Investieren in Dubai nach dem Preissturz: Unterwegs mit unserem Finanzberater | Umair Zaffar
Umair Zaffar October 26, 2026 0 Comments

EU AI Act Article 4: AI Literacy, Implemented in Practice

EU AI Act Article 4 requires providers and deployers to take care of the AI literacy of the people who operate and use their AI systems. We implement that not as mandatory e-learning but as a competence register attached to the model registry: if you do not hold the module for your role, the gateway does not let you in.

Article 4 is the least conspicuous article of the EU AI Act and the first one that applies. It covers every AI system, not only high-risk ones, and addresses providers and deployers. An insurer rolling out an LLM assistant to 500 employees is a deployer from the first login, and has been under this obligation since February 2025. Article 4 of the AI Regulation does not appear in the catalogue of fines in Art. 99. The leverage sits elsewhere: with sector supervisors, who in regulated firms already ask about staff qualifications, and with the damage done when an employee trusts a model output they cannot interpret.

The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Article 4 belongs to the first phase and has applied since 2 February 2025. The Omnibus moved deadlines; it removed no obligation. It did adjust the wording of Article 4, more on that in a moment. The overview is in EU AI Act in practice.

What Article 4 requires

  • Art. 4(1): Providers and deployers of AI systems must take measures to support the AI literacy of their staff and of all other persons who operate or use AI systems on their behalf, which includes a service provider’s staff. Three things have to be taken into account: the technical knowledge, experience, education and training of those persons; the context in which the systems are used; and the persons or groups of persons on whom the systems are used. The original wording required them to ensure, “to their best extent, a sufficient level of AI literacy”. The wording after the amending Regulation speaks of measures to support literacy and clarifies that no specific level has to be guaranteed for any individual. In both versions it is a duty to take measures, with the burden of proof, not an exam per head.
  • Art. 4(2): The Commission and the Member States support these measures, in particular for SMEs. The Commission publishes practical examples on the single information platform under Art. 62(3)(b).
  • Art. 4(3): The AI Board issues recommendations on promoting AI literacy, taking European competence frameworks into account, where appropriate with common objectives for the Commission and the Member States.
  • The definition: Art. 3(56): skills, knowledge and understanding that allow providers, deployers and affected persons to make an informed deployment of AI systems and to be aware of the opportunities, risks and possible harm. Recital 20 adds the interpretation of outputs and the knowledge of how AI-assisted decisions affect the people concerned.
  • The connection: For high-risk systems, Art. 26(2) (oversight by persons with competence, training and authority) and Art. 14(4) build on this foundation.

The technical example

An insurer rolls out an LLM assistant to 500 employees, from claims handling to actuarial. The assistant answers questions about policy wordings and work instructions through RAG, summarises case files, drafts customer letters and extracts information from documents. It is not a high-risk system under Annex III as long as it stays out of risk assessment and pricing in life and health insurance (Annex III, point 5(c)). Every user has to know that line, because it does not run through the tool but through how they use it.

The data flow: every request carries a use-case ID and the user’s identity and passes through the policy gateway. The gateway resolves the registry record, checks the data zone, replaces names, policy numbers and addresses with IDs, retrieves documents through RAG with the user’s access rights at query time, with provenance per passage, and writes one entry per answer to the evidence log. Three data zones: green is policy wordings and internal instructions, gray is pseudonymised case data, red is health data from benefit claims and everything that never belongs in a prompt.

Where Article 4 bites shows up in four situations. A claims handler pastes the assistant’s summary of a claim into the file unchecked and misses that the assistant quoted a deadline from an outdated work instruction. A colleague copies a doctor’s letter into the chat; pseudonymisation catches the name, not the diagnosis. A department head, acting as business owner, approves a use case without knowing what the evaluation set she is signing is. After a model change the assistant abstains more often, and users read that as an error. None of these is a model problem. Each is a competence problem, and each ends up with a customer.

How we implement it

We at SIFAMO treat AI literacy like any other requirement on the governance layer: as something the system enforces and that you can export. The principle is the same as for the model registry. There it says: no registration, no route. Here it says: no proof, no access.

Four role tiers instead of one course. Each tier has a competence profile written as can-do statements, not a slide deck. Users, meaning all 500: what the assistant can and cannot do, the three data zones, checking sources before adopting an output, reading an abstention correctly, knowing that manipulated documents are an attack path, the reporting route. Business owners: reading and owning their registry record, interpreting the monitoring, pulling the kill switch. Reviewers: interpreting outputs, automation bias, reason categories, the seeded error cases we do not announce to them. Engineers: gateway, evidence log, evaluation set, regression and adversarial tests, rollback. The user module runs on real cases, the reviewer module ends with a practical test on cases with a known answer.

