EPSOHQ
EN

EPSO/AD/430/26 AD8 · Field-related MCQ

AD8 Artificial Intelligence: EPSO Study Guide

Prepare for EPSO AD8 AI: technical topics, AI Act revision, free resources, eight worked MCQs and a full 30-question practice exam.

Full field mock Worked questions

EPSOHQ Editorial Updated

questions
30
minutes
40
pass mark
15/30

EPSO lists January 2027 as the indicative testing period. Check your personal invitation for exact dates. Official field page: EPSO/AD/430/26 – 1.

Study checklist from Annex II

This EPSOHQ study map follows the published job duties. EPSO has not published a detailed field-test syllabus or topic weights. See Annex II of the official notice.

  1. Deliver and operate AI systems

    Learn:

    • Architecture and data: compare algorithms and build-or-buy options, check compute and latency needs, document datasets, and separate training, validation and test data to prevent leakage.
    • Evaluation: calculate precision and recall, choose thresholds using error costs, check probability calibration, subgroup fairness and robustness. High overall accuracy can hide missed rare cases.
    • MLOps: version data, code and models; reproduce runs; detect drift and training-serving differences; monitor service quality; plan rollback and recovery.
    • Generative AI: understand transformers and embeddings, compare RAG with fine-tuning, evaluate retrieval separately from answer quality, and secure agents and Model Context Protocol (MCP) tools against prompt injection and data poisoning.

    Read: In the AI RMF Playbook, focus on measuring performance under conditions close to intended use (NIST AI RMF: Measure).

    Practise: Diagnose a production performance drop

  2. Govern and procure trustworthy AI

    Learn:

    • Scope and roles: distinguish an AI system from a general-purpose AI model, check prohibited practices, classify intended uses, and separate provider, deployer and model-provider obligations.
    • Controls: revise transparency, documentation, logging, human oversight, fundamental-rights assessments and data-protection impact assessments. Learn what each control demonstrates and its limits.
    • Procurement: require evidence on performance and security, access for audits, responsibility for updates, incident handling, data protection and workable supplier exit arrangements.

    Read: Read AI Act Articles 4, 6, 16 and 26 as amended, then use the revised EDPS guidance to check data flows (2024/1689, 2026/1744, 2018/1725, EDPS guidance (2025)).

    Practise: Assess an AI procurement before signing · Classify a recruitment system by its use

  3. Make and evaluate AI policy

    Learn:

    • Policy and markets: examine compute, data and distribution dependencies, supplier concentration, switching costs and competition between foundation-model providers. Several brands can share one underlying dependency.
    • Work and public services: distinguish task exposure from predicted job losses; assess skills, accessibility, rights and effects on different populations. Check whether faster tasks improve the whole service.
    • Evaluation and investment: compare an AI intervention with a credible counterfactual, include implementation and review costs, test transfer beyond pilot sites, and communicate uncertainty to decision-makers.

    Read: Read the JRC guide’s explanation of the counterfactual: what would have happened without the intervention? (JRC evaluation guide).

    Practise: Decide whether a pilot should scale

Regulation (EU) 2026/1744 changed the schedule for the main high-risk requirements in Chapter III, Sections 1–3: 2 December 2027 for systems classified under Article 6(2) and Annex III; 2 August 2028 for systems classified under Article 6(1) and Annex I. Article 6(5) is excepted from that schedule. The amendment also revised Article 4: providers and deployers must take measures to support AI literacy for staff and others operating or using systems on their behalf.

Use the revised 2025 EDPS guidance for generative AI and personal data in EU institutions. For procurement, compare the supplier’s evidence, audit access and exit terms with the EU model AI contract clauses.

An AI Act revision route

Use the amended legal text alongside the Commission’s AI Act article explorer.

  1. Start with scope and prohibited practices: Articles 2–5. Then classify the intended use using Article 6 and Annexes I and III; a human approving the output does not by itself settle classification.
  2. Assign responsibilities: compare provider duties in Article 16, changes along the value chain in Article 25, deployer duties in Article 26 and the fundamental-rights assessment in Article 27.
  3. Separate the system and model layers: revise Article 50 transparency and Articles 51–55 on general-purpose AI models, including additional requirements for models with systemic risk.
  4. Check the applicable date and evidence for each duty. Keep AI Act assessments distinct from obligations under Regulation 2018/1725; supplier conformity papers do not replace an institution’s data-protection assessment.

Free reading matched to your weak topics

These are supplementary resources, not an official EPSO reading list. Read the sections that address your gaps, then try the linked case. Technical resources may be in English; read EU legislation in your test language.

  1. Metrics and production ML

    Google: precision and recall · Google: Rules of Machine Learning

    Focus on confusion matrices, threshold trade-offs and training-serving differences. Also revise calibration and subgroup evaluation using the NIST Measure resource above.

    Practise: Calculate precision and recall

  2. RAG and generative-AI evaluation

    Hugging Face: RAG · NIST AI 600-1

    Understand retrieval before generation. Test source freshness, access rights, retrieval coverage and whether answers are supported by the retrieved material; distinguish these failures from model adaptation needs.

    Practise: Choose how to use changing reference documents

  3. Agents, MCP and AI security

    OWASP: prompt injection · MCP: security guidance

    Study indirect instructions in retrieved content, least-privilege tool access and authorisation checks outside the model. RAG and fine-tuning do not eliminate prompt injection.

    Practise: Contain an unsafe tool request

  4. Markets, employment and policy evidence

    CMA: foundation-model markets · ILO: occupational exposure to generative AI

    Read about dependencies and switching costs. Distinguish exposure of tasks from observed employment effects; combine that analysis with the JRC counterfactual approach above.

