The latest wake-up call comes from Deloitte: according to its annual back-to-school survey, half of the parents interviewed fear their children “rely on AI too much.” This data point captures widespread unease as ChatGPT and similar platforms slip into students’ digital backpacks, no longer confined to university labs. But the discussion risks stopping at the symptom without addressing the structural cause: who designs and controls the behavior of these intelligent assistants?

The concern is understandable. A student receiving ready-made answers without having to formulate reasoning develops a dependency that undermines deep learning. AI tutors, if configured to maximize engagement rather than cognitive autonomy, become digital crutches. The critical point is that most of these tools run in the cloud, managed by providers that optimize for fluidity and speed, not for long-term pedagogical benefit. Every interaction a minor has with a hosted LLM ends up on remote servers, often outside the school’s jurisdiction, with all the data sovereignty implications that the GDPR strictly requires to be considered.

Shifting deployment to on-premise – that is, onto local servers within the school or district – would radically change the equation. With a self-hosted infrastructure, educators could intervene on the model’s parameters: limit the number of tokens per response, set constraints that push the student to reformulate the question instead of receiving an immediate solution, or even block the output when the AI is used to bypass an exercise. This is not about advanced fine-tuning techniques, but about controlling the rules of engagement. In practice, a local LLM can be programmed to say “have you thought about it enough?” rather than “here is the answer.”

This perspective has second-order consequences for the industry. Cloud providers that currently dominate the educational market would see their competitive advantage eroded if schools began to favor on-premise solutions, perhaps certified for GDPR compliance and designed with open-source logic. Parental demand – still informal but growing – could translate into pressure on public decision-makers so that investments in AI for education include clauses for self-procurement, local installation, and independent audits. It is not science fiction: some European regions are already experimenting with contracts for dedicated AI servers in schools, precisely to avoid dependence on foreign platforms and ensure that minors’ data never leaves the national perimeter.

The flip side is cost and complexity. Managing an on-premise infrastructure requires skills that most schools lack, and purchasing hardware with sufficient VRAM to run modern LLMs weighs on budgets. But what the Deloitte survey signals is not just parental fear: it is a warning for a redefinition of who should have the final say on AI in schools. As long as the model runs in the cloud, control remains with the provider. Bringing inference closer to the school desks – literally, inside the building – could transform a dependency risk into a tool for active education.