A stem asks what a deployer must do FIRST. Which reading does the signal word require?

Sequencing — the earliest step, not the most important one.

Priority — the most important step, whenever it occurs.

Exclusion — the step that is not required of a deployer.

Frequency — the step performed most often in practice.

You have eliminated two options and the remaining two both look defensible. What is the most productive next move?

Name the single dimension on which the two differ.

Re-read both options until one begins to look weaker.

Pick the longer option, as it usually carries more detail.

Pick the one using the vocabulary from the stem.

Which of the following are reliable signals that an option is a distractor? (Select all that apply.)

It states a real duty but attaches it to the wrong operator.

It answers a part of the scenario the question did not ask about.

It contains an absolute such as "any", "must" or "always".

It names a genuine concept from one level down a taxonomy.

It is the shortest of the four options presented.

On a select-all item, you are confident about two options and unsure about a third. What does the scoring imply?

Judge each option independently, since partial credit is not available.

Select only the two you are sure of, to protect the marks you have.

Select all three, since more selections raise the chance of a match.

Skip the item, because an uncertain select-all is unlikely to score.

Exam technique

The AIGP tests judgment more than recall. Most questions present several defensible-sounding options and ask for the best one, so content knowledge gets you to the final two — technique picks between them. Drill this alongside the reference material, and re-read it the week before your exam.

Where this comes from. The patterns below are not guesses about how the exam is written — they are the patterns our own 200-question bank is built from, checked against the IAPP Body of Knowledge v2.1 and the primary sources. Every distractor we write encodes a specific misconception, and this page is the catalogue.

Work the STEM

Four passes over the question, in order, before you commit. The name is the point: you work the stem first, because the moment you start reading options they begin anchoring you.

S Signal word

What is the question actually asking?

Find the word that changes the task — MOST, FIRST, EXCEPT, LEAST — before you read a single option. Options are written to anchor you, so read the stem twice while they are still out of view.

T Territory

Which lens applies?

Fix four coordinates: which domain, which framework, which role, which lifecycle phase. Most wrong answers are true statements from the wrong coordinate, so pinning them down eliminates distractors before you evaluate content.

E Eliminate

Which options are built wrong?

Run the distractor patterns below. Absolutes, shifted accountability, single safeguards, wrong role, wrong tier, wrong framework. This usually takes four options down to two.

M Margin call

What separates the final two?

Both survivors are defensible — that is the design. Name the single dimension they differ on, then apply a tiebreaker. Commit and move on: the budget is about 1.8 minutes per question.

S · Spot the signal words

Read the stem twice before looking at the options. These words change what the question is asking:

"MOST important" / "PRIMARY" / "BEST"

Several options may be valid — pick the root cause or most foundational one, not merely a correct one.

"FIRST" / "BEFORE"

A sequencing question. Pick the earliest step in the lifecycle, not the most important overall.

"EXCEPT" / "NOT"

Flip your thinking: three options are correct — find the odd one out. Missing this word makes you pick the exact opposite of the right answer.

"LEAST likely" / "LEAST relevant"

Find the weakest connection. The other options will all be clearly connected, so you are ranking, not judging true/false.

Role words (provider, deployer, controller, processor, importer)

The answer depends on WHO. Verify the obligation belongs to that role before judging whether the obligation itself is real.

Framework words (EU AI Act, GDPR, NIST, OECD, ISO)

Match the answer to that specific framework — a real requirement from the wrong framework is the single most common distractor we write.

A qualifier inside one option ("adopted in 2019", "under Art. 27")

Read the whole option, not just the institution or instrument it names. The same body issues instruments decades apart on different subjects.

T · Fix the territory

Four coordinates, settled before you judge any option. Most wrong answers are true statements sitting at the wrong coordinate, so this pass alone removes them:

Which framework?

Controller and processor are GDPR. Provider, deployer, importer and distributor are the AI Act. Govern, Map, Measure and Manage are NIST. A real duty quoted from the wrong instrument is a distractor, however accurate it is.

