AI literacy: what the AI Act asks and how to train your staff

What Article 4 of the AI Act asks

Since 2 February 2025 the AI literacy duty has applied to providers and deployers of AI systems. The Digital Omnibus of July 2026 changed the wording: a “sufficient” level no longer has to be ensured, but organisations must take measures to support the development of AI literacy. Deployers of high-risk AI must additionally make sure the people exercising oversight are trained for it.

According to the European Commission’s questions and answers:

• no certificate is required; an internal record of training and initiatives is enough;
• you take into account the organisation’s role, the risks of the AI, and people’s knowledge and experience;
• it also covers contractors and service providers working with AI on your behalf;
• using ChatGPT for text falls under it too: staff must know, for example, that the model can make up facts.

What belongs in AI training?

1. How a language model works, in plain language: it predicts text, it does not know what is true. See AI hallucinations.
2. What may and may not go in: the data classes from your AI policy, with examples from your own work. See Writing an AI policy.
3. Checking output: verifying sources and figures, and knowing when AI output must not leave the building without human review.
4. Security: recognising manipulation through documents and emails, and what to do after a mistake.
5. Role-specific depth: lawyers on citations and professional secrecy, HR on decisions about people, IT on agents and rights.

Keep the basics short, for example an hour, and make the depth role-specific. Repeat for a new model, a new tool or an incident.

How to record it

A supervisor or auditor will ask what you did, not whether you have a certificate. So record:

• which AI systems you use and for which roles;
• which training goes with which role, and when it was completed;
• which policy and instructions exist, and where staff find them;
• when you review the programme.

That links to an AI register. See Setting up an AI register.

Training works better in an environment that helps

Training teaches staff what to do. A well-configured environment catches what they forget anyway. Safeguards that stop an ID number or IBAN, with an explanation why, are a lesson at the moment it matters.

In werqly AI Control you see, in aggregate, which kinds of questions are asked and how often safeguards step in, without tracking individuals. That shows where more explanation is needed, and becomes your evidence that the programme works. See What is private AI?