Article 4 AI literacy duty applies
Until 26 July 2026: measures to ensure, to their best extent, a sufficient level of AI literacy. From 27 July 2026: measures to support its development, without a duty to guarantee a specific level for any individual.
AI training is often treated as a one-off awareness session: a prompt-engineering lunch, a slide deck on hallucinations, a short note saying employees should be careful. That is not enough for responsible adoption.
Article 4 of the EU AI Act has applied since 2 February 2025, in two successive versions:
- 2 February 2025 to 26 July 2026: providers and deployers had to take measures to ensure, to their best extent, a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf.
- From 27 July 2026: Regulation (EU) 2026/1744 replaced Article 4. Providers and deployers must take measures to support the development of AI literacy of those people. The obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual. The Commission and the Member States must support these efforts, in particular for SMEs, and the Commission must publish practical examples of how to comply on the single information platform referred to in Article 62(3)(b).
Both versions take account of technical knowledge, experience, education and training, the context of use and the people on whom the systems are used. When you assess what was required in the past, apply the wording in force at that time.
The duty applies to providers and deployers within the scope of the AI Act. For a Swiss organisation, check the connecting factors in Article 2, such as placing an AI system on the EU market or the use of an AI system’s output in the Union. Not every Swiss organisation is in scope.
In plain language: AI literacy is operational. It has to match the role, the tool, the workflow, and the risk.
Almatrafi and colleagues' 2024 systematic review treats AI literacy as broader than tool familiarity, including evaluation, application and ethical navigation, and reviews assessment approaches. Use it to frame learning goals. The four-level model below is Ada Studio’s proposal rather than a validated course for Article 4, so test it against the checks the actual task needs.
Where generic training may be insufficient
A finance analyst, HR manager, legal assistant, clinician, compliance officer, and board member do not need the same AI training. They need a shared foundation, but their decisions are different.
Check whether general training leaves any of these gaps:
- people learn vocabulary but not decision rules;
- high-risk roles receive the same guidance as low-risk roles;
- the organisation cannot show that training matched the actual use context.
AI literacy should be designed like a control, not like a motivational seminar.
A practical AI literacy model
Level 1: General AI literacy
Everyone using AI should understand what generative AI does and does not do: probabilistic output, hallucination, data sensitivity, limits of prompts, overconfidence, bias, and the need for review.
Level 2: Workflow literacy
People using AI in a specific workflow need to understand the permitted tool, permitted data, review standard, escalation path, and documentation requirement for that workflow.
Level 3: Decision literacy
Managers and accountable owners need to understand when an AI use case affects rights, obligations, safety, finance, employment, health, reputation, or customer trust. They need to know when the answer is not "train harder" but "change the workflow" or "do not deploy yet".
Level 4: Governance literacy
Leaders, risk teams, compliance teams, product owners, and AI champions need to understand the governance system: inventory, risk classification, vendor review, human oversight, monitoring, incident handling, and periodic review.
AI literacy programme starter checklist
0/6The training evidence problem
Article 4 does not prescribe a single training format or a specific form of documentation. That flexibility is useful, but you should still be able to show which measures you took.
Evidence can include:
- role-based training matrix;
- training materials;
- attendance records;
- tool-specific guidance;
- workflow playbooks;
- AI policy acknowledgements;
- refresh cycles;
- escalation examples;
- records of updates after incidents or tool changes.
This is where ISO/IEC 42001 becomes useful. It frames AI governance as a management system: not a one-time project, but a maintained set of policies, processes, responsibilities, evidence, and improvement loops.
AI literacy is not about making everyone a technologist. It is about making people competent enough for the AI decisions their role actually requires.
What to do next
Start with three groups: everyone, frequent users, and accountable owners. Define what each group must know, which tools they use, which data they handle, and what evidence you will keep. Test whether participants can apply the programme to their actual decisions.