IAPP AIGP · Domain II · Laws, Standards & Frameworks

Know the Law, Map the Rules

Domain II is the legal layer of the AIGP: which rules bind an AI system, who owes what duty, and which assessment proves you did the homework. GDPR, the EU AI Act, the US patchwork, ISO 42001, OECD, UNESCO, Singapore — each with a plain-words example so the law stops being jargon.

GDPR Art. 22 · 35 · Recital 26 EU AI Act fine tiers + dates NIST G-M-M-M · ISO PDCA SG PDPA · MGF · FEAT 20 flip cards · 8 traps
How to use this module: read each section, then flip the cards and run the trap quiz. The compare tables at §7 are the highest-yield revision — the exam loves blurring GDPR vs PDPA, DPIA vs FRIA, and provider vs deployer. Every topic carries an IN PLAIN WORDS example, because the exam asks you to explain these rules to a non-technical person.
Scroll to begin ↓
0
The territory at a glance

Domain II Mindmap

Eight branches, one exam domain. Trace each line: the law, its structure, and the exam tell that identifies it.

Domain II Laws · Standards · Frameworks GDPR Art. 22 · 35 · Recital 26 EU AI Act Fines · prohibitions · dates US landscape Sectoral · NIST RMF · states ISO · OECD · UNESCO PDCA · 5 principles · ethics Singapore PDPA · MGF · AI Verify · FEAT Other laws IP · consumer · employment Roles Provider vs deployer Assessments DPIA · FRIA · AIA
1
BoK Domain II.A — The EU privacy baseline

GDPR: The Rules AI Must Live With

The GDPR is the privacy law the AIGP assumes you know cold. The exam tests four things: automated decisions (Art. 22), impact assessments (Art. 35), what counts as personal data (Recital 26), and who owes what duty (controller vs processor).

1.1 Automated decision-making — Art. 22

Exam anchor — Art. 22Article 22 gives a data subject the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them. Three exceptions: (a) necessary for a contract, (b) authorised by law, (c) based on explicit consent. For the first two exceptions, the controller must provide suitable safeguards — including the right to human intervention, to express views, and to contest the decision.[GDPR Art. 22]
In plain words: a bank's AI denies your loan application with no human involved. Under Art. 22 you can demand a human re-look at the decision, explain your side, and challenge it. The bank must have a human-review path built in — not as a favour, but as a legal requirement.
Exam trap — "solely"Art. 22 only bites when the decision is solely automated. If a human meaningfully reviews the decision before it takes effect, Art. 22 does not apply. The exam loves a scenario where a human clicks "approve" without reading — that is still solely automated in substance, because the human adds no real judgement.

1.2 DPIA — Art. 35

When a DPIA is requiredA Data Protection Impact Assessment is mandatory where processing is likely to result in a high risk to individuals' rights and freedoms — especially new technologies, systematic evaluation/profiling, large-scale special-category data, or large-scale public monitoring. The DPIA must describe the processing, assess necessity and proportionality, and identify and mitigate risks.[GDPR Art. 35]
In plain words: a hospital wants an AI that screens all patient records for clinical-trial eligibility. That is large-scale special-category data — a DPIA is legally required before the system touches real records. The DPIA is the paperwork that proves the hospital thought about the risks before building.
Exam anchor — the DPIA checklistWhat a DPIA must contain: (1) a systematic description of the processing and its purposes; (2) an assessment of necessity and proportionality; (3) an assessment of risks to individuals; (4) the measures to address those risks. If the residual risk stays high, the controller must consult the supervisory authority before processing (Art. 36).

1.3 Personal data & Recital 26

ConceptDefinitionPlain-words example
Personal dataAny information relating to an identified or identifiable natural personA customer's name, email, location, purchase history, or voice recording — alone or combined
IdentifiableCan be identified directly or indirectly, using means reasonably likely to be usedA pseudonymised user ID is still personal data if the company holds the key
Anonymised (Recital 26)Information that no longer relates to an identifiable person — GDPR does not applyAggregated statistics with no way back to individuals are outside the GDPR
Pseudonymised (Art. 4(5))Data that can no longer be attributed without additional information kept separatelyReplacing names with codes — still personal data, but a strong safeguard
Exam trap — pseudonymised ≠ anonymisedPseudonymised data is still personal data under the GDPR. Only true anonymisation takes data out of scope. The exam asks this directly — if the key exists anywhere, it is pseudonymised, not anonymised.[GDPR Recital 26; Art. 4(5)]

