needs proportionally heavier governance scrutiny.
Governance intensity is set objectively by harm severity and legal obligation, which a firm's risk tolerance and size cannot lower below that floor. Risk tolerance and organisational size shape how governance is implemented, but the objective intensity required is anchored by the stakes of potential harm and by binding legal obligations, which apply irrespective of an organisation's chosen appetite.

Why A is correct: Correct: the severity of potential harm to affected individuals is an objective driver of governance intensity that the firm's own appetite cannot lower.

Why B is correct: Correct: legal obligations set a floor of required governance that applies regardless of how much risk the founders are willing to accept.

Why C is wrong: Tempting because risk tolerance does shape governance, but appetite is a subjective preference that cannot reduce the objective intensity the context demands.

Why D is wrong: Tempting because proportionality is real, yet size never makes light governance automatically sufficient for a high-impact, regulated system.

Why E is wrong: Tempting since capability sounds relevant, but marketing claims about sophistication are not what objectively drives required governance intensity.

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.