Understand what AI is and why it needs governance.
- ✓AI = a machine-based system that infers how to produce outputs that influence environments.
- ✓AI ⊃ machine learning ⊃ deep learning ⊃ generative AI.
- ✓RPA and rules defined solely by people are automation, not AI; expert systems that infer from encoded knowledge can be AI.
- ✓Supervised, unsupervised, semi-supervised, reinforcement and self-supervised learning are the main ways machines learn.
- ✓Agentic AI acts autonomously toward goals, so it raises the governance stakes.
- AI system (OECD / EU AI Act):
- A machine-based system that infers from inputs how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments.
- Machine learning:
- AI that learns patterns from data rather than being explicitly programmed with rules.
- Robotic process automation (RPA):
- Software that mimics repetitive human steps using fixed rules. It automates tasks but does not learn or infer.
- Expert system:
- A knowledge-based system that reasons over facts and if-then rules supplied by human experts to reach new conclusions; it can count as AI because it infers from encoded knowledge.
- Foundation model:
- A large model trained on broad data that can be adapted to many downstream tasks.
- Agentic AI:
- AI that plans and takes multi-step actions, often using tools, to pursue a goal with limited human intervention.
Exam tips: If an option says RPA 'learns', 'relies on human judgment' or 'increases human involvement', it's a trap. RPA follows fixed rules to REDUCE human effort. · When asked what makes something AI, look for inference (learning from data, or reasoning from encoded knowledge), not automation, speed or data volume. · Rules don't automatically mean 'not AI'. An expert system that infers from a knowledge base can be AI (EU AI Act Recital 12); software that only executes rules defined solely by people is not. Real systems are judged case by case. · Labeled data means supervised; no labels means unsupervised; a few labels plus lots of unlabeled data means semi-supervised; rewards mean reinforcement. The exam loves these one-word tells.
- ✓AI harms land on individuals, groups, organizations and society. Classify by who absorbs them.
- ✓The BoK's named risk patterns: misalignment with objectives, ethics and bias risk, complexity and scalability.
- ✓Scale turns one flaw into thousands of harms before anyone notices.
- ✓AI sharpens privacy harms like inference and aggregation; subjective and objective privacy harms both count.
- Individual harm:
- Damage to a specific person's rights, safety, civil liberties or economic opportunity.
- Group harm:
- Damage to a community or population sub-group, such as systematic discrimination or targeted surveillance.
- Organizational harm:
- Damage the organization itself absorbs: disruption, breaches, fines, lost business, reputational or cultural damage, and acceleration risk from rushing AI out untested.
- Societal harm:
- Damage to shared goods like democratic participation, trust in information, access to education or the environment.
- Misalignment:
- When an AI system pursues a measurable proxy or purpose that diverges from what its owners actually intended.
- Inference harm:
- A privacy harm in which an AI derives sensitive traits a person never disclosed.
Exam tips: Classify harm by WHO absorbs it, not by the technology. Surveillance 'feels' societal, but if it targets one community, the IAPP answer is group harm. · When asked which harm is MOST directly raised, pick the first and biggest hit. Organizational harm (fines, reputation) is usually a consequence, not the primary harm. · 'Each decision is individual' is a classic distractor when the facts describe a pattern across a community. · Misalignment means optimizing the proxy, not the goal. Look for clicks, watch time or 'tickets closed' rising while the real outcome gets worse.
- ✓CAPS POD: complexity, autonomy, probabilistic outputs, speed and scale, potential for harm or misuse, opacity, data dependency.
- ✓AI gives likelihoods, not guarantees, so governance sets thresholds and keeps humans involved where stakes are high.
- ✓Data dependency means models drift as the world changes; monitoring never ends.
- ✓Speed and scale multiply ordinary errors into mass harms.
- Opacity (black box):
- When an AI system's internal reasoning can't readily be inspected or explained, even by its builders.
- Probabilistic output:
- A likelihood or best guess that can vary and can be wrong, rather than a guaranteed result.
- Deterministic output:
- A result that is always the same for the same input, as with conventional rule-based software.