1.4 Controller vs processor

RoleWhoCore dutyPlain-words example
ControllerDetermines the purposes and means of processingOwns lawful basis, notices, DPIA, data-subject rights, breach notificationA hospital decides to deploy a diagnostic AI and how patient data will be used — the hospital is the controller
ProcessorProcesses data on the controller's behalfActs only on documented instructions; must have a contract (Art. 28)The cloud vendor hosting the hospital's data — it cannot use the data for its own purposes
Joint controllersTwo or more jointly determine purposes and meansMust arrange responsibilities transparently between themA hospital and a university jointly design a research AI — both decide how data is used, so both are controllers
Why this matters for AI: when a company buys an AI tool from a vendor, the vendor is usually the processor and the buyer the controller — but if the vendor trains the model on the buyer's data for its own benefit, the vendor becomes a controller too. The exam tests who carries the breach-notification duty: the controller.

1.5 Data subject rights the exam tests

Right of access (Art. 15)

Individuals can ask what data is processed, why, and who it is shared with.

Example: a customer asks the bank what the credit model knows about them — the bank must answer within one month.

Right to rectification & erasure (Art. 16-17)

Correct inaccurate data; erase data when grounds apply (the "right to be forgotten").

Example: a job applicant asks a recruiter to delete their profile after a failed application.

Right to object (Art. 21)

Object to processing based on legitimate interests or for direct marketing.

Example: a user opts out of a retailer's AI-driven personalised ads.

Right to data portability (Art. 20)

Receive data in a machine-readable format and transfer it to another provider.

Example: a fitness-app user exports their health data to switch to a rival app.

Right not to be subject to automated decisions (Art. 22)

Human intervention, express views, contest decisions made solely by automation.

Example: an AI denies a loan — the applicant demands a human review.

Right to lodge a complaint (Art. 77)

Complain to a supervisory authority in the member state of residence, work, or alleged infringement.

Example: a data subject whose AI-rejected claim was never reviewed files a complaint with the DPA.
2
BoK Domain II.A — Exam-relevant summary only

EU AI Act: The Fines, The Bans, The Dates

You have the deep EU module — this is the exam-ready skeleton: fine tiers, the eight prohibited practices, the key dates (including the Digital Omnibus shifts), and the traps that separate a pass from a fail.

2.1 Fine tiers — Art. 99

TierViolationFineMnemonic
Top tierProhibited practices (Art. 5)€35M or 7% of total worldwide annual turnover"7% for the seven deadly practices"
Middle tierMost operator obligations, incl. Art. 50 transparency€15M or 3%"3% for the three letters of Art. 50"
Lower tierSupplying incorrect/incomplete/misleading info to authorities€7.5M or 1%"1% for the one thing you should have told them"
Exam anchor — the SME lower-of rule (Art. 99(6))For SMEs and start-ups, the fine is the lower of the fixed amount or the percentage of turnover — not the higher. For everyone else it is the higher. The exam tests this with a small company and a big turnover figure: pick the smaller number.[EU AI Act Art. 99(6)]
In plain words: a big bank runs a prohibited social-scoring system — fine is the higher of €35M or 7% of its global turnover. A 20-person startup does the same — the fine is the lower of the two. The law protects small players from being wiped out.

2.2 The eight prohibited practices — Art. 5(a)-(h)

(a) Subliminal manipulation

Techniques that distort behaviour and cause harm, exploiting vulnerabilities.

Example: an app flashes hidden cues to push a vulnerable user into harmful purchases.

(b) Vulnerability exploitation

Exploiting age, disability, or socio-economic situation to distort behaviour.

Example: a toy chatbot manipulates a child into revealing family details.

(c) Social scoring

Public or private social scoring leading to detrimental treatment.

Example: a government scores citizens' "trustworthiness" and denies benefits based on the score.

(d) Individual predictive policing

Predicting a person's risk of offending based solely on profiling.