Which role?

Providers carry the pre-market weight; deployers carry duties of use; importers verify; distributors check. Settle who the stem is asking about before you judge what they owe.

Which phase?

Use-case assessment decides whether to build. Conformity assessment precedes market placement. Monitoring, incidents and drift are post-deployment. Distractors are typically the right activity one stage out.

Which tier or level?

Prohibited, high-risk, limited or minimal — and for taxonomies, top-level category or subtype. A term one level down from the one asked for is a favourite distractor.

E · Eliminate by pattern

Wrong options are not random. They are built from a small set of recurring shapes, and if an option matches one it is almost certainly a distractor:

✗ 1Absolutes

"Always", "never", "all", "any", "must", "eliminates", "guarantees", "solely". Governance is proportionate, so an option can name the right treatment and still be wrong because it states it as a rule — "any catastrophic harm must eliminate the design" fails on the "must", not the idea.

✗ 2Shifted accountability

"The vendor is solely responsible", "the contract transfers liability". Contracts move duties and money; they never move answerability to a regulator.

✗ 3Single safeguard

"Anonymise the data and the concern is resolved", "add a disclaimer". Governance is layered — one control never closes everything.

✗ 4Right concept, wrong role

A genuine obligation pinned on the wrong party — the FRIA offered as a provider duty when Art. 27 places it on certain deployers.

✗ 5Right concept, wrong phase

A valid activity one or two stages too late. Use-case assessment before building; conformity assessment before market; monitoring after release.

✗ 6Right concept, wrong framework

NYC Local Law 144 bias audits offered as an ECOA duty; the GDPR’s 4% / €20M tier offered as the AI Act’s 7% / €35M.

✗ 7Right concept, wrong level

A real term from one level down. Sampling and temporal bias are real, but they sit inside NIST’s statistical/computational category rather than beside it.

✗ 8Wrong risk tier

Social scoring is prohibited, not high-risk. CV screening is high-risk, not prohibited. Chatbot disclosure is a transparency duty, not a high-risk requirement.

✗ 9Baseline dressed as an addition

Asked what applies BEYOND a baseline, an option names a genuine baseline duty — the GPAI training-content summary offered as a systemic-risk extra.

✗ 10Aggregate metrics

Overall accuracy used to justify deployment. Aggregates conceal subgroup failure, and on imbalanced data 98% accuracy can mean zero recall.

✗ 11Dismissing the concern

"Public data carries no privacy protection", "that is not really a risk". Options that wave a governance concern away are almost never right.

✗ 12Disproportionate response

"Cancel the project", "ban AI entirely". Governance seeks proportionate mitigation. Note the inverse trap too: doing nothing because a control exists on paper.

M · The margin call

On a well-written item the last two options are both defensible — that is the design, not a trick. The move is to name the single dimension they differ on, rather than re-reading both hoping one starts to look wrong. Six pairs the exam returns to:

Opacity vs Complexity Axis: Tense

Engineers with full model access cannot explain one specific credit refusal.

Complexity is forward-looking — too many interacting parts to predict what the system will do. Opacity is backward-looking — a specific output already exists and cannot be explained. The decision has been made, so it is opacity.

Data drift vs Concept drift Axis: What moved

A live model starts mis-predicting months after a clean launch.

Ask whether the correct answer for a given input changed, or only which inputs arrive. New customer segment, same notion of a good outcome, is data drift. Fraud tactics evolving so old signals stop meaning fraud is concept drift.

FRIA (Art. 27) vs DPIA (GDPR Art. 35) Axis: Which instrument asks

A public body deploys a high-risk system to triage benefits claims.

Both are probably required, which is what makes this hard. The AI Act’s answer for a public-body deployer is the FRIA; the DPIA is the GDPR’s. They are complementary, and organisations usually run one exercise satisfying both.

Deployer vs "User" Axis: Which draft you learned from

A company licenses a CV-screening tool and runs it on its own applicants.