- Data drift:
- A change in the input data a deployed model receives compared with its training data.
- Concept drift:
- A change in the relationship between inputs and the correct output, so old patterns stop holding.
- Hallucination:
- Fluent, confident output from a generative model that is false or fabricated.
Exam tips: 'Configure it to be right every time' signals a probabilistic-vs-deterministic question. The IAPP answer adds thresholds, human escalation and monitoring. · Drift is a monitoring problem (decay after deployment); overfitting is a training problem (fails on new data from day one). · On LEAST/NOT questions about AI's unique traits, the trap is a true statement that applies to ALL software, such as patching or uptime. · Reusing existing controls is good; reusing them 'unchanged' or 'as-is' for AI is the trap.
- ✓Fair Ships Pass Through Any Harbor: fairness, safety & reliability, privacy & security, transparency & explainability, accountability, human-centricity.
- ✓Transparency (what, and that AI is in use) and explainability (how a specific result came out) are related but distinct.
- ✓Accountability can't be outsourced to a vendor or a model.
- ✓OECD, NIST and others use different labels for the same core values.
- Fairness:
- Ensuring AI outcomes don't unjustly disadvantage individuals or groups, with harmful bias identified and managed.
- Transparency:
- Disclosing what an AI system is, what it does and that AI is in use; NIST frames it as answering 'what happened'.
- Explainability:
- The ability to describe how an AI system produced a particular output or decision.
- Interpretability:
- In NIST's terms, understanding why a system produced an output and what it means in context for the user.
- Accountability:
- Identifiable people and organizations answer for an AI system's outcomes and have the authority to act.
- Human-centric AI:
- AI designed to augment people, respect their rights and keep meaningful human control over consequential decisions.
Exam tips: Transparency answers WHAT (and that AI is in use); explainability answers HOW a specific output came about. An option treating them as the same is wrong. · Accountability stays with the organization. 'The vendor is responsible' or 'the algorithm decided' are trap answers. · Profit, speed, efficiency and innovation are never responsible-AI principles, however nicely an option dresses them up. · Human review that starts only after a complaint is remediation, not meaningful human control.
- ✓Principles become real through policy, procedure, control and evidence.
- ✓Common collisions: fairness vs. privacy, accuracy vs. explainability, transparency vs. security, accuracy vs. fairness, efficiency vs. human control.
- ✓Resolve collisions explicitly: legal floor, quantify, decide cross-functionally, document, monitor.
- ✓Match oversight (in, on or out of the loop) to stakes and reversibility, and guard against automation bias.
- Principle trade-off:
- A situation where advancing one responsible-AI principle, such as accuracy, costs another, such as fairness or privacy.
- Human-in-the-loop:
- A person must review and approve each decision before it takes effect.
- Human-on-the-loop:
- The system acts autonomously while a person monitors it and can intervene or halt it.
- Automation bias:
- The tendency of people to over-rely on automated outputs and stop exercising their own judgment.
- Operationalize:
- Turning a principle into concrete policies, procedures, controls and evidence.
Exam tips: When principles conflict, the IAPP answer is a documented, cross-functional balance with the law as the floor. Options that maximize one principle 'whatever it costs' are traps. · 'A human is somewhere in the workflow' isn't meaningful oversight. Look for competence, information, time and authority to override. · Leaving an ethical trade-off to the data science team alone is the technical-only trap. · Accountability = named owner + authority + records. Disclaimers and indemnities don't create it.
Establish and communicate organizational expectations for AI governance.
- ✓Create the structure first: check what exists, name a champion, decide who maintains it. Then implement.
- ✓Key roles: board, executive champion, governance committee, governance lead, system owners, specialist functions, internal audit.
- ✓Accountability needs one named owner per decision plus cross-functional oversight. Use RACI to make it explicit.
- ✓Three lines: owners manage risk, oversight functions set and challenge, internal audit assures.
- Executive champion:
- A senior leader who sponsors AI governance and gives it visible backing, authority and budget.