Example: police flag a person as a future criminal based on their profile alone — banned; crime-pattern analysis is a different, high-risk category.

(e) Untargeted facial scraping

Scraping the internet or CCTV for facial images to build or expand databases.

Example: a company harvests social-media photos to train a facial-recognition database.

(f) Emotion inference at work/school

Inferring emotions in workplaces and education.

Example: a call-centre AI reads agents' emotions from their voices to score performance — banned in the workplace.

(g) Sensitive biometric categorisation

Categorising people by race, religion, sexual orientation, etc., from biometrics.

Example: a system classifies faces by presumed sexual orientation — banned.

(h) Real-time remote biometric ID in public

Real-time identification in public spaces, with three narrow exceptions (terrorism, serious crime, search for victims) and judicial authorisation.

Example: police may use live facial recognition to find a kidnapped child — but only with a court order and time/place limits.
Exam trap — emotion recognition in retailEmotion recognition in a retail store is limited-risk (Art. 50 transparency duty), NOT prohibited. Only emotion inference in workplaces and education is banned under Art. 5(f). The exam sets this trap deliberately — read the context before you pick "prohibited".

2.3 Key dates — including the Digital Omnibus shifts

ObligationOriginal dateOmnibus shift
Prohibitions (Art. 5)2 Feb 2025Unchanged
GPAI transparency (Art. 50)2 Aug 2025Unchanged
General application + Art. 502 Aug 2026Unchanged
Annex III high-risk obligations2 Aug 2026→ 2 Dec 2027
Annex I (embedded in regulated products)2 Aug 2027→ 2 Aug 2028
Exam memory hook — "timeline relief, not repeal"The Digital Omnibus (in force 27 Jul 2026) delayed the Annex III and Annex I obligations but did not remove them. If a question says the EU "dropped" or "scrapped" high-risk rules, that is wrong — the obligations still land, just later.[Digital Omnibus, Council final approval 29 Jun 2026]

2.4 Systemic-risk GPAI & the 10^25 threshold

Exam anchor — Art. 51A general-purpose AI model is presumed to pose systemic risk when its training compute exceeds 10^25 FLOPs. Systemic-risk models face extra duties: risk assessment, adversarial testing, incident reporting, and cybersecurity protection.[EU AI Act Art. 51]
In plain words: a frontier lab trains a model on a massive cluster — past the 10^25 FLOPs line, the model is presumed dangerous enough to need extra testing and reporting. The number is the exam tell: 10 to the power of 25.
3
BoK Domain II.A — The US patchwork

The US: Sectoral Laws, No Federal Omnibus

The US has no single federal AI law. Instead, a patchwork of sectoral statutes, state laws, and voluntary frameworks. The exam tests which law covers which activity — and the fact that NIST's framework is voluntary, not mandatory.

3.1 The sectoral statutes

LawWhat it coversPlain-words example
HIPAAHealth data — privacy, security, breach notification for covered entitiesA hospital's AI that reads patient records must follow HIPAA's security rules for that data
FCRACredit reporting and consumer reports — accuracy, disputes, adverse-action noticesAn AI that scores a loan applicant is a "consumer report" — the applicant must be told and given a dispute path
ECPAWiretapping and electronic communications interceptionAn AI that records and analyses customer-service calls must respect consent rules for interception
FTC Act §5Unfair or deceptive acts or practicesAn AI chatbot that hides it is a bot, or a model that silently discriminates, can be an FTC enforcement target
ADADisability discriminationAn AI hiring tool that filters out candidates with disabilities can violate the ADA
Title VIIEmployment discriminationAn AI resume screener with disparate impact on protected groups faces Title VII liability
Exam anchor — FCRA and AIWhen an AI system produces a consumer report used for credit, employment, or insurance decisions, the FCRA's adverse-action notice duty applies: the consumer must be told the decision and the reasons, and given a chance to dispute. The exam pairs this with GDPR Art. 22 — same idea, different statute.

3.2 State laws

CCPA / CPRA (California)

Consumer privacy rights: access, deletion, opt-out of sale/sharing, and (CPRA) opt-out of automated decision-making in some cases.

Example: a Californian user asks a retailer's AI to stop using their data for personalised pricing.

