The Digital Omnibus rewrote Article 4 of the AI Act. Where you previously had to ensure AI literacy, you now have to take measures supporting its development. We have taken that apart in detail elsewhere. This piece is about a question that sits next to it and almost never comes up in a compliance conversation: where does that literacy actually get built?
Not in the training session. Training makes it visible and evidenceable, that is its job and that is why it is necessary. But a morning in a seminar room does not create judgement that was not there before. Anyone hiring a developer or an administrator today is hiring someone whose habits with these tools formed years ago, in places no company is responsible for.
The standard the regulation itself sets
It is worth reading closely what is being asked. Article 4 names technical knowledge, experience, education and training, and the context in which the systems will be used. And Article 26(2) requires, for high-risk systems, that human oversight be assigned to natural persons who have the necessary competence, training and authority.
The word "training" sitting next to "competence" is not an accident. The legislator assumes part of it arrives with the person. A company can supply tools, set rules and refresh knowledge. Whether somebody notices when a plausible answer is wrong gets decided earlier than that.
The same distinction, in two places
In companies we keep having the same conversation. It is rarely about whether AI may be used, and almost always about where in the workflow it shows up. Used beforehand, to get to the bottom of something, weigh options or find a structure, quality rises and the person can still answer for what they hand over. Used afterwards, to fill a gap, you get output nobody in the building can defend the moment somebody asks a follow-up question.
That same distinction decides whether a sixteen-year-old learns anything from her homework. Have a topic explained and then write it yourself, and you understand it. Have the text handed over, and you submit something that does not survive being asked about in class. It is the same rule, minus the compliance vocabulary.
We have called it create, do not consume for years, and the reason the phrase holds up is that it does not depend on the tool. It survives every model update because it describes not what the machine can do, but at which point a person thinks.
What actually applies at school, and what does not
This is where it gets interesting for anyone who works with regulation, because the structure is the one from Article 4, only a notch softer.
We checked the official sources on 1 August 2026. In Germany, the standing conference of education ministers adopted a recommendation on the use of AI in school education on 10 October 2024, addressed expressly to the education administration. In Austria, the ministry states plainly that pedagogical responsibility stays clearly with the teacher. In Spain, the ministry published guidance through INTEF, with competence divided between regions and individual schools.
So in none of the three countries is there a national rule settling whether a pupil may use AI for homework. Read that from a compliance seat and the pattern is instantly familiar: a recommendation to the administration, wide discretion beneath it, and the actual decision taken where it is least documented. It is the position Article 4 occupied for companies before the Omnibus, only without a market surveillance authority at the end of it.
Why this is a business matter
Three reasons, without the pedagogical uplift.
First, hiring. The people who in five years are meant to supervise your high-risk systems are currently in a classroom where nobody has bindingly settled when a tool may be used. Whatever habits they build there, they bring with them.
Second, your own staff. The adults you train have children at home and the same question. Training that explicitly permits taking the material home lands better than training that stops at the gate. That is not a courtesy, it is the difference between a box-ticking exercise and actual interest.
Third, the evidence. When you document your Article 4 measures, the most convincing element will not be the attendance list but a recognisable position that shows up in your rules, your examples and your materials. An organisation that can explain why it uses AI at one point in a workflow and not at another has an answer. One that has merely booked a course has a receipt.
Where this is written out
The part that concerns families and schools has no business on a B2B site, so it lives elsewhere. The methodology for the school context, the same distinction worked out didactically, is at whileyoucreate.com/school. The version for home, without the jargon, is at Learn AI as a family. Both are freely accessible and require no book.
If you want to see the series behind it, it is on our AI books overview.
For the workplace side, what Article 4 requires in its current wording still stands: take measures and be able to evidence them. How that gets built is in our EU AI Act training, and what actually applies on 2 August 2026 we have written up separately.