EU AI Act Article 4: What AI Literacy Training Must Cover

Practitioner guidance reflecting the Act as amended by the Digital Omnibus (Regulation (EU) 2026/1744); not legal advice.

Article 4 is the EU AI Act's quietest obligation and its most misquoted one. It has applied since 2 February 2025, it touches every company that provides or deploys AI systems, and half the training vendors selling against it describe a requirement that is not in the text.

What Article 4 Actually Requires

The current wording, as adjusted by the Digital Omnibus, asks providers and deployers of AI systems to take measures to support the development of AI literacy among the staff and other persons dealing with AI systems on their behalf, taking into account their technical knowledge, experience, education and the context the systems are used in.

Read that carefully, because two popular claims are not in it. There is no mandated training program, no certification, and no defined competence level that staff "must" reach: it is an obligation of effort, not of result. And it is not satisfied by nothing: "take measures to support" still means you can show what you did. The defensible position is documented, role-appropriate measures; the indefensible positions are both "we bought everyone a course" with no connection to their actual roles, and a shrug.

Be equally careful with vendors who sell "Article 4 compliance certificates". No such certificate exists in the regulation. What exists is your own documentation of proportionate measures.

Who Needs Measures

Everyone "dealing with" AI systems on your behalf, which is broader than the AI team: the recruiter using a screening tool, the analyst pasting data into a copilot, the support agent working alongside a bot, the manager approving AI-assisted decisions, and external contractors doing any of the above for you. The proportionality cuts the other way too: the depth owed to a model developer differs from the depth owed to an occasional copilot user, and Article 4 explicitly lets you calibrate.

Training by Role, Not by Course

The structure that holds up in practice groups people by what they can break:

• All staff using AI tools: what the tools can and cannot do, confidentiality rules (what never goes into a prompt), how to recognise and report errors, and when AI output needs human verification.
• People in AI-assisted decisions (hiring, credit, assessments): the system's known limitations and failure modes, their authority and duty to override, and what gets logged.
• Builders and operators: evaluation, monitoring, drift, incident response, and the Act's obligations relevant to the systems they run.
• Leadership: risk classification, the compliance timeline, and what they are accountable for signing off.

Our AI audit checklist includes the literacy question in context, and the broader obligations sit in our EU AI Act guidance for 2026.

How to Document It

Documentation is the part a supervisor can actually inspect, so treat it as the deliverable: a short written assessment of which roles deal with AI and what they need; the measures chosen per role (training sessions, written guidance, tool-specific onboarding); attendance or acknowledgement records with dates; and a review cadence, because new tools keep arriving. A single living register with those four columns beats a binder of certificates.

A Training Plan Template

A proportionate first iteration for a mid-market company, from engagements we have run: week one, a role mapping and a two-page AI usage policy everyone acknowledges. Weeks two to three, a general session for all AI-tool users, recorded for joiners. Week four, role-specific sessions for decision-adjacent teams and builders. Quarterly, a short refresh tied to whatever tools actually changed. The whole first pass is measured in days of effort, which is also the strongest argument against ignoring it.

If you want the mapping, materials and the documentation register built with you, that is what our AI literacy training service covers, calibrated to Article 4's actual wording rather than to the version vendors wish it said.