AI Literacy Is a Construction Problem

There is a document travelling through more and more inboxes these weeks, and it is a certificate of attendance. It confirms that an employee took part in a mandatory training on artificial intelligence, that she now knows about the risks of generative systems, about hallucinations, about data protection, about copyright. The morning after, that same employee pastes a full client contract into a freely accessible chat window because she wants it summarised faster, and both things are true at the same moment, the certificate and the pasted contract.

The certificate is not a lie. She really did understand what she was told, she could tick the box in an exam saying that contract data should not go into public models, and she would be right. It is only that on Tuesday morning she is not acting in an exam but in an environment where the public chat window is one click away, the approved internal tool is three approvals and a VPN away, and the deadline for the summary runs out in forty minutes.

For about twenty years the answer to every competence problem in organizations has been the same, and it is training. Compliance, diversity, information security, agile working, and now artificial intelligence: the moment a capability is missing, a course is booked, a completion rate is defined, a certificate is issued. Since the second of February 2025 this reflex even has a legal basis, because Article 4 of the EU AI Act requires every company that uses AI to ensure a sufficient level of AI literacy among the people who work with it, and it names training explicitly as one way to meet that duty. Obligation spotted, course booked, rate achieved, task done.

The pressure behind it is real, and it is well documented. The OECD reports in its Employment Outlook 2026 that vacancies in the AI-exposed industries have risen more sharply than elsewhere, that there are no signs of widespread labour displacement, and that the real bottleneck is a different one: companies cannot find enough people who can actually handle the tools. At the American association NACE, more than one in three entry-level roles now demands AI skills, nearly triple the figure from autumn 2025. The demand is there, the regulation is there, the course is booked.

This is the point to pause and name the thing differently from how it is usually named. What goes wrong in the certificate example is not a knowledge deficit. The employee knows what would be right. What is missing does not sit in her head, it sits in the environment in which her head works on Tuesday morning.

What a course changes and what it does not

A course changes knowledge, and that is no small thing, but it is a bounded one. Knowledge sits in the declarative, retrievable part of memory, the part that shows up in an exam. Action happens elsewhere, in the moment itself, under time pressure, with whatever tools are within reach, along the paths the environment holds open or shut. Between what someone knows and what they do lies neither a motivation problem nor a memory problem, but an environment.

The obvious objection is that one should simply train better, more often, more practically, with worked examples and repetition. That, too, only moves the line, it does not remove it. You can train a person as thoroughly as you like in using the approved internal tool, and as long as that tool sits three approvals and a VPN away while the public one sits a single click away, under time pressure the nearest option gets chosen, not the trained one. This is not a matter of character and not a matter of discipline, it is a matter of path lengths.

The layer nobody designed

Every organization designs its environments with some care, only not this one. There is an architecture for money, with controlling and a chief financial officer, there is one for information, with an IT function and a security officer, there is one for law. For the environment in which people actually act, in which knowledge becomes behavior or is left lying, there is none, and so no one designs it. This undesigned layer is what I call behavioral architecture, and the case of AI literacy exposes it with unusual sharpness.

The order that gets overlooked here has a name. Not form follows function, as the architect Louis Sullivan once put it, but the other way round, Function follows Form: the function, meaning the actual behavior, follows the form of the environment in which it takes place. Whoever does not build the form leaves the behavior to the accident of shortest availability, and shortest availability in most companies currently means the public chat window.

For the people in learning and development who are planning these mandatory courses right now, this is less a criticism than an enlargement of what their work is about. Their job does not end at the knowledge they convey, it begins there. Whoever wants to produce AI literacy in an organization designs the environment in which the conveyed knowledge becomes the nearest action available, the internal tool one click away instead of three approvals, the safe path the shortest path, the right choice the most convenient one. That is not course content, that is a blueprint.

The test that takes one minute

There is a simple way to tell a construction problem from a knowledge problem, and it takes one minute. Imagine everyone in the company had passed the course with top marks and knew everything about the correct handling of AI. Does behavior then change on Tuesday morning? If the answer is yes, it really was a knowledge problem, and the course is the right answer to it. If the answer is no, because the safe path is still the longer one and the public chat window is still one click away, then it was never a knowledge problem, and no course in the world will solve it.

The completion rate is such a seductive metric precisely because it measures the thing that is easy to measure, and not the thing that matters. It answers the question whether people were present while something was explained to them. It does not answer the question whether anything changed in the environment in which those people act afterwards. An organization can reach a rate of one hundred percent and not have shortened a single path, and it will wonder why the contracts keep landing in the public window.

The uncomfortable part is not that training is useless, because it is not. The uncomfortable part is that training takes care of the visible thing and leaves the effective thing undone, and that a completion rate that has been met feels like a solved problem, while the environment that really governs the behavior runs on untouched as before.

Perhaps the most honest question an organization can ask itself before the next mandatory course is therefore not how many will attend, but a different one: if everyone knew everything on Monday, which path would they still find first on Tuesday?