- AI governance committee:
- A cross-functional body that approves AI policy, reviews higher-risk use cases and resolves escalations.
- AI system owner:
- The business leader accountable for a specific AI system's purpose, performance and compliance.
- RACI:
- A chart that marks who is Responsible, Accountable, Consulted and Informed for each decision or task.
- Three lines model:
- A governance pattern splitting risk ownership (first line), oversight (second line) and independent assurance (third line, internal audit).
Exam tips: Creating the structure = (1) does one already exist? (2) name an executive champion (3) decide who maintains and implements it. Mapping laws, data lineage and training are implementation, so on a NOT question they're your answer. · Privacy and security experts are core stakeholders from day one. Options that leave them out, or add them 'later', are wrong. · 'One AI owner, no committee' and 'everyone is responsible' are opposite traps. The IAPP answer pairs an accountable executive with a cross-functional body. · Independent assurance comes from internal audit, not from the team that designed or runs the controls.
- ✓The BoK's two reasons for cross-functional governance: efficacy, and diversity of expertise and perspective.
- ✓Privacy and security are permanent seats; other functions join as AI reaches their territory.
- ✓Social scientists surface non-obvious bias that metrics and legal checks miss.
- ✓Go faster with tiered review, not by removing perspectives.
- Cross-functional collaboration:
- Bringing legal, privacy, security, technical, business, HR and other functions together to govern AI jointly.
- Diversity of expertise and perspective:
- Including different disciplines and backgrounds so the group spots risks a single viewpoint would miss.
- Non-obvious bias:
- Unfairness that standard technical and legal checks miss, such as culturally loaded language or subtle proxies.
- Tiered review:
- Scaling the depth of governance review to a use case's risk, so low-risk uses move fast and high-risk ones get full scrutiny.
Exam tips: When a competent technical and legal team keeps missing subtle bias, the IAPP answer is to add social scientists (psychologists, sociologists, linguists). · Watch for the true-but-irrelevant option: an executive sponsor or bigger budget sounds responsible but doesn't fix a detection gap. · Collapsing review into one executive is the LEAST effective way to speed up. Tiered, risk-based review keeps the diversity.
- ✓Train all stakeholders: board to contractors, not just the people who build models.
- ✓Cover AI terminology, strategy, governance and ethics; tailor depth by role on a shared baseline.
- ✓Training builds role skills; awareness keeps AI risk visible every day.
- ✓The EU AI Act expects providers and deployers to take measures to build staff AI literacy. Treat it as the floor.
- AI literacy:
- The skills, knowledge and understanding people need to use and oversee AI appropriately and to recognize its risks.
- Role-based training:
- Training whose depth and content match what a person's role requires, layered on a shared baseline.
- Awareness program:
- Ongoing communication that keeps AI risks, rules and reporting channels front of mind for everyone.
- Automation bias:
- The human tendency to over-trust an automated output instead of exercising independent judgment.
Exam tips: Any option that limits training to IT, data science or a committee is the too-narrow trap. The IAPP answer is all stakeholders, organization-wide. · The BoK's content list is AI terminology, strategy and governance, plus ethics. An option missing most of these isn't 'comprehensive'. · One-off launch training, or training only after an incident, is reactive. Look for continuous awareness and periodic refreshers.
- ✓PROMIS: products and services, risk tolerance, objectives, maturity, industry, size.
- ✓Centralized, decentralized or hybrid: hybrid fits large multinationals.
- ✓Leverage existing governance programs and add AI-specific pieces.
- ✓Maturity grows from ad hoc to optimized; start with inventory, policy, owner and risk triage.
- Risk tolerance:
- How much AI-related risk an organization is willing to accept in pursuit of its objectives.
- Proportionality:
- Scaling governance effort to the organization's context and the risk of each AI use.
- Centralized governance:
- One central function sets AI policy and makes AI decisions for the whole organization.
- Decentralized governance:
- Each business unit governs its own AI with little central coordination.
- Hybrid (federated) governance:
- Central policy and oversight combined with local assessment of use cases; the usual fit for large multinationals.
