A reflection based on the editorial by Fernández-Andújar et al. published in Anales de Pediatría on the emerging risks of generative AI for children and adolescents.
A few days ago, Anales de Pediatría published an editorial that, read from within the education sector, is hard to digest. It is signed by Marina Fernández-Andújar, Joaquín González-Cabrera, Alejandro Romero, Diego Hidalgo and María Angustias Salmerón-Ruiz, and the title leaves no room for interpretation: “Inteligencia Artificial Generativa en la Población Infantojuvenil: Riesgos Emergentes y Retos para la Salud Pública”. What the authors argue, in essence, is that what we are dealing with in generative AI is no longer a technological or strictly educational question, but a public health problem. And I think they are right.
I have been working on the application of AI in primary school classrooms at NeurekaLab for some time, and precisely for that reason, reading texts like this does not leave me indifferent. When a clinical and academic team of this standing points to risks to children’s cognition, social-emotional development and mental health, those of us who design technology for them have an obligation, at the very least, to stop and think about what we are doing.
The uncomfortable question
There is a question that public debate still avoids with a certain skill: what happens when a child stops thinking because a machine thinks for them? We are not talking about a distant future or a dystopia. We are talking about the present of a very considerable proportion of pupils who already live alongside conversational assistants capable of solving, in a few seconds, exactly the kind of tasks that had historically been the engine of their learning.
The cognitive process we activate when we face a question with no immediate answer — the doubt, the searching, the mistakes, the reformulation — is precisely what builds the mental structures that will later allow us to think autonomously. If we systematically cut this process short during the years in which these structures are being formed, what we are doing is not facilitating learning. It is replacing it.
The risk is not the technology, it is the design
I believe we must be very careful with apocalyptic discourse, because it ends up being useless. Generative AI has enormous potential and it will not disappear from the educational landscape. Demonising it is therefore as sterile as idealising it. The relevant discussion is not whether we use AI in classrooms, because that is already a fait accompli. The discussion is which AI, designed with what criteria and with what pedagogical intention behind it.
When a pupil receives a finished answer without having gone through the process of constructing it, it is not the technology that is failing: it is the design of that technology. A tool that always responds, immediately and with every appearance of competence, is optimised for the productivity of an adult user, not for the cognitive development of a child. These are two radically different objectives, and confusing them has consequences.
The Anales de Pediatría editorial frames it correctly: the risks are not confined to the academic sphere. There are implications for inhibitory control, sustained attention, frustration tolerance, and also for adolescents’ mental health and relational dynamics. It is a broader view than the usual conversation about “cheating on homework”, and it is the view we ought to take.
Learning is not solving tasks quickly
A confusion is taking hold that I find dangerous: the idea that learning and solving are the same thing. They are not. Solving is reaching a result. Learning is being transformed while reaching it. A pupil who hands in twenty impeccable essays generated by AI has not written twenty essays; they have commissioned twenty. And the difference, even if the final product looks similar, is structural.
Real learning requires making mistakes, trying again, reformulating hypotheses, enduring the discomfort of not knowing. If we offer our pupils a permanent shortcut that sidesteps all of that, what we will find in a few years will not be a better-prepared generation, but a more dependent one. And that dependence is, at bottom, a form of vulnerability.
From AI that answers to AI that trains
The approach we advocate at NeurekaLab, and which I believe should be the framework for any tool designed for children, is simple to state and demanding to implement: educational AI should not answer for the pupil, it should make the pupil think. That means not always offering the solution, adapting the difficulty to the real level of the person in front of it, introducing progressive challenges, giving useful feedback on the process and not only on the result, and, above all, designing with a developmental model of the learner in mind, not with a generic user model.
It is a deeper paradigm shift than it seems. It means accepting that a good educational AI, if it works, will often be slower, more frictional and less “satisfying” in the immediate sense than a general-purpose AI. But that friction is not a defect: it is exactly where learning happens.
A decision we cannot postpone
The editorial argues that this is a public health challenge, and I fully share that reading. We are not deciding which app our children use this school year. We are deciding, collectively and often by omission, how the coming generations will think, how they will make decisions and how they will relate to knowledge.
It seems to me that the education sector, public administrations, families and those of us who design these tools have, right now, a narrow window to get things right. It is not about holding back AI; it is about not delegating to it precisely what must not be delegated: the process by which a child becomes a person capable of thinking for themselves.
Because in education, the underlying question is never what technology can do. It is what we want our pupils to be able to do when the technology is not there.