BIPA (Illinois)

Biometric privacy — written consent before collecting biometric identifiers.

Example: a gym's facial-recognition check-in must get written consent from members first.

State AI laws (2024-2026 wave)

Colorado AI Act, Utah AI Policy Act, California AI transparency bills — a growing patchwork with notice, disclosure, and risk-assessment duties.

Example: a Colorado chatbot must disclose it is AI; a California deepfake must be labelled.

3.3 NIST AI RMF 1.0 — the voluntary framework

Exam anchor — G-M-M-MNIST AI RMF 1.0 (Jan 2023) has four functions: Govern – Map – Measure – Manage. Govern is the core that wraps the other three. It is voluntary — the exam contrasts it with the EU AI Act's mandatory compliance.[NIST AI RMF 1.0]
FunctionWhat it doesPlain-words example
GovernRisk-management culture, policies, roles, accountability — the coreThe board approves an AI risk appetite and names an AI governance officer
MapUnderstand context: use case, stakeholders, benefits, risksThe team documents who the hiring model affects and what could go wrong for each group
MeasureTest and evaluate: metrics, bias testing, red-teaming, baselinesModel accuracy is measured per demographic group, not just overall
ManageAct on findings: mitigate, implement controls, monitorA bias finding triggers retraining with rebalanced data and a new threshold
Exam trap — voluntary vs mandatoryNIST AI RMF is voluntary. The EU AI Act is mandatory. A question that implies NIST compliance is legally required is wrong — unless a specific law (like a state statute) incorporates it.
4
BoK Domain II.B — The standards layer

ISO 42001, OECD Principles, UNESCO

Three international instruments, three different legal natures: a certifiable management standard, a set of non-binding principles, and an ethics recommendation. The exam tests which is which.

4.1 ISO/IEC 42001:2023 — the certifiable AI management system

Exam anchor — PDCA + certifiableISO/IEC 42001:2023 is the first certifiable AI management-system standard. It follows the Plan – Do – Check – Act (PDCA) cycle, like ISO 27001 for information security. Annex A lists controls across the AI lifecycle — from data acquisition to model deployment and monitoring.[ISO/IEC 42001:2023]
PDCA phaseWhat happensPlain-words example
PlanSet the AI policy, objectives, risk appetite, and scope of the management systemA consultancy writes its AI policy and decides which systems the management system covers
DoRun the AI lifecycle with the controls: risk assessments, data governance, documentationThe team builds a credit model with documented data lineage and bias tests
CheckAudit, measure, and review performance against objectivesAn internal audit finds the model's monitoring thresholds were never set
ActCorrect, improve, and feed findings back into the systemThe team fixes the thresholds and updates the policy so it cannot recur
In plain words: ISO 42001 is like a quality-management badge for AI. A company that gets certified proves it has a working system of policies, risk assessments, and audits — not just a one-off checklist. The exam tell: "certification" and "management system" point to ISO 42001.

4.2 OECD AI Principles

Exam anchor — five principles, non-bindingThe OECD AI Principles (2019, updated 2024) are non-binding recommendations: (1) inclusive growth and sustainable development; (2) human-centred values and fairness; (3) transparency and explainability; (4) robustness, security and safety; (5) accountability. They are the international baseline most other frameworks reference.[OECD AI Principles]
In plain words: when a government writes its national AI strategy, it usually starts by saying "we follow the OECD principles" — a shared vocabulary for trustworthy AI. They are guidance, not law: no regulator fines you for breaching an OECD principle.

4.3 UNESCO Recommendation on the Ethics of AI

Exam anchor — the ethics layerUNESCO's Recommendation on the Ethics of AI (2021) is the first global normative instrument on AI ethics. It sets values and principles — human dignity, human rights, transparency, fairness, accountability — and asks member states to implement them through law, policy, and education. Like the OECD principles, it is non-binding but influential.[UNESCO Recommendation on the Ethics of AI, 2021]
In plain words: UNESCO is the "ethics charter" layer — the values that laws and standards are supposed to operationalise. If a question asks which instrument is about ethics values rather than enforceable rules, the answer is UNESCO.
5
BoK Domain II.A — The Singapore stack

Singapore: PDPA, MGF, AI Verify, FEAT

Singapore's approach is framework-first: a privacy law (PDPA), two governance frameworks (MGF 2020 + GenAI 2024), a testing toolkit (AI Verify), and a sectoral fairness guide for finance (FEAT).

