An organisation says its training set is "anonymised" because direct identifiers were replaced with random tokens held in a separate key file. Is the GDPR engaged?

Yes — a retained key makes this pseudonymisation, so it stays personal data.

No — replacing direct identifiers takes the data outside the Regulation.

No — provided the key is held by a different team from the model builders.

Yes — but only while the key exists, and not once it is eventually deleted.

A processor discovers a personal-data breach in a system it runs for a controller. What does the GDPR require of the processor?

Notify the controller without undue delay.

Notify the supervisory authority within 72 hours of awareness.

Notify the affected individuals where the risk to them is high.

Nothing, since breach duties attach only to the controller.

GDPR for the AIGP

The AIGP is not a privacy exam, but it assumes you can apply the GDPR to AI. This page covers every article the exam realistically tests — in the order the exam reasons about them.

1 · Scope and definitions (Arts. 2–4)

Art. 3 territorial scope — applies to an EU establishment, or to offering goods/services to, or monitoring the behaviour of, people in the Union. Behaviour monitoring is what pulls non-EU AI products in.
Art. 4(1) personal data — any information relating to an identified or identifiable natural person, directly or indirectly.
Art. 4(4) profiling — automated processing to evaluate personal aspects, in particular to analyse or predict performance, economic situation, health, preferences, reliability, behaviour, location or movements.
Art. 4(5) pseudonymisation — still personal data. Anonymous data falls outside the GDPR entirely (Recital 26).
Art. 4(7)/(8) controller and processor — controller decides purposes and means.

2 · The seven principles (Art. 5)

(a) Lawfulness, fairness, transparency	Needs a lawful basis, no unjustified adverse effects, and people must understand what happens. "Fairness" is where discriminatory model outcomes get caught even when a basis exists.
(b) Purpose limitation	Collected for specified, explicit, legitimate purposes; no incompatible further processing. The classic AI failure: reusing operational data as training data without a compatibility assessment.
(c) Data minimisation	Adequate, relevant, limited to what is necessary. Directly in tension with "more data is better" — the exam expects you to side with necessity.
(d) Accuracy	Personal data must be accurate and kept up to date. Applies to training data and to inferences/outputs about a person — a wrong prediction can be inaccurate personal data.
(e) Storage limitation	Kept identifiable no longer than necessary. Drives retention schedules for training sets and logs.
(f) Integrity and confidentiality	Appropriate security. Ties to Art. 32 and to AI-specific threats (model inversion, membership inference, training-data extraction).
Art. 5(2) Accountability	You must not only comply but demonstrate compliance. This is the hook for RoPAs, DPIAs and documentation — and the reason "we did it, trust us" fails.

3 · Lawful bases (Art. 6) and special categories (Art. 9)

(a) Consent	Freely given, specific, informed, unambiguous; withdrawable as easily as given (Art. 7). Rarely workable for training on historic data at scale.
(b) Contract	Necessary to perform a contract with the data subject, or pre-contractual steps. "Necessary" is strict — not merely useful.
(c) Legal obligation	Required by EU/member-state law.
(d) Vital interests	Life-or-death. Rare.
(e) Public task	Official authority or task in the public interest.
(f) Legitimate interests	The usual basis for AI development. Requires a three-part LIA: identify the interest, test necessity, balance against the data subject’s rights and reasonable expectations. Not available to public authorities performing their tasks.

Art. 9 prohibits processing racial/ethnic origin, political opinions, religious or philosophical beliefs, trade-union membership, genetic data, biometric data processed to uniquely identify someone, health data, and sex life/sexual orientation — unless an Art. 9(2) condition applies (most often (a) explicit consent, (g) substantial public interest, or (j) research).

4 · Transparency (Arts. 12–14)

Art. 12 — concise, transparent, intelligible, easily accessible, plain language. Rights requests answered within one month, extendable by two further months for complexity.
Art. 13 — what to tell people when you collect data from them.
Art. 14 — what to tell people when you did not get it from them (scraped or third-party training data), at the latest within one month. Art. 14(5)(b) exempts you where notification is impossible or involves disproportionate effort — the provision every web-scraping argument leans on, and one regulators read narrowly.
Arts. 13(2)(f) / 14(2)(g) — where there is automated decision-making under Art. 22, you must disclose its existence, meaningful information about the logic involved, and the significance and envisaged consequences.

5 · Data subject rights (Arts. 15–22)

Art. 15	Access	Copy of the data plus purposes, recipients, retention — and, where applicable, meaningful information about automated decision logic.
Art. 16	Rectification	Correct inaccurate data. Hard against a trained model; usually met at the input/output layer.
Art. 17	Erasure	Not absolute. Raises the unresolved question of whether a model that memorised data must be retrained or unlearned.
Art. 18	Restriction	Pause processing while accuracy or objection is resolved.
Art. 20	Portability	Only where processing is by consent or contract AND automated. Covers data provided by the subject — not derived inferences.
Art. 21	Object	Applies to legitimate-interests and public-task processing, incl. profiling. Absolute for direct marketing.
Art. 22	Automated decisions	The centrepiece — see below.

6 · Article 22 — solely automated decisions

The single most examined GDPR provision on the AIGP. Work through it as a four-step test.

