EU AI Act Article 4: AI Literacy Requirements Explained
Article 4 of the EU AI Act on AI literacy: who it covers, what changed with the AI Omnibus, what suitable training includes and how to document your measures.
Article 4 of the EU AI Act requires companies that provide or use AI systems to take measures to support the development of AI literacy among staff and others who operate or use those systems on their behalf. Since the AI Omnibus, this is an obligation of effort: the law no longer asks you to ensure a sufficient level, and no specific level is required for any individual. The duty has applied since 2 February 2025 and is not limited to high-risk AI: a company whose employees use a general AI assistant at work is covered. The Act does not prescribe a specific course or certificate. Which measures are appropriate depends on people's skills, experience and training, their role and the context in which AI is used. In practice, that means role-based training on how the tools work, where they fail, the company's rules, and when to escalate, plus a record of what was done.
This article is general information, not legal advice. It reflects Article 4 as amended by the AI Omnibus, which entered into force on 27 July 2026. According to the European Commission, the previous requirement for companies was simplified, and the Commission and Member States now take a stronger role in promoting AI literacy; see the Commission's AI Omnibus announcement and its questions and answers on AI literacy. Guidance continues to develop, so check the current official text before relying on details.
What the law says, in plain terms
The key elements of Article 4:
| Element | Meaning |
|---|---|
| Who must act | Providers and deployers of AI systems |
| Whose AI literacy is supported | Their staff and other persons dealing with the operation and use of AI systems on their behalf, such as contractors |
| What is required | Measures to support the development of AI literacy: an obligation of effort, with no specific level mandated |
| What the measures take into account | Skills, experience, education and training of the people; the context of use |
| Public support | The Commission and Member States support and facilitate these efforts, in particular for SMEs |
The Act defines AI literacy broadly as the skills, knowledge and understanding that allow providers, deployers and affected persons to make an informed deployment of AI systems, and to gain awareness of the opportunities and risks of AI and the possible harm it can cause.
Who is covered
If your company uses AI systems in a professional context, you are a deployer. That includes:
- employees using AI assistants for drafting, research or summarising,
- AI features in software you use, such as CRM, office, HR or accounting tools,
- AI systems you have built or integrated, such as chatbots or document automation.
If you develop AI systems and place them on the market or put them into service under your name, you are a provider, and the same requirement applies to your staff involved.
The EU AI Act deployer obligations article explains the provider and deployer roles in more detail.
What appropriate measures look like
There is no single standard, so think in terms of roles. A reasonable approach distinguishes at least three groups:
| Group | Typical people | What they need to understand |
|---|---|---|
| All users | Anyone using AI tools at work | What AI can and cannot do, hallucinations, data rules, approved tools, checking outputs, where to ask |
| Power users and reviewers | People who use AI daily, write prompts, review outputs, or oversee automated processes | Prompting techniques, systematic review, recognising failure patterns, escalation, documentation |
| Decision-makers and owners | Managers, process owners, IT, compliance | Risk categories, legal obligations, vendor assessment, governance, incident handling |
People involved with high-risk AI systems, especially those assigned to human oversight, need deeper, system-specific training on that system's capabilities, limits and correct use.
Core topics for a basic AI literacy programme
1. How generative AI works, at a useful level
- AI models predict likely text; they do not look up truth.
- Output can be fluent and wrong.
- Quality depends on the instructions and the context provided.
- Models have knowledge cut-offs and do not know your company's internal information unless it is provided.
2. Limits and risks
- Hallucinations: invented facts, figures and sources. See how to reduce AI hallucinations for practical techniques.
- Bias: outputs may reflect biases in training data.
- Over-reliance: the tendency to accept AI output without enough scrutiny.
- Prompt injection: instructions hidden in documents or web pages can manipulate AI tools.
- Inconsistency: the same prompt can produce different answers.
3. Company rules
- Which tools are approved for what.
- What data may and may not be entered.
- Review requirements before outputs are used.
- Uses that are prohibited or need approval.
- How to report problems.
These rules usually live in an AI policy. Our guide to writing an AI acceptable use policy includes an outline and a one-page summary you can use in training.
4. Using AI well
- Writing clear prompts with task, context, format and rules. The prompt engineering guide is a good basis.
- Checking outputs efficiently.
- Knowing when not to use AI.
5. Regulation and accountability
- The basic risk categories of the AI Act. EU AI Act risk categories explained gives an overview.
- Why some uses, such as evaluating job applicants, need special care.
- Who in the company is responsible for AI questions.