A competence register linked to the model registry. Per person it holds: roles, completed modules with version and date, result, validity, for external staff as for employees. Per model record in the registry it states which competence is required for which action: use, approve, review, operate. Hard rules follow from that. A registry record without an owner holding valid business-owner competence cannot be created. The three-way sign-off by business, risk and IT accepts only signatures from people with a matching entry. Only someone with reviewer competence for that system class gets into the reviewer pool.

Access through the gateway, not through the learning platform. On every call the gateway looks up the user’s competence status the same way it looks up scope: missing or expired module, no answer, but a notice with a link to the module; for reviewers without valid competence, no cases in the queue. The proof is not a document but a property of the system. Whoever is not trained cannot act. That is the sentence a supervisor wants to hear, and it is technically true.

A refresher on every relevant change. When the model version, prompt version or scope changes in the registry, that creates a delta module. Its content comes from the evaluation diff: what changed in behaviour, which new weaknesses the tests showed, which tools were added. For users it is short and has a deadline, for reviewers and business owners it is blocking. Known weaknesses from the evaluation thus become training content, not just a line in the test report.

The affected persons in the curriculum. For customer-facing use cases the user module covers what the output means for the customer: that a letter drafted by the assistant counts as a letter from the insurer, what rights affected persons have under the GDPR and Article 50, where a complaint goes. Almost every programme leaves this part of Art. 4(1) out.

What the supervisor sees. MaGo and VAIT require insurers to have staff who are qualified for their tasks, and BaFin checks in regulated firms whether the people who operate, approve and review AI systems are competent to do so, and wants to see it demonstrated. We answer with four exports: the competence register per role, the join with the model registry, the join with the evidence log (every reviewer action provably comes from a person who held valid competence at that moment) and the gateway statistics on rejected access attempts. On top of that the effectiveness metrics before and after each module: hit rate on seeded error cases, red-zone attempts, disagreement rate of the blind re-check.

Where the law leaves room, this is what we chose. We define “sufficient” per role as a competence profile, not per person as an exam grade, in line with the clarified wording. We read “measures” as something that changes permissions, because otherwise they cannot be proven. And we train the external staff too, because paragraph 1 explicitly covers persons who work with the system on the organisation’s behalf. I am an engineer, not a lawyer. But a join across three registers answers questions that a folder of certificates does not.

Typical mistakes

  • One e-learning for everyone. Forty-five minutes, the same content for the intern and the reviewer. Article 4 requires prior knowledge, context and affected persons to be taken into account. A uniform module takes none of them into account.
  • Competence without permission. Training completed, access granted independently of it. Two lists that know nothing of each other, and the proof is an assertion.
  • Once and never again. The model changes twice a year, the training is still the one from the rollout.
  • The forgotten externals. The service provider’s case handlers use the assistant every day and are in no register, because the HR platform only knows employees.
  • Only the users, not the decision makers. Five hundred people trained, but the business owners sign registry records they cannot read.

EU AI Act Article 4 checklist

  • Four roles, each with a competence profile of testable can-do statements.
  • A competence register holds role, module version, date, result and validity per person, for external staff as well as employees.
  • Every record in the model registry names the required competence per action; registration without a qualified owner is technically impossible.
  • The gateway checks competence status on every call; there is no access without valid proof.
  • A change of model, prompt or scope creates a delta module with a deadline, with content taken from the evaluation diff.
  • The proof is exportable: competence register, model registry and evidence log can be joined per decision.
  • Effectiveness is measured, not only attendance: seeded error cases, red-zone attempts, disagreement rate of the re-check.

Open questions

  • “Sufficient” and “measures” are not defined, and there is no harmonised standard for them. The Commission’s practical examples under paragraph 2 and the Board’s recommendations under paragraph 3 only help if they get more concrete than “offer training”.
  • A completion rate measures attendance, not competence. We measure behaviour, but there is no recognised metric for AI literacy, and whether our measures capture the right thing is an open research question.
  • Competence data and usage data per person are employee data. How granular a proof per head a supervisor may see, and what works councils and data protection allow, has to be settled in every organisation. We export metrics in aggregate; individual entries only carry access decisions.

Frequently asked questions

Does this apply to us as a deployer if we only buy the assistant?

Yes. Art. 4(1) addresses deployers explicitly, regardless of the system’s risk class. The provider takes care of the literacy of its own staff, not of yours. It owes you information about the system; you owe your employees the ability to handle it.

Is a mandatory e-learning with a certificate enough?

It is a measure, so formally a start. But Article 4 requires prior knowledge, context of use and affected persons to be taken into account, and a uniform module does not do that. The proof only becomes robust when it is linked to permissions and behaviour measurably changes. More on this under AI Governance.

What happens when the model changes?

The registry notices first: the version change creates a delta module from the evaluation diff. Users get a deadline; reviewers and business owners only continue working once they have completed the module. The other articles in the series: all articles at a glance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment