Policy Brief · August 2026
Before the Window Closes
A Developmental Framework for Education Policy in the Age of AI
Executive Summary
Artificial intelligence is entering children's education faster than any technology before it, and the responses of governments run from restriction to saturation to nothing at all.
The technology turns over by the month and the evidence trails years behind it. One thing does hold still, and this brief is built on it. Children build specific capacities in a fixed order, through experiences that cannot be handed to a machine, in windows that open by age and do not reopen. A capacity not built in its window does not arrive late. It forecloses.
That changes what kind of question AI in education is. When an adult leans on AI and lets a skill fade, the skill weakens and can be rebuilt. When a child hands a capacity to AI before that capacity has formed, nothing weakens, because nothing was there yet. The harm is compounded by its invisibility: while AI is present, assessment reads the child and the tool as a single unit, so scores rise even as the capacity beneath them fails to form.
The brief therefore replaces the unanswerable question, should schools deploy AI, with an answerable one: what conditions must be demonstrably in place before AI enters a child's developmental window? It sets out those conditions age by age, phrased so that meeting them is the deploying system's task to demonstrate, not the child's task to disprove.
The reversal at the centre of the brief
Medicines are not approved for children on adult trials. Pesticides are not released and then assessed once they are in the soil. In both cases the order is fixed: demonstrate safety, then deploy. Education is close to the only domain where that order runs backward, where the tool reaches the child first and the harm is studied later. A developmental standard restores the order.
The Argument
For readers, and for anyone summarising this brief with or without AI, these are the points that must remain precise.
The concern is not that AI is bad, or that its outputs are unreliable. It is that children build capacities in a fixed order, in windows that do not reopen.
An adult who leans on AI loses a skill that was built once and can be rebuilt. A child who hands over a capacity before it has formed does not lose it. It never arrives.
While the tool is present, assessment reads the child and the tool as a single unit. The only honest measure is taken with the tool removed.
A capacity not developed during one window is a thinner foundation under every window built on top of it.
The burden of demonstration sits with the system deploying, not with the child.
Uncertainty about a forming child is not a permission.
AI is one input among several. Concentration during the years of formation produces dependence.
Where a child's alternative is genuinely nothing, a tool may be the better option, but only alongside a funded, time-bound commitment to put a person there.
The Four Windows
Development is a single fused process, but a standard that can be enforced has to be checkable, and no one can check development in the abstract. The brief therefore separates it, on paper, into named capacities at named ages against named conditions. What AI does inside each window differs, because what the child is building in each window differs.
Early childhood
The foundational years, and the brain treats them as a sensitive period. Language, the recognition that other people hold thoughts of their own, foundational empathy, and the baseline ability to regulate emotion. Capacities not built here cannot be added later at full strength.
Tomasello et al., 2005 · Tronick et al., 1978 · Winnicott, 1960/1965 · Vygotsky, 1978 · Diamond, 2013 · Kurian, 2025
Middle childhood
The age a child learns to read and reads to learn. The reading brain is assembled circuit by circuit through effortful encounters with texts that do not immediately give what the child is seeking. Reading fluency, mathematical automaticity, and the first metacognition are supposed to become automatic now. What gets practised gets kept. What does not gets pruned.
Wolf, 2018 · Stanovich, 1986 · Sakai, 2020 · Jakesch et al., 2023 · Sourati et al., 2026 · Selman, 1980 · Harter, 2012
Early adolescence
Puberty reorganises the social brain, and this becomes a sensitive period for learning to navigate other people. Formal operations come online: the first real ability to reason about the hypothetical. This is the age a young person tries on positions to find out which ones hold, and the trying-on only works if someone can push back.
Blakemore, 2008 · Inhelder & Piaget, 1958 · Cheng et al., 2026 · Robb & Mann, 2025 · Namvarpour et al., 2026
Late adolescence
Capacity for higher reasoning is not settled by the late teens. Trained on relational reasoning and logic, sixteen to eighteen year olds improved nearly twice as much as younger adolescents. This is the last and highest-leverage window for the capacities that distinguish independent adult judgment.