5.1 PDPA — the privacy law

Exam anchor — PDPA basicsThe Personal Data Protection Act (PDPA) governs personal data in Singapore: consent (or deemed consent / legitimate interests), purpose limitation, notification, access and correction, retention limits, and transfer restrictions. The PDPC enforces it. Unlike the GDPR, the PDPA has no general right to object to automated decisions — but the PDPC's advisory guidelines address AI and automated decision-making.[PDPA 2012, as amended]
In plain words: a Singapore retailer's AI loyalty engine must tell customers what data it collects and why, get consent where needed, and delete data when the purpose ends. The PDPC can fine for breaches — up to 10% of annual turnover in Singapore (or S$1M, whichever is higher) for serious breaches.

5.2 Model AI Governance Framework (MGF) 2020 + GenAI 2024

FrameworkStructurePlain-words example
MGF 2nd ed. (2020)Four areas: internal governance, human oversight, operations management, stakeholder interactionA fintech sets human oversight by weighing harm severity × probability against commercial considerations
GenAI Framework (2024)Nine dimensions: accountability, safety, transparency, fairness, data governance, security, incident reporting, testing, international cooperationA chatbot vendor documents its model's limitations, tests for jailbreaks, and publishes an incident-reporting channel
Exam anchor — risk-weighted human oversightThe MGF's signature concept: the level of human oversight should be proportional to the risk — severity × probability of harm. High-risk, high-impact AI gets human-in-the-loop; low-risk gets lighter oversight. The exam contrasts this with the EU's fixed risk tiers.

5.3 AI Verify & MAS FEAT

AI Verify

Singapore's testing toolkit and governance framework for AI — a self-assessment and testing suite covering transparency, explainability, fairness, robustness, and safety.

Example: a company runs AI Verify's tests on its hiring model and publishes the results as a transparency artefact.

MAS FEAT

The Monetary Authority of Singapore's principles for Fairness, Ethics, Accountability, and Transparency in financial AI — aimed at banks and insurers.

Example: a bank's credit-scoring AI must be explainable to customers and fair across demographic groups, per FEAT.
Why this matters: Singapore is the exam's "framework-first" jurisdiction — voluntary frameworks plus a privacy law, rather than a dedicated AI statute. If a question asks which Singapore instrument is legally binding, the answer is the PDPA; the MGF, AI Verify, and FEAT are guidance.
6
BoK Domain II.C — The surrounding law

Other Laws That Touch AI

AI does not live in a privacy bubble. Copyright, trade secrets, consumer protection, employment, and sectoral rules all apply — and the exam tests which law answers which problem.

6.1 IP & copyright

IssueThe rulePlain-words example
Training on copyrighted worksDepends on jurisdiction: fair use (US), text-and-data-mining exceptions (EU), or licence requirementsA generative-AI vendor must confirm its training corpus is licensed or within an exception — or face infringement claims
AI-generated outputCopyright in AI output is unsettled; human authorship is usually requiredA company cannot assume its AI-generated marketing images are copyrightable — check the jurisdiction's authorship rules
Trade secretsProtecting model weights, training data, and prompts as confidential informationA vendor's model weights are its crown jewels — the licence must restrict reverse engineering and disclosure
Exam anchor — the IP governance angleBoK v2.1 added IP governance: organisations need policies for what AI may ingest, what it may output, and who owns the rights. The exam tests the practical question — "which clause matters in this AI vendor contract?" — and the answer is usually data ownership and IP indemnities.

6.2 Consumer protection & product liability

Consumer protection

Unfair or deceptive practices, product safety, and advertising rules apply to AI products and services.

Example: an AI "health coach" that makes unverified medical claims is a consumer-protection problem, not just a privacy one.

Product liability

Who is liable when an AI product causes harm — the manufacturer, the software vendor, or the deployer?

Example: a defective AI-assisted medical device injures a patient — the liability chain runs through the manufacturer and the deployer's duties.

Liability reform

New rules (e.g., EU AI Liability Directive proposals) change how damages from AI systems are allocated.