What AI literacy looks like in a daily role
Training quizzes on Article 4 often ask which behaviour shows that an employee is AI literate. The answer follows from the Act's definition: AI literacy is about informed use and awareness of opportunities, risks and possible harm. It is not about technical skill.
| Behaviour | AI literate? | Why |
|---|---|---|
| Evaluating an AI system's output for errors and potential bias before including it in a report | Yes | The person understands that outputs can be wrong or skewed and takes responsibility for the result |
| Approving all AI outputs to increase efficiency | No | This is over-reliance; speed replaces judgement |
| Sharing personal employee or customer data with a public AI tool to get a quick summary | No | It ignores data rules and the tool's terms |
| Writing complex machine-learning algorithms | Not required | Useful for developers, but literacy does not mean being able to build AI |
| Telling a customer that a reply was drafted by an AI assistant where company rules require it | Yes | The person knows the transparency rules that apply to their role |
| Escalating when an AI tool is about to be used for a new purpose, such as ranking applicants | Yes | The person recognises that a new use can change the risk category |
A useful test for your own training: after the session, can each participant name two typical failure modes of the tools they use, the data they must not enter, and the person to ask when unsure? If so, the measure did its job, and that is what you record.
Formats that work
AI literacy does not have to mean long courses. Effective formats include:
- Short live sessions with real examples from your own work, including examples of AI getting things wrong.
- Hands-on exercises: improve a weak prompt, find the errors in an AI-generated summary, decide whether data may be entered into a tool.
- Quick reference cards with the data rules and approved tools.
- Onboarding modules for new employees.
- Refreshers when tools, rules or use cases change.
- Peer champions in each team who help colleagues and collect questions.
Training that uses your actual tools and tasks is usually more effective than generic material.
Documenting your measures
The AI Act does not prescribe a specific form of documentation for AI literacy, but because the obligation is about taking measures, a record of those measures is the most practical way to show what you did. A simple record could include:
| Field | Example |
|---|---|
| Measure | "AI basics session for all staff" |
| Target group | All employees using AI tools |
| Content | Topics covered, materials used |
| Date and duration | |
| Participants | List or count |
| Trainer or source | Internal, external provider |
| Next review |
Keep materials and attendance records together with your AI policy and your inventory of AI tools and use cases.
A practical 30-day plan
| Week | Action |
|---|---|
| 1 | Inventory AI tools and use cases; identify who uses what |
| 1 | Define target groups and what each needs |
| 2 | Finalise or update the AI use policy and data rules |
| 2–3 | Run a basic session for all users; hand out a reference card |
| 3–4 | Run deeper sessions for power users, reviewers and owners |
| 4 | Document measures, collect questions, set a review date |
Small companies can compress this considerably. The point is to start with the basics for everyone and go deeper where responsibility is greater.
Linking literacy to other obligations
AI literacy supports other parts of the AI Act and good practice generally:
- Human oversight of high-risk systems requires people with the competence and training to do it.
- Transparency duties require staff to know when disclosure is needed, for example for deepfakes.
- Avoiding prohibited practices requires managers to recognise where the lines are.
It also has business value beyond compliance: trained staff get better results from AI tools and make fewer costly mistakes, which shows up in the time savings you can measure with the AI ROI calculator.
Common mistakes
- Assuming it only applies to high-risk AI. Article 4 applies to all providers and deployers.
- Assuming the AI Omnibus removed the duty. It was softened to an obligation of effort, but providers and deployers still have to take measures.
- One generic e-learning for everyone. Needs differ by role and context.
- Only theory. Without hands-on practice with your own tools, behaviour rarely changes.
- No record. Without documentation, it is hard to show what you have done.
- One-off training. Tools and rules change; plan refreshers.
- Forgetting contractors and temporary staff who use AI on your behalf.
Summary
AI literacy under the EU AI Act means taking measures that help the people who use AI on your behalf understand it well enough for their role: what it does, where it fails, what the rules are and when to escalate. Start with an inventory, train by role, use real examples, keep a record and update regularly. To see which of your AI uses may need deeper training because of their risk level, try the EU AI Act risk checker.
FAQ
What does Article 4 of the EU AI Act require?
Since the AI Omnibus, it requires providers and deployers of AI systems to take measures to support the development of AI literacy among their staff and other persons operating AI systems on their behalf, taking into account their skills, experience, education, training and the context of use. It is an obligation of effort: no specific level of AI literacy is required for any individual.
Since when does the AI literacy requirement apply?
Article 4 has applied since 2 February 2025. The AI Omnibus, in force since 27 July 2026, changed it from ensuring a sufficient level of AI literacy to taking measures that support its development, and made the Commission and Member States responsible for supporting these efforts, in particular for SMEs.
What is an example of AI literacy at work?
An employee who checks an AI system's output for errors and possible bias before using it in a report is applying AI literacy. Approving every output to save time, or pasting personal data into a public AI tool, shows the opposite. Staff do not need to be able to write machine-learning code.
Is a certificate required for AI literacy?
The AI Act does not prescribe a specific certificate or course. Companies decide which measures are appropriate. Keeping a record of the measures taken, for whom and when, is the most practical way to show what you did.
Does AI literacy apply if we only use ChatGPT or similar tools?
Yes. Using AI systems in a professional context makes you a deployer, and the literacy requirement applies regardless of the risk level of the system.
Related articles
- AI Agents and the EU AI Act: Which Rules ApplyRegulation & governance8 min
- AI Inventory Template: How to Record Every AI SystemRegulation & governance9 min
- EU AI Act for Deployers: What Businesses Must DoRegulation & governance10 min