Exam tips: When two options sit at opposite extremes (centralize everything vs. every unit for itself) and a third reinvents the wheel, the IAPP answer is the balanced one that leverages existing structures. · 'Build AI governance from a blank page' is a trap when mature privacy, security or risk programs exist. Extend them. · Proportionate never means 'none'. Watch for options that exempt small or 'low-stakes' businesses entirely.
- ✓Developer builds, provider makes available, deployer puts to work, user interacts.
- ✓Each role has its own responsibilities, opportunities and needs, and information must flow between them.
- ✓One organization often holds several roles; duties attach to every role it holds.
- ✓Under the EU AI Act, provider and deployer are legal roles, and Art. 25 can turn a deployer into a provider.
- Developer:
- The party that designs, builds, trains and tests an AI model or system.
- Provider:
- The party that makes an AI system available to others; under the EU AI Act, it places the system on the market or puts it into service under its own name.
- Deployer:
- An organization that uses an AI system under its own authority, outside purely personal use.
- User:
- The individual who interacts with an AI system or acts on its outputs.
- Substantial modification:
- A significant change to an AI system after it reaches the market; under EU AI Act Article 25 it can turn a deployer into a provider.
Exam tips: Developer, provider, deployer and user are task labels, not exclusive identities. The IAPP answer usually says an organization holds several roles and owes each role's duties. · Integration or configuration is deployment, not development. Options that 'promote' integration work into development are traps. · Rebrand, substantially modify, or repurpose into high-risk = becomes the provider (EU AI Act Art. 25). · Deployers can't offload all responsibility onto the vendor, and calling yourself a 'mere user' doesn't change that.
Establish policies and procedures to apply throughout the AI life cycle.
- ✓Policies should cover every life-cycle stage, from use-case assessment to incident management.
- ✓Intake triages every AI proposal into a risk tier and records approved uses.
- ✓Risk policy sets the method, the tolerance and who may accept residual risk.
- ✓Ethics by design starts at planning; data rules cover rights, provenance, quality, minimization and retention.
- Use-case assessment:
- The intake review that captures an AI proposal's purpose, owner, data and affected people, and triages it by risk.
- Risk tier:
- A category (such as prohibited, high, medium or low) that sets how much review and control an AI use receives.
- Residual risk:
- The risk that remains after mitigations, which someone with the right authority must formally accept.
- Ethics by design:
- Identifying and resolving ethical issues from planning and design onward, so values shape the system rather than being checked at the end.
- Data provenance:
- A documented record of where data came from and how it has been transformed.
Exam tips: Planning order: business problem → use case → applicable laws → risks and gaps → data. Data-first or development-first options are wrong. · Ethics by design means from the design phase onward. A pre-launch ethics gate, even one with veto power, is the near-miss. · Laws attach to a defined use. 'Map all AI laws' before the use case exists is a tempting but premature answer.
- ✓Development: document, secure and vet components. Testing: separate data, criteria first, adversarial testing.
- ✓A readiness review gates deployment; approved uses and monitoring keep the system on track.
- ✓The inventory covers every system; a model card summarizes one model; decision records limit misuse and liability.
- ✓Incidents: detect, triage, contain, fix, learn, with notification running alongside.
- Readiness review:
- The go/no-go check before deployment that confirms testing, documentation, oversight, training and incident plans are in place.
- Scope creep:
- An AI system gradually being used for purposes beyond those it was assessed and approved for.
- AI inventory:
- An organization-wide register of AI systems with their owners, purposes, risk tiers and status.
- Model card:
- A concise document summarizing one model's purpose, intended use, performance and limitations.
- Red teaming:
- Deliberately attacking your own AI system, as an adversary would, to find weaknesses before release.
- AI incident:
- An event where an AI system's development, use or malfunction causes or nearly causes harm.
Exam tips: A policy that stops at deployment approval is incomplete. The IAPP answer adds monitoring, incident management, documentation and reporting. · Adversarial testing before launch is responsible. Options calling it unethical, or deferring it to production, are inversion traps. · An approved system used for a new purpose = scope creep. The fix is documented approved uses plus mandatory re-review, not accuracy monitoring. · Deactivation triggers are regulatory or performance problems, never profit and never 'a new version exists'.
