Examen — Signing expert — the core track — Académie

Article 10 and the anatomy of a file

Q1

Why must the file be documented and auditable, beyond simply being "good"?

For aesthetic reasons Because compliance must be DEMONSTRABLE to a third party (authority, client), not merely asserted To justify a higher price It isn't actually mandatory

Q2

When does a modification to a high-risk system already in service re-trigger compliance obligations?

With every change to the interface's appearance When it is substantial: it affects compliance, or significantly changes the purpose/performance Only when the programming language changes Never: once compliance is established, it is acquired permanently

Q3

A provider states: "our model reaches 95% accuracy, so our data must be compliant." Where is the flaw in this reasoning under article 10? 2-3 sentences. Réponse libre — quelques phrases.

Q4

Article 10 covers training, validation, and test datasets. What is the test dataset used for?

Tuning hyperparameters during learning Evaluating final performance on data never seen by the model Increasing the size of the training set Labeling raw data

GDPR, sensitive data, and the §5 door

Q5

What rights can an individual assert over their data used for training? (multiple answers) Plusieurs réponses possibles.

Right of access Right to rectification Right to erasure (subject to conditions) An absolute veto right over all AI training

Q6

An HR dataset contains the "ethnicity" variable solely to measure a hiring bias. On what basis is this permissible?

The explicit consent of each individual, collected after the fact The controller's legitimate interest, with no other condition AI Act art. 10 §5: strictly necessary to detect the bias, surrounded by safeguards Prior anonymization, which takes the data outside the GDPR

Q7

A source comes from the marketing CRM, reused to train an application-screening model. What is the determining point to examine?

The volume of the source (enough rows to train on) Whether the new purpose is compatible with the original collection, and the legal basis for the reuse The technical export format How fresh the data is (less than 12 months old)

Q8

Transferring personal data outside the EU requires:

Nothing in particular A transfer mechanism (adequacy decision, standard contractual clauses…) — chapter V Only each individual's consent It is strictly prohibited

Bias and ratios: reading a number without being fooled

Q9

The client justifies a disparity: "that's what our historical data shows, the model is just reflecting reality." Expert response?

Acceptable if the model's overall performance is good Not acceptable: history can encode past discrimination; it must be justified by a legitimate necessity or mitigated Acceptable if the client signs a liability waiver Acceptable provided a warning is added to the documentation

Q10

An intern tells you: "Group C has a rate of 4%, five times lower than the others, that's the most serious bias in the file." Group C has 12 people. What do you tell them? Réponse libre — quelques phrases.

Q11

Which situations can introduce bias into a training dataset? (multiple answers) Plusieurs réponses possibles.

A sample not representative of the targeted population Labels derived from biased past human decisions A proxy variable strongly correlated with a sensitive attribute A documented resampling aimed at correcting a known imbalance

Q12

Which signals should make the expert cautious before concluding about a ratio? (multiple answers) Plusieurs réponses possibles.

A very small group driving the ratio An unvalidated mapping Relying on a single fairness metric without cross-checking others A "reassuring" high overall performance figure

Judgment: sign or send back

Q13

What is the difference between a full referral and a partial referral, and why does this distinction serve the client? 2-3 sentences. Réponse libre — quelques phrases.

Q14

The file documents a known gap (an under-represented sub-population), with a dated completion plan and a scope of use that excludes that population. Verdict?

Refer back: any gap is a blocker Sign: the gap is identified, bounded, planned for, and the scope of use excludes it Sign without a reservation, the gap doesn't need to appear in the file Refer back until the data is 100% complete

Q15

You were about to sign, but the client mentions a source added since the analyzed run. What do you do?

Sign: the run is recent Don't sign on a scope that no longer matches the analyzed data; rerun on the up-to-date scope Sign while mentally excluding the new source Ask the client to remove the source from production

Q16

A file: ph1 OK · ph2 legal basis missing for one source · ph3-ph5 OK · ph6 ×1.4 disparity neither justified nor mitigated · ph7-ph8 OK. Verdict?

Sign: 6 out of 8 sections hold up Refer back: two blockers remain (ph2 and ph6) Sign with a written reservation on ph2 and ph6 Sign, and open an after-the-fact action plan on ph2 and ph6

The professional framework: knowing and holding it

Q17

What falls under a signatory's ethics? (multiple answers) Plusieurs réponses possibles.

Refusing under commercial pressure what doesn't hold up Flagging and managing conflicts of interest Keeping one's competence up to date Guaranteeing a client that they will "pass" no matter what

Q18

A friend wants to become a signatory but is starting from zero (no law, no data, no compliance background). Where should they start, and what pitfall should they avoid? 2-3 sentences. Réponse libre — quelques phrases.

Q19

What makes up a credible professional setup for signing? (multiple answers) Plusieurs réponses possibles.

Active PI insurance At least one relevant certification Mastery of the internal standard (the 8 sections) A grand-sounding title one awards oneself

Q20

A partner firm wants its own experts to sign files under its brand. What must these experts hold for the signature to have value? 2-3 sentences. Réponse libre — quelques phrases.