Article 4 AI literacy: what it asks for after the Omnibus reworded it

The AI literacy duty has applied since 2 February 2025, covers every organisation using AI regardless of risk tier, and was reworded rather than deferred. What changed, and what to document.

Article 4 is the shortest obligation in the EU AI Act and the one most often missed entirely. It has applied since 2 February 2025. It is not tied to a risk classification, so it reaches organisations whose systems are all low-risk. And it was reworded by the Digital Omnibus rather than deferred, which several summaries have reported as the same thing.

What it asks for

Providers and deployers must take measures to support the development of AI literacy among their staff and other people dealing with the operation and use of AI systems on their behalf. What that takes is judged against those people's technical knowledge, experience, education and training, against the context the systems are used in, and against the people the systems are used on.

The Omnibus made three changes, and the first is the one most summaries miss. The duty itself was rewritten: it used to require measures to ensure, to their best extent, a sufficient level of AI literacy, and it now requires measures to support the development of it. The amended text adds that the obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual. And the Commission and member states are now positioned to support and facilitate those efforts, with SMEs named. The duty to take measures remains exactly where it was.

The rewording lowers the evidential bar. It does not remove the obligation. 'We took no measures' is still non-compliance; 'we trained people and one of them still misused a tool' is a materially better position than it was before the amendment.

What the measures look like by role

The Act prescribes no curriculum. It ties the measures to the roles involved and the context of use, which means the answer differs sharply between an engineer wiring a model into a product and a support agent using a summarisation button.

What to document

• Who received what training, when, and at what depth
• The material itself, versioned, so you can show what was taught at a given date
• The policy telling staff which tools are approved and what may be put into them
• Role-specific content for anyone carrying oversight duties under Article 26(2)
• A refresh schedule, and evidence that it actually ran

Article 4 sets no retention period for these records. Keep them for as long as the systems they relate to are in use, and hold them with the rest of your accountability evidence rather than in a training platform nobody exports from.

The cheapest obligation in the Act to close

Most of what Article 4 asks for is half a day of writing and a slot in an existing induction. Set against classification, logging design and oversight arrangements, it is the one obligation you can finish this month. It is also the one an authority can test in a single question, because the answer is either a training record or nothing.

The organisations that will struggle with it are the ones that cannot say which tools their staff are using, which is the same gap that makes an AI system inventory hard to start.

This post describes the regulation as amended and is not legal advice. Article 4 sets no threshold and no curriculum, so what counts as enough turns on your own roles and context, and the Omnibus reworded the duty in July 2026 in terms that summaries written before then still get wrong. Check the consolidated text or take advice before treating a training record as sufficient evidence.