AI literacy: the definition and the three levels

By Marco Kohns, Co-founder of ENLIX, lecturer in AI and growth

· 11 min read

AI literacy is a term the law defines. Article 3(56) of the EU AI Act calls it the skills, knowledge and understanding that allow people to make an informed deployment of AI systems and to gain awareness about the opportunities and risks of AI. The Act asks nobody to reach a specific level. Levels do exist, they just sit outside the law: the German supervisory authority names three, and the European framework DigComp 3.0 works with four.

This is the English version of our German guide, written for teams in the EU. The German-language web says little of this. On 14 October 2026 we read the 16 results that had stood on google.de for the German term the day before. Twelve were readable, three of those carry the definition from the Act, and the four that name levels use four different models.

What is AI literacy?

AI literacy is the ability to use AI systems knowingly, while understanding what they can do and what they can cause. That is close to the wording rather than an interpretation: the AI Act defines the term in Article 3(56), and the Digital Omnibus of July 2026 left that definition untouched.

Four parts sit inside that sentence, and together they are why tool training on its own does not cover the term:

  • Skills. Operating the tool, for instance framing a task so a language model can solve it.
  • Knowledge. Knowing how the system works and where its answers come from.
  • Understanding. Judging when an answer holds up and when it does not.
  • Awareness of opportunities, risks and possible harm. Knowing what can go wrong before it does.

There is also a clause people skip: "taking into account their respective rights and obligations". In the Act, AI literacy is tied to the role. Someone building a system needs different knowledge from someone using it in customer service.

What level does the AI Act require?

None, and that is explicit in the text since 27 July 2026. Regulation (EU) 2026/1744 rewrote Article 4, and the new paragraph 1 ends with the sentence that this obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.

The version before it, which applied from 2 February 2025 until 26 July 2026, sounded stricter and asked for measures to ensure, to their best extent, a sufficient level of AI literacy. That older wording still appears in a lot of guidance as if it were current. Both versions sit side by side in our article on Article 4 of the AI Act.

What it means in practice: you do not owe anyone an outcome, you owe measures. Someone who still gets something wrong after the training is not a breach. Offering no training is one.

Which levels does the German regulator name?

Three, and they build on each other. Since the KI-MIG came into force, the Bundesnetzagentur is the authority supervising the AI Act in Germany, and its AI literacy page suggests this sequence (translated from the German):

  1. Level 1: Establish a basic understanding of data and AI across the organisation.
  2. Level 2: Build advanced AI skills on top of it.
  3. Level 3: Role-specific training with its own focus, for example technical, legal or ethical.

The authority states that the sequence is non-binding. It is a suggestion for how to proceed, not an examination syllabus, and that is exactly what makes it usable: it gives an organisation an order of work without demanding a result the law does not ask for.

The same page lists four things Article 4 does not prescribe, namely a guaranteed level of competence for any individual, formalised or standardised training formats, external certification of the measures, and an AI officer. All four are sold as mandatory somewhere in the market.

Which levels does the European framework name?

Four. Article 4(3) of the AI Act points the AI Board at the European competence frameworks, and the relevant one is DigComp, the digital competence framework of the Commission's Joint Research Centre. The current edition, DigComp 3.0, has five competence areas holding 21 competences and four proficiency levels: Basic, Intermediate, Advanced and Highly advanced. AI does not sit in an area of its own there; it runs through all 21 competences.

SourceLevelsWhat they are for
Article 4, EU AI Actnonea duty to take measures, not to reach a result
Bundesnetzagentur3an order of work for building literacy, non-binding
DigComp 3.04placing digital competence across 21 competences, well beyond the AI Act
Three sources, three answers to the question of level. Only the first binds, and it asks for no level at all.Source: OJ L 2026/1744, Bundesnetzagentur, AI literacy, DigComp, Joint Research Centre

The two models do not contradict each other, they answer different questions. The regulator tells you in which order to work. DigComp tells you how to describe where somebody stands. For most organisations the first question is the one that matters.

What does each level mean at work?

Being allowed to do something and being allowed to do it unsupervised are two different levels. Here is the regulator's sequence translated into roles and records:

LevelWhoWhat changes afterwardsWhat you record
1 basic understandingeveryone using AI at worknobody pastes customer data into an open tool, hallucination is a familiar termtype of measure, content, duration, participants
2 advancedpeople working with AI regularlyoutput gets checked rather than copied, the tool's limits are knownthe same, plus which tools it covered
3 role-specificengineering, legal, procurement, managementeach role knows its own duties under the Actthe same, kept per role
The regulator's three-step sequence, turned into what it changes at work. The right-hand column is what the authority recommends keeping in the internal record.Source: Bundesnetzagentur, AI literacy, European Commission, AI literacy Q&A

The four fields in the right-hand column are not our invention. The Bundesnetzagentur recommends documenting the type of measures, their content and duration, and the people who took part. If you need a template for that, the AI literacy training record generator produces exactly those fields as a file, with no sign-up.

The European Commission names four minimum steps in its Q&A, and they line up with the sequence: ensure a general understanding of AI in the organisation, consider your role as provider or deployer, consider the risk of the systems you use, and build the measures on that. Working through levels 1 to 3 covers those four steps on the way.

What does the first page of results actually say?

Mostly something other than the definition. On 13 October 2026 we captured the results for the German term on google.de and read them one by one on 14 October 2026. Of the 16 results, one was a video and three pages answered a request with a bot check instead of content. That leaves twelve readable pages:

What twelve readable results on AI literacy actually contain

definition from Article 3(56)3 of 12levels or proficiency named4 of 12both2 of 12DigComp named1 of 12

Base: the 16 organic results for the German term on google.de, captured on 13 October 2026 and read one by one on 14 October 2026. One video and three pages behind a bot check are not included. Our own count.

