The Emperor's New Clothes: Gambling with Children's Minds

There is a collective pretence currently taking hold in our education system. We are watching an endless parade of AI-assisted lesson plans, automated grading platforms, and generative tools placed directly in front of primary students, and we are being told to admire the fine threads of progress. The social pressure to be "future-ready" is immense, and the fear of appearing Luddite is greater than the impulse to question that pressure.

Beneath the layers of this magical new tool, however, is a more fundamental question that remains unanswered. In our rush to celebrate this great technological leap forward and streamline for efficiency, are we certain of the long-term outcome?

We Have Done This Before

In the beginning of 2009, the One Laptop per Child initiative deployed cheap laptops across rural Peru. The evaluations were exhaustive. Cristia and colleagues (2017) randomised 318 rural primary schools over 15 months, tracking an expansion from 0.12 to 1.18 computers per student. They found no statistically significant effect on mathematics or language achievement. In urban Lima, Beuermann et al. (2015) evaluated home laptop access and found no academic benefit, noting that teachers reported a decrease in student academic effort. When Cueto et al. (2025) tracked a separate sample of 531 primary schools from 2007 through 2019, the long-run academic impact across the complete educational trajectory remained flat.

The point is not that a laptop and a language model are the same object. They are not. The point is what the education sector demanded of itself before putting each one in front of a child, and the answer to that has changed beyond recognition in fifteen years.

But the sharper problem is who has been studied. I opened on primary students. The evidence that follows comes from a secondary school in Turkey and thirty-six undergraduates.

Klarin and colleagues (2024) examined generative AI use against executive function and described their own work as a first exploration of the relationship. Their sample had mean ages of fourteen and seventeen. Campos and Koff (2026), writing in Frontiers in Developmental Psychology, put the position plainly. Studying what AI does to adult brains is interesting, but studying what it does to brains that are still being built is necessary. They scope their own concern to ages ten to twenty.

Below the age of ten, there is nothing. Reviews of AI in primary education code their outcomes as cognitive, affective, and psychomotor. They measure what the child produced. Not what the child became able to do without it.

In one of the few controlled field trials to date, Bastani et al. (2025) evaluated nearly a thousand students across three arms in a single Turkish high school. They tested an unrestricted standard generative model, a system with pedagogical safeguards designed to guide rather than answer, and an unassisted control group.

While students using the standard model performed strongly during practice sessions, they scored roughly 17 per cent lower than the unassisted control group once the tool was removed for the evaluation exam. In the safeguarded tutor arm, that deficit was essentially eliminated, though the authors observed no positive learning effect relative to controls either.

Proponents argue that AI can act as a scaffold, enabling students to engage with ideas beyond their current level of independent mastery. That possibility is real, though the difficulty is that scaffolds and crutches often look identical in short-term performance data. Distinguishing one from the other requires longitudinal evidence that largely does not exist.

The distinction is particularly important in children. An adult checking an output is testing it against a model of the world they already possess. A child is still building that model. The same tool may therefore play a fundamentally different role in each case. What serves as productive assistance for one user may become developmental substitution for another.

The paper’s core conclusion was straightforward. Design choices determine whether automated assistance undermines learning. Yet the sector continues down the path of a haphazard technological arms race.

Outsourcing cognitive labour is like outsourcing the ability to swim to a motor. You cross the pool effortlessly. The second the power cuts out, you find out what you can actually do.

Who Is in the Room?

If this is a gamble, where is the bet being placed and who is taking the money?

Look at what currently passes for evidence when a school signs an enterprise software contract. Is it an experimental trial? Has anyone involved read the methodology of an independent study, or are decisions made on a product demonstration, an overly zealous conference keynote, and the fear of losing students to a rival school across the city that just put a generative AI tool on their admissions billboard?

Who is asked to weigh in on this? Senior administrators, technology coordinators, and business managers are in the room. Are the educators who teach foundational literacy to seven-year-olds present to evaluate cognitive risk?

What is the sequence of adoption? Does a coherent institutional and ethical policy precede the contract, or does the tool arrive in the classroom months before staff assemble to figure out what safe usage looks like? Who is left to manage it? Almost always, the task of driving adoption falls to a classroom teacher who was never consulted, who did not ask for the platform, and who is handed no criteria for success beyond making sure students log in.

What does renewal look like when the year ends? Does anyone measure whether independent capability increased once the software was turned off? Or does the renewal conversation only place importance on login metrics and licence burn?

And what happens if the tool fails to deliver? Once a school has marketed an AI-powered campus to prospective parents, how do they U-turn on an expensive subscription that got them through the door?

Will We Do It Again?

We have no trouble understanding why we protect a child’s physical body. International laws forbid child labour because the damage is visible in a stunted spine, scarred hands, or a body worn down before its prime. We recognise that a child is not merely a small adult who can be pushed into a factory for the sake of economic efficiency.

The mind gets no such protection. There is no photograph of a capacity that failed to develop, and so there is nothing for anyone to be moved by.

In developmental medicine, the standard for new interventions is rigorous. Evidence of safety and efficacy is balanced carefully against potential risks before an intervention reaches a patient. In educational technology, that process has been inverted. Tools are shipped at scale, and the public is left to wonder whether independent researchers will detect consequences years later.

In its landmark global analysis Students, Computers and Learning, the OECD (2015) evaluated PISA data across 64 countries and found that intensive investment in school ICT had produced no appreciable improvement in reading, mathematics, or science. In fact, students who used computers very frequently at school scored lower in reading comprehension than those with moderate use.

Yet the capital continued to flow, and the vocabulary of "future-readiness" was adopted wholesale by educational leadership.

If the concern is unfounded and we take our time, we lose a few years of adoption. If it is founded and we rush ahead, we cannot go back and rerun childhoods. Campos and Koff put it more precisely than I can. The developmental window does not pause for the publication cycle.

The Grounded Mind

The assumption that the physical interface does not matter to the mind is similarly unproven.

When Van der Weel and Van der Meer (2024) used high-density EEG to examine handwriting versus typing, they observed widespread theta and alpha oscillatory connectivity across parietal and central regions during handwriting that was absent during keystrokes. Yet reading the literature honestly requires noting the boundaries of the finding. The study examined 36 university students copying isolated words on a screen, with no downstream measure of long-term memory retention. As Pinet and Longcamp (2025) subsequently pointed out, that neural divergence may simply reflect the higher working memory load demanded by a mechanically slower task.

Thirty-six adults, copying single words, contested in print within a year. That is the evidence base. And on foundations of roughly that weight, we are overhauling the primary modalities through which children learn to write.

The Neuro-Debt Hypothesis

Consider this. If a child offloads a small increment of cognitive synthesis every morning, such as a sentence structured, a query organised, or an ambiguity resolved by an algorithm rather than the child, the immediate effect on the day's lesson is imperceptible. It looks like a few pennies deducted from an account.

If this micro-withdrawal occurs across five subjects, five days a week, for twelve years of schooling, it represents an unmeasured cumulative shift in how habits of thought are formed.

The child cannot see what was never built. Neither can we. We do not know if the mind is losing unrecoverable ground or simply adapting.

The technological parade continues through the school gates, plastered on school websites, flaunted at education fairs, and dangled to prospective parents. Contracts are signed, software suites are provisioned, and guidelines are updated to accommodate the arrival of tools whose developmental footprint has never been audited.

The enterprise representative points to the screen. The school leadership nods in agreement. And in the classroom, the child sits quietly in front of the glass, waiting to see what we have actually bought for them.

Do we wait for the data, or continue the gamble?