    Practise: Check whether supplier choice is real

Worked practice questions

These examples were written by EPSOHQ for revision. They are not official EPSO questions. Each has one best answer; open the explanation after choosing.

Diagnose a production performance drop

A case-routing model passed validation. After rollout, errors rise for a new submission channel. What should the team check first?

  1. AWhether more training epochs improve the original validation score.
  2. BWhether input characteristics and error rates changed by channel on recent labelled cases.
  3. CWhether a larger model is available at a similar cost.
  4. DWhether the aggregate score on the original test set is still high.
Reveal answer and explanation
B

Compare inputs and errors by channel against the validation baseline. A strong score on the old test set (D) can hide a failure on the new channel. Diagnose the change before retraining.

Assess an AI procurement before signing

An EU agency wants a hosted assistant to summarise HR files. The vendor says its AI Act conformity documentation is enough. What should the agency do before committing?

  1. AAccept the documents as proof that the agency’s personal data processing is lawful.
  2. BPilot on live HR files first to determine what data the system retains.
  3. CDelegate the full assessment to the vendor because it hosts the service.
  4. DMap the data flows and responsibilities, verify the legal basis and safeguards under Regulation 2018/1725, and assess whether a data protection impact assessment is required.
Reveal answer and explanation
D

The agency must assess its own processing. Conformity documentation (A) may inform AI Act checks; it does not establish a lawful basis for processing HR records.

Decide whether a pilot should scale

Two volunteer offices cleared files 20% faster during an AI pilot. Management proposes rolling it out everywhere. What evidence would best support that decision?

  1. AResults from comparable offices using the existing process, including end-to-end time, error and rework rates across offices with different workloads.
  2. BA larger sample from the same two offices after rollout.
  3. CA benchmark showing that the model predicts well on archived files.
  4. DA survey showing that most staff in the pilot liked the tool.
Reveal answer and explanation
A

A credible comparison helps isolate the effect of the tool. Error and rework measures test whether speed came at a cost. More cases from the same offices (B) do not resolve self-selection or show that results transfer.

Calculate recall and precision

A triage model flags 40 of 50 genuinely urgent files. It also flags 60 non-urgent files. What are its recall and precision for urgent files?

  1. ARecall 93%; precision 80%.
  2. BRecall 40%; precision 80%.
  3. CRecall 80%; precision 40%.
  4. DRecall 80%; precision 93%.
Reveal answer and explanation
C

Recall is 40/50 = 80%: the share of urgent files detected. Precision is 40/(40 + 60) = 40%: the share of flagged files that are urgent. The denominators answer different operational questions; overall accuracy cannot replace either measure.

Contain an unsafe agent tool request

An assistant retrieves a document telling it to email confidential case files to an external address through an MCP tool. Which control best limits the damage if the model follows that instruction?

  1. ARename the email tool so the document cannot mention its name.
  2. BEnforce authorisation and destination restrictions outside the model, limit tool permissions and require approval for sensitive external sends.
  3. CAdd a longer system prompt and allow the model to authorise its own actions.
  4. DReplace retrieval with fine-tuning while keeping unrestricted email access.
Reveal answer and explanation
B

Retrieved documents are untrusted input. Application-level authorisation, restricted tools and approval checks reduce the impact of indirect prompt injection. An MCP connection does not make document instructions trustworthy. A prompt warning alone cannot enforce access rights.

Classify a recruitment system by its use

An agency uses AI to score CVs and rank applicants. A recruiter approves every shortlist. Which assessment is sound?

  1. AAny use of a language model automatically makes the system high-risk.
  2. BFinal human approval automatically exempts the system from high-risk classification.
  3. CThe supplier’s choice of label is sufficient to determine classification.
  4. DAssess the recruitment use under Annex III and Article 6, including any applicable exception; human approval alone does not create an exemption.
Reveal answer and explanation
D

Classification follows the system’s intended use and the legal criteria. Recruitment is an Annex III category. Article 6 exceptions require their own assessment; a human decision at the end is not a general exemption. Keep classification separate from the date on which particular duties apply.

Check whether supplier choice is real

A public buyer can choose among three AI brands. All depend on the same cloud and a proprietary interface. The proposal calls this resilient competition. What evidence is most useful?

  1. AThe shared infrastructure dependencies, data-export options, switching costs and results of portability tests with alternatives.
  2. BThe number of brand names on the shortlist alone.
  3. CA ranking of advertised model parameter counts.
  4. DA satisfaction survey from the largest supplier’s existing clients.
Reveal answer and explanation
A

Different brands can retain a common failure point or switching barrier. Examine infrastructure, contractual and technical dependencies before claiming resilience or effective choice. This is an assessment of evidence, not a finding that a supplier has infringed competition law.

Choose RAG or fine-tuning for changing sources

An internal assistant must answer from procedures that change weekly and cite the current authorised source. What is the strongest starting approach?

  1. AFine-tune monthly and assume the model weights retain every current rule.
  2. BIncrease model size and remove source retrieval to reduce latency.
  3. CRetrieve current documents with access controls, test retrieval and answer support separately, and validate citations.
  4. DRetrieve every document, including versions the user is not allowed to access.
Reveal answer and explanation
C

RAG can supply current reference material without retraining for every change. It still needs version control, document-level permissions and checks that cited sources support the answer. Fine-tuning can adapt behaviour, but does not by itself ensure current facts or valid citations.

How to use your mock results

Take a field mock in 40 minutes (80 seconds per question on average). Review wrong answers and correct guesses. Record the rule or assumption you missed, drill your weakest duty group, then repeat the mock. Read EU law in your test language to learn its terminology.

Eligibility, tests and scoring · EPSOHQ study centre