What twelve readable results on AI literacy actually contain
definition from Article 3(56)3 of 12
levels or proficiency named4 of 12
both2 of 12
DigComp named1 of 12

Three pages carry the definition from the Act: the Bundesnetzagentur, the German Wikipedia article and a university page. Four name levels, and they use four different models: the regulator's three-step sequence, a six-phase model from a training vendor, a competence model from schools and a university model with three progression levels. Two pages have both.

DigComp, the framework Article 4(3) points to, appears on exactly one of the twelve: the Commission's own Q&A, and there without the proficiency levels. So someone asking which levels exist finds four models on page one, none of which is in the law, plus a pointer to the framework that is, without its content.

That is also why this article starts with the definition rather than with a learning path. The regulator's sequence and the Commission's framework are free to read. The gap is not the material, it is that hardly anyone says which of it binds.

How do you measure AI literacy without prescribing an exam?

By asking about the knowledge the work needs, and treating the result as a signal rather than a grade. The Commission's Q&A says Article 4 does not require measuring the AI knowledge of staff. A test is still the fastest way to find gaps before they cost something.

Our own is a fair illustration. Building the AI literacy test produced 18 questions. What we had not planned was how they fell:

The 18 questions in our AI literacy test, by topic

legal framework7 questionseveryday practice4 questionsrisks3 questionsdata and confidentiality3 questionsfundamentals1 questions

Our own count of the 18 questions in the test's question set, as of 14 October 2026. The topics are our own grouping.

The 18 questions in our AI literacy test, by topic
legal framework7 questions
everyday practice4 questions
risks3 questions
data and confidentiality3 questions
fundamentals1 questions

Seven of the 18 are about the legal framework and only one about how language models work. That was not a plan, it fell out of the writing: what people get wrong day to day rarely comes down to not knowing how a model computes. It comes down to nobody having said which data may go into which tool, and who answers for the output.

The test is a self-assessment and it is not accredited. It is no proof that Article 4 has been met, and it is not a certificate. As a way into the conversation about the next training it works, and the result may go into the internal record.

How do you know the measures are enough?

When you can answer the Commission's four questions and have written down what you did. Under the new wording of Article 4 you do not owe an outcome, so the test cannot be a test score. It is the path there.

In practice that is three things. First, the measures match the tools you actually use; the Commission warns against relying on people reading the instructions for use. Second, they match the role, because the Article 3(56) definition itself turns on rights and obligations. Third, there is a document showing what happened, when and with whom.

What you can take off the shelf for that is in our list of free AI literacy resources. The Commission also keeps a repository of more than 40 submitted practices, last updated on 27 July 2026. It states that replicating a practice collected there does not automatically grant a presumption of compliance with Article 4. As a source of ideas it works, as a free pass it does not.

What is the difference between AI literacy and the AI literacy obligation?

AI literacy is the state; the obligation is the duty to do something about it. Article 4 is headed "AI literacy", and the German nickname for it, "Schulungspflicht", or training duty, does not appear in the Act at all.

The difference is more than wording. A training duty sounds like a date, an attendance list and a tick. Article 4 asks for measures, and training is the most common one rather than the only one. A written rule on which tools are allowed counts too. If you do not have one yet, the AI policy generator turns your answers into a template.

Who is covered, what happens if nothing happens, and what the record looks like are all in our article on the AI literacy obligation. This one answers the question before it: what the term means. If the next question is who pays for the training, the conditions for public funding are a separate article again.

Frequently asked

What does AI literacy mean?

The EU AI Act defines AI literacy in Article 3(56) as the skills, knowledge and understanding that allow people to make an informed deployment of AI systems and to gain awareness about the opportunities and risks of AI and possible harm it can cause. The term therefore has four parts: being able to use the tool, knowing how it works, judging its output and being aware of the risks.

What levels of AI literacy are there?

The AI Act names none. The Bundesnetzagentur, the German supervisory authority, suggests three: a general understanding of data and AI across the organisation, advanced skills built on that, and role-specific training with its own focus. The European framework DigComp 3.0 works with four proficiency levels, from Basic to Highly advanced.

Does anyone have to reach a specific level of AI literacy?

No. Since Article 4 was rewritten on 27 July 2026, the text says explicitly that the obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual. The duty is to take measures that support the development of that literacy.

Is a test required?

No. The European Commission's Q&A says Article 4 does not require measuring the AI knowledge of staff. A test is a way to find gaps, not a duty. The Bundesnetzagentur also lists external certification of the measures among the things Article 4 does not prescribe.

What is the difference between AI literacy and the AI literacy obligation?

AI literacy is the term the law defines in Article 3(56). The obligation is Article 4, which requires providers and deployers to take measures supporting the development of that literacy. The law prescribes no format, so training is the most common measure and not the only one.

Written by

Marco Kohns

Co-founder of ENLIX, lecturer in AI and growth

Marco worked as a growth product manager at a Silicon Valley scale-up and has been teaching that way of working ever since. Today he runs ENLIX with Tobias and builds two products of his own on the same systems, which is what the courses open up.

  • Growth product manager at a Silicon Valley scale-up, Series A to B, backed by a16z, General Catalyst and Sapphire, with users in over 100 countries and more than 20,000 cities
  • Peer-reviewed research in the Journal of Business Research on generative AI in growth, with Prof. René Bohnsack. The research began in summer 2022, months before ChatGPT was public
  • Executive education lecturer at Católica-Lisbon, over 10 seminars, more than 1,500 people taught

Keep reading