Early AI Act drafts called deployers "users", so much study material still says it. The final text reserves the person-facing concept for the affected person — here the applicant, not the company.

RAG vs Fine-tuning Axis: Knowledge or behaviour

An assistant must answer from monthly-updated manuals and cite its source.

Fine-tuning changes how a model behaves — style, format, vocabulary — and every document change would mean retraining. Retrieval supplies current knowledge at inference time and makes citation natural.

Post-market monitoring vs Continuous monitoring Axis: Activity or obligation

What is the provider’s documented programme for reviewing field performance?

Continuous monitoring is the day-to-day practice of watching a live system. Post-market monitoring under Art. 72 is a documented plan a regulator can ask to see. One is what you do; the other is what you must be able to show.

Then apply a tiebreaker if the axis alone does not settle it:

→ Read what the question asked, not what the scenario described. A distractor is often a true statement about a part of the stem you were not asked about.
→ Rights and people beat performance and organisational convenience.
→ Proactive beats reactive — prevent before deployment rather than patch after.
→ Upstream beats downstream — data and design fixes beat output-level patches.
→ Layered controls beat single safeguards.
→ "It depends on risk and context" beats "always" or "never".
→ Accountability stays with the organisation even when the work is delegated.
→ A broad category usually beats its own specific example (the umbrella rule).

The clock

100 questions in 180 minutes is about 1.8 minutes each. A repeatable routine is what keeps you disciplined when the options look similar and the clock is visible:

~15s Signal word. Read the stem twice, options still hidden. Name what is being asked.
~10s Territory. Which domain, framework, role and lifecycle phase?
~20s Eliminate. Run the distractor patterns — usually takes four options to two.
~15s Margin call. Name the axis the final two differ on, apply a tiebreaker, commit, move on.

Question formats to expect

Beyond standard single-best-answer items, these formats appear — each with its own discipline:

Multi-question scenarios

A long business scenario — named company, named people, often international — followed by two to four questions. Read it once, carefully, noting jurisdictions, each actor's role, the lifecycle phase, and any red-flag facts such as special-category data, automated decisions or children. Each question probes a different angle of the same facts, so consult your mental map rather than re-reading the whole scenario each time.

"Select all that apply"

You are not told how many options are correct, and there is no partial credit. Judge each option independently as true or false against the stem, then select every true one. Do not second-guess yourself by expected count — three correct out of five is as legitimate as two.

Ordering and sequencing

Occasionally you must put lifecycle steps in order, scored all or nothing. Two anchor sequences to internalise: business problem → use cases → applicable laws → necessary data → gaps and risks for planning, and plan → design and data → build and train → test → readiness → deploy → monitor → retire for the lifecycle. Strategy precedes law precedes data precedes risk.

Exam-day reminders

Don't argue with the question — accept its premise and answer within it.
Select-all items don't reveal how many answers are correct, and there is no partial credit. Judge each option on its own merits.
Don't try to spot unscored pilot questions — answer everything with full effort.
Flag and move on rather than stall; return with fresh eyes in the review pass.
Scoring is reported on a 100–500 scale with a pass mark of 300, so you need consistency rather than perfection. Verify current details in the IAPP Candidate Handbook.
The night before: skim the cheat sheet, then rest. Cramming past midnight costs more than it gains.

Technique is worth little without the material behind it. Work through the blueprint map to see what each domain is worth, and the cheat sheet for the facts most often tested.

Test yourself

Five questions on what is above, with every option explained. Your score is kept in this browser and shown on your dashboard, and saved to your account if you are signed in.

Quick check: spot the distractor pattern1 / 5 · score 0

An option reads: "Anonymising the training data resolves the privacy concerns." Which distractor pattern is this?

AShifted accountability — responsibility moved to another party.
BWrong risk tier — a use case placed in the wrong category.
CAbsolutes — an unqualified claim stated as a universal rule.
DSingle safeguard — one control presented as closing everything.