- ✓Evaluate and update existing policies for AI rather than writing a parallel rulebook.
- ✓Privacy, security, data governance and IP each need AI-specific additions.
- ✓Out-of-cycle review triggers: AI into an existing program, a new AI program, a new regulation.
- ✓Inferences can be personal data, and AI outputs aren't automatically owned by the prompter.
- Gap analysis:
- Reviewing existing policies against AI's characteristics to find what they fail to address.
- Out-of-cycle review:
- An immediate policy review triggered by AI entering an existing program, a new AI program, or a new regulation.
- Prompt injection:
- An attack that hides malicious instructions in a model's input to hijack its behavior.
- Data poisoning:
- An attack that corrupts training data so the model learns harmful or wrong behavior.
- Data lineage:
- The traceable path of data from its source through every transformation into a model.
Exam tips: Three out-of-cycle triggers: AI into an existing program, a new AI program, a new regulation. Marginal changes ride the normal cadence. On LEAST/NOT questions, pick the marginal change. · The BoK's four examples of policies to update are data privacy, security, data governance and intellectual property. Expect them as options. · Training on content you don't have rights to is an IP problem first. Security is the tempting near-miss.
- ✓Third-party AI follows a life cycle: identify, assess, contract, onboard, monitor, exit.
- ✓Due diligence asks about training data rights, testing, documentation, data use, updates and sub-suppliers.
- ✓Contracts need AI-specific clauses, not just commercial boilerplate.
- ✓Map the supply chain behind the vendor: foundation models, data suppliers, open-source components.
- Third-party AI risk:
- Risk from AI systems, models, data or components supplied by vendors and their own suppliers.
- Vendor due diligence:
- Risk-scaled investigation of a supplier's AI: training data, testing, documentation, security, data use and update practices.
- Change-notification clause:
- A contract term requiring the vendor to tell you before material model updates or new AI features go live.
- Fourth party:
- Your vendor's own supplier, such as the company whose foundation model the vendor builds on.
- AI bill of materials:
- An inventory of the models, datasets and components that make up an AI product.
Exam tips: 'We couldn't tell which users were affected' → the answer is a third-party AI inventory plus documented decisions. Encryption or incident playbooks are near-misses. · The AI-specific clause set covers data-use limits, output ownership, change notification, testing evidence, incident cooperation and audit rights. Confidentiality and uptime are generic. · You can outsource the tool, never the accountability. 'The vendor is responsible' is a trap answer.
- ✓Third-party risk also enters through people: staff using outside AI tools, and vendors' AI used on staff.
- ✓The AUP binds people; contracts bind vendors; HR policies govern AI decisions about workers.
- ✓A good AUP covers approved tools, prohibited inputs and uses, human review, disclosure, accounts, reporting and consequences.
- ✓AI in employment decisions needs assessment, notice, human review, a way to contest, and vendor flow-down.
- Acceptable use policy (AUP):
- Rules binding employees and others on how they may use AI tools: approved tools, prohibited inputs and uses, review and reporting.
- Shadow AI:
- Employees using AI tools the organization hasn't approved or assessed.
- Prohibited inputs:
- Categories of data, such as confidential, personal or regulated information, that must never be entered into external AI tools.
- Meaningful human review:
- Review by a person with the competence, information and authority to change an AI-informed decision.
Exam tips: Employee conduct with unsanctioned tools → acceptable use policy. Procurement clauses are the near-miss when no contract exists. · An AUP clause claiming approved tools are 'error-free' or need 'no review' is always a trap. · Human editing of individual outputs doesn't cure systematic bias. 'A human reviews it' is a cure-all trap. · In employment, consent demanded as a condition of the job is a weak basis. Look for assessment, notice and human review instead.
Independent study aid, not affiliated with the IAPP. Written with AI from primary sources; laws change, so check the sources. Last updated 7 Nov 2026.