Example: a proposed EU directive shifts the burden of proof so a harmed consumer does not have to explain the AI's inner workings.

6.3 Employment & sectoral law

AreaHow it appliesPlain-words example
Employment lawHiring, monitoring, and termination decisions by AI must not discriminate; notice and transparency duties applyAn AI that screens resumes must be tested for disparate impact on protected groups
Workplace monitoringRecording and analysing workers (calls, keystrokes, emotion) is restricted by privacy and labour rulesA call-centre AI that scores agents' emotions may breach workplace-privacy rules
Sectoral regulatorsFinance (MAS, SEC, FCA), health (HSA, FDA, EMA), and other regulators impose their own AI expectationsA bank's credit model must satisfy the MAS's FEAT principles on top of the PDPA
In plain words: the same AI system can trip five different laws at once — privacy (PDPA/GDPR), discrimination (employment law), consumer protection, IP, and a sectoral rule. The exam asks you to pick which law answers a specific harm, so learn the mapping, not just the names.
7
Easily confused — exam differentiators

Compare & Contrast

Four comparison tables that decide Domain II questions. If you can fill these from memory, the legal layer is largely yours.

7.1 GDPR vs PDPA vs US

GDPR (EU)PDPA (SG)US (federal)
NatureComprehensive privacy regulationComprehensive privacy regulationSectoral statutes + state laws; no omnibus
Automated decisionsArt. 22 right not to be subject to solely automated decisionsNo general right; PDPC guidelines address AIFCRA adverse-action notice; state laws (e.g., CPRA opt-out)
Impact assessmentDPIA where high risk (Art. 35)DPIA-style assessments encouraged; PDPC guidanceNo general federal DPIA; state AI laws add risk assessments
EnforcementDPAs; fines up to €20M/4%PDPC; fines up to 10% of turnover or S$1MFTC, sectoral agencies, state AGs
Exam tell"solely automated", "DPIA", "Art. 22""consent", "PDPC", "notification""sectoral", "FCRA", "no federal omnibus"

7.2 DPIA vs FRIA vs AIA

DPIA (GDPR Art. 35)FRIA (fundamental rights)AIA (algorithmic impact assessment)
FocusPersonal-data protection risksFundamental rights and freedomsAlgorithmic/systemic risks and impacts
TriggerHigh risk to individuals' rights from processingHigh-risk AI per EU AI Act (Art. 27)Organisational policy or law (e.g., NYC Local Law 144)
WhoControllerDeployer of high-risk AIDeployer / organisation
Exam tell"personal data", "Art. 35""fundamental rights", "Art. 27""algorithmic", "bias audit"

7.3 ISO 42001 vs NIST RMF

ISO/IEC 42001NIST AI RMF
TypeManagement-system standard (certifiable)Risk-management framework (voluntary)
StructurePlan–Do–Check–ActGovern–Map–Measure–Manage
Best forFormal certification and auditsDesigning risk processes
Exam tell"certification", "management system""functions", "GMMM"

7.4 Provider vs deployer obligations

ProviderDeployer
WhoDevelops the AI system and places it on the marketUses the AI system in its operations
Core dutiesConformity assessment, technical documentation, quality management, risk management, registrationUse per instructions, human oversight, monitoring, data governance, transparency to users, incident reporting
Exam tell"developed", "placed on the market", "conformity""uses", "oversight", "monitoring", "instructions"
Exam trap — the distributor is not the providerUnder the EU AI Act, a distributor or importer is a separate operator role in the supply chain — not the provider. The provider is the entity that developed the system and placed it on the market. If a question says "the provider is the company and its distribution network", that is wrong.[EU AI Act Art. 3]
8
Active recall · 20 cards

Flip Cards: The Legal Layer

Front = term. Back = definition + exam tip. Click to flip; keyboard Enter/Space works too. Best score persists in this browser.

9
Self-test · Exam traps

Eight Exam Traps

TRUE / FALSE — instant feedback per question, score tallied at the bottom. Best score persists in this browser.

10
Exam-day readiness

Before Thursday — The Checklist

Tap each item as you master it. The goal: fill every box from memory, not recognition.

§
Citation ledger

Sources