Supporting is not understanding

Regulation (EU) 2026/1744, the Digital Omnibus on the AI Act, has been in force since 27 July. Everyone looked at the postponed high-risk deadlines: 2 December 2027 for Annex III, 2 August 2028 for Annex I. I stopped on a smaller sentence.

Article 4 required providers and deployers to ensure, “to their best extent”, a sufficient level of AI literacy among their staff. The new text says they shall “take measures to support the development of AI literacy”. No longer a level to reach: measures to adopt.

From result to ritual

I understand the reasons. A duty of result on something as vague as “sufficient literacy” was hard to verify and easy to contest. But the chosen fix moves the law onto ground I know well: rituals.

Consumer and privacy law is full of rituals. The published notice. The ticked box. The completed course. We know from the data what they produce. Fewer than one user in a thousand opens the licence before accepting it (Bakos et al.). 98% of participants in one study missed a clause asking for their first-born child (Obar and Oeldorf-Hirsch). The ritual is performed. Understanding is not.

The Habeas Mentem Charter has an article I wrote with exactly this in mind: “No ritual produces understanding merely by having been performed.” The new Article 4 is not wrong. It is exposed. Everything now depends on how “support” is read.

The lazy reading and the serious one

We will see the lazy reading in the coming months: a forty-minute e-learning module, a certificate, a line in the training log. Measures taken, duty discharged.

The serious reading starts with a question: support towards what? If the aim is for users of a system to know when to trust it and when not to, measures are judged by what they produce. And that can be observed:

• do users recognise a wrong output when they see one?
• can they say what the system does not do?
• do they know whom to report a problem to, and do they?

You do not need to measure people. In my work the rule is precise: we measure documents, environments and conditions, never the worth or ability of readers. One can assess whether a training path is built to be understood, whether instructions for use clear a readability threshold, whether the interface makes the system’s limits visible.

Why it matters more than it seems

Literacy underpins almost everything else in the AI Act. Human oversight under Article 14 presumes people able to understand. Transparency under Article 50, applicable since 2 August, presumes recipients able to interpret a label. If the foundation shrinks to a process to document, everything on top of it becomes a process to document too.

The EU legislator chose not to define the result. That does not stop anyone from defining it for themselves. Whoever does so first will have a strong argument before an authority, a client or a court. Not “we ran the course”, but “here is what the course changed”.