Is it solely automated? No meaningful human involvement. A human who cannot realistically override the output does not break "solely".
Legal or similarly significant effect? Credit refusal, job rejection, insurance pricing, benefits decisions — yes. Ordinary ad targeting — usually not.
If both are true, it is prohibited unless one of Art. 22(2) applies: (a) necessary for a contract, (b) authorised by EU/member-state law, or (c) explicit consent.
Then apply Art. 22(3) safeguards: at minimum the right to obtain human intervention, to express a point of view, and to contest the decision.

Art. 22(4): such decisions may not be based on special-category data unless Art. 9(2)(a) or (g) applies with suitable safeguards. Note that the widely cited "right to an explanation" appears in Recital 71, not in the binding article text — the articles require meaningful information about the logic, which is a lower bar.

7 · Article 35 — DPIAs

Required where processing is likely to result in a high risk, in particular when using new technologies. Art. 35(3) lists three mandatory triggers:

(a) systematic and extensive evaluation of personal aspects based on automated processing, including profiling, on which decisions producing legal or similarly significant effects are based;
(b) large-scale processing of special-category or criminal-offence data;
(c) systematic monitoring of a publicly accessible area on a large scale.

Art. 35(7) sets the minimum content: a systematic description of the processing, an assessment of necessity and proportionality, an assessment of risks to individuals, and the measures addressing those risks. Consult the DPO (Art. 35(2)); where high risk remains after mitigation, Art. 36 requires prior consultation with the supervisory authority before processing begins.

8 · Accountability and security (Arts. 24–39)

Art. 24	Controller must implement appropriate technical and organisational measures and be able to demonstrate them.
Art. 25	Data protection by design and by default — the article behind "governance at design time, not after launch". Default settings must be the privacy-protective ones.
Art. 26	Joint controllers must agree and make available the essence of who does what.
Art. 28	Processors: written contract with mandated terms (documented instructions, confidentiality, security, sub-processor consent, assistance, deletion/return, audits). Your AI vendor is usually a processor — until it uses your data to improve its own model, at which point it is likely a controller.
Art. 30	Records of processing activities (RoPA). Your AI inventory should map to it.
Art. 32	Security appropriate to risk: pseudonymisation, encryption, CIA + resilience, and regular testing.
Arts. 33–34	Breach: notify the supervisory authority within 72 hours of awareness unless unlikely to result in risk; notify individuals without undue delay where high risk.
Arts. 37–39	DPO mandatory for public authorities, or where core activities involve large-scale regular and systematic monitoring, or large-scale special-category/criminal data.

9 · International transfers (Arts. 44–49)

Art. 44	General principle — the protection must travel with the data.
Art. 45	Adequacy decision (the simplest route where one exists).
Art. 46	Appropriate safeguards — Standard Contractual Clauses, Binding Corporate Rules, codes, certifications.
Art. 47	Binding Corporate Rules for intra-group transfers.
Art. 49	Derogations for specific situations — narrow, occasional, non-repetitive.
Post-Schrems II	A transfer impact assessment is expected alongside Art. 46 tools: assess destination-country law and add supplementary measures. Relevant whenever a model or API sits outside the EEA.

10 · Enforcement (Art. 83)

€10M / 2%

Art. 83(4) — controller/processor obligations: Arts. 8, 11, 25–39, 42, 43. Covers by-design, DPIAs, records, security, processors and the DPO.

€20M / 4%

Art. 83(5) — the serious tier: principles (5, 6, 7, 9), data subject rights (12–22) and transfers (44–49).

Whichever is higher, based on total worldwide annual turnover of the preceding financial year. Remember the split: breaching a principle or a right is the 4% tier; failing a process obligation like a DPIA is the 2% tier.

11 · GDPR × EU AI Act

They apply cumulatively — the AI Act does not displace the GDPR. The AI Act is product-safety law organised around risk tiers and provider/deployer duties; the GDPR is data-protection law organised around personal data. An AI Act conformity assessment does not satisfy Art. 35, and a DPIA does not satisfy the AI Act’s Art. 27 fundamental rights impact assessment. Where high-risk AI processes personal data, expect to owe both.

12 · Traps that cost marks

“Anonymised” training data	Only outside GDPR if re-identification is reasonably impossible. Pseudonymised data (Art. 4(5)) is still personal data — a very common distractor.
Basis vs. condition	Special-category data needs BOTH an Art. 6 lawful basis AND an Art. 9(2) condition. Answers offering only one are wrong.
Biometrics	Only special category when processed for the purpose of uniquely identifying a person. Face detection that never identifies anyone is not automatically Art. 9.
“Solely” automated	Rubber-stamp human review does not defeat Art. 22. The reviewer must have authority and competence to change the outcome.
DPIA vs. FRIA	DPIA is GDPR Art. 35 and protects personal data. The AI Act’s fundamental rights impact assessment (Art. 27) is a separate, additional obligation for certain deployers.
Controller vs. processor	Determined by who decides purposes and means — not by what the contract calls you.
72 hours	Runs from awareness, not from occurrence, and applies to the authority. Individuals get "without undue delay", only if high risk.

Primary sources

Article numbers and thresholds are stated as in Regulation (EU) 2016/679. Verify current procedural details against the IAPP Candidate Handbook before your exam.

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: GDPR for the AIGP1 / 5 · score 0

A team wants to process special-category data to test a model for bias. Under the GDPR, what does it need?

ANeither, because testing for bias is not processing personal data.
BBoth an Art. 6 lawful basis and an Art. 9(2) condition.
CAn Art. 6 lawful basis, since bias testing is a legitimate interest.
DAn Art. 9(2) condition, which supplies the lawful basis itself.