I am in favour of AI in education.
In favour, under conditions.

The next problem

Once we say “yes”, we must decide who frames the machine’s voice.

A child hears a fluent, courteous answer and has every reason to believe there is knowledge on the other side. The voice sounds certain. But that certainty belongs to the form of the answer, not to its truth.

This question emerged from a small family moment. I asked a system about Nikos Kazantzakis in front of my eight-year-old daughter. Its answer was smooth and deeply misleading. That moment did not make me fear AI. It reminded me what education requires: presence, context and a person who helps the child stand before the machine’s voice.

A child needs to learn to ask who is speaking, where a claim comes from, what is missing and when to stop. A school needs to ensure that an output does not appear with borrowed authority.

The EU AI Act moves this discussion from good intentions to institutional responsibility.

The European regulation follows a risk-based approach. Certain uses in education, including the assessment of learning outcomes or decisions that affect access and educational pathways, may fall into the high-risk category. In these cases, human oversight, documentation, traceability and the ability to intervene acquire concrete weight. At the same time, the obligation to ensure AI literacy concerns those who operate and use such systems within organisations.

Having a human somewhere in the process is not enough. That person must know what happened, have time to judge and possess the authority to change the outcome before it affects the child.

Human approval solves part of the problem. But it opens another. If the entire workflow belongs to one provider, how real is the school’s control?

The problem that remains hidden

What if AI is not visible at all?

A learner may never open a chatbot and still have an educational experience shaped by AI.

A system may recommend the material they will see, draft the feedback they will receive, classify an assignment or influence a decision about support and access. AI participates in the background. Its consequences remain real.

Here we need to examine two things together. How recognisable is AI’s participation, and how much institutional force does its output retain? The most troubling case arises when the participation is difficult to see but the outcome materially affects the child.

In that case, disclosure after the event comes too late. We need to know the chain before an output becomes a decision.

We now have purpose, human judgment, independence from the provider and visibility of influence. We can finally see how all of these become one concrete pedagogical workflow.

The problem I am focusing on now

The next difficulty appears when we have to assess learning.

This is where my research is moving now. The question that interests me is not whether a machine touched the final text. I am interested in what the human learned, what they checked and what they can defend.

AI detection does not prove understanding, nor does it prove misconduct on its own. A final text can be flawless and contain almost no trace of a learning process. This is why AI-RAW shifts the weight from the product to the evidence.

An assignment connects the learning objective with specific evidence. Sources, choices, prompts, outputs, mistakes, corrections, reasoning and oral transfer of knowledge. The use of AI is declared within the workflow of the assignment, not in a general note at the end.

AI participation is made clear before the assignment begins.

The journey finally returns to where it began. To the real classroom, the teacher and the children who must do something meaningful with what the machine produces.