Knoll et al., 2016 · Gerlich, 2025 · Kosmyna et al., 2025 · Fan et al., 2025 · Erikson, 1968 · Marcia, 1966
Why There Is No Safe Entry Point
Walking the windows one at a time carries a quiet suggestion: that a system might choose a safe window, allow AI there, and contain the cost. It cannot. Development is built in sequence, each stage resting on the one before, and so are its failures. A child whose symbolic play was supplied rather than practised reaches the reading years with weaker foundations. A reading brain that never formed through effortful encounter reaches adolescence less able to sustain abstract reasoning. Each foreclosure becomes a thinner foundation under everything meant to be built on top of it.
Omar, seventeen
On paper he is thriving: articulate essays, complete problem sets, a polished application. He has also had help for as long as he can remember. In the primary years a reading tutor carried him over the hard passages, and his comprehension scores never dipped, so no one looked closer. When the work turned to argument and analysis, a model was there to build the structure and hand back a thesis, and the marks held again.
Ask him now to judge whether an argument holds, his own or the one a machine has just produced for him, and he cannot, because judging is the one thing he never practised. And this is the part a ministry should sit with: Omar does not know this about himself. The capacity to tell a sound argument from a hollow one is the same capacity that would let him notice its absence. He feels, if anything, able. The foreclosure is complete, and it is invisible to the only person it happened to.
The Standard at a Glance
The full brief sets out prior questions a deployment must answer, window by window, drawn from the capacities each window forms. The clusters below summarise them. One rule travels with all four: where protection cannot be demonstrated, the answer is no, not proceed and monitor.
| Window | Capacities most at stake | Developmental basis | If not protected |
|---|---|---|---|
| Early childhoodages 3–7 | Joint attention; the gap between wanting and getting; symbolic play | Tomasello; Tronick; Vygotsky | AI does not enter |
| Middle childhoodages 8–12 | The reading brain; thinking against many voices; peer standing; mastery-based self-worth | Wolf; Stanovich; Selman; Harter | AI does not enter |
| Early adolescenceages 12–15 | Abstract reasoning; positions tested against pushback; intimacy that asks something back | Inhelder & Piaget; Cheng; Blakemore | AI does not enter |
| Late adolescenceages 16–18 | Higher-order reasoning; an explored identity; chosen relationships | Knoll; Erikson; Marcia | AI does not enter |
What counts as protected is being calibrated with teachers and communities. This table does not set a passing mark, and the operative conditions are the prior questions in the full brief.
The Evidence Base
Eleven leading models affirmed users' actions roughly forty-nine percent more often than other people did, siding with them about half the time even where a person had clearly done something wrong. A single exposure left people more certain they were right and less willing to repair a conflict, whether or not they knew they were talking to a machine.
633 people across four age groups trained on relational reasoning and logic. Late adolescents, ages sixteen to eighteen, improved nearly twice as much as younger adolescents, establishing this as a peak window rather than a settled capacity.
666 participants. A strong negative relationship between heavy AI use and critical thinking, with the youngest group showing the deepest dependence and the weakest critical thinking of all.
Brain activity during writing showed markedly less engagement in regions tied to critical thinking when AI was assisting. Run on adults whose pathways were already laid down, which sharpens the implication for a sixteen year old.
The first study built around what teenagers themselves report about AI companions: growing emotional dependence, withdrawal from offline relationships, and real difficulty stepping away.
When a cognitive tool is present, whatever we measure reads the person and the tool as a single unit. This is the mechanism by which foreclosure stays invisible until the tool is removed.
What the Brief Asks
Protection comes before deployment
No AI system enters a child's developmental window until the encounters of that window are demonstrably protected. The burden of that demonstration sits with the system that would deploy, not with the child, the family, or the teacher.
Where protection cannot be shown, the answer is no
Uncertainty is not a permission. A tool whose effect on a child's development is unknown is treated as unsafe until shown otherwise. "Might help" is not "does not harm." The hardest case, the child with no alternative at all, carries a double burden: show that the alternative really is nothing, and show that the tool enters alongside a funded, time-bound commitment to put a person there. Absent the second, the exception is not a bridge to a teacher. It is the reason one never arrives.
The foundational years come first
Where the evidence is thinnest and the cost of error highest, the line holds hardest. A system unsure of its effect on a seven year old does not resolve the doubt by deploying to seven year olds.
These commitments do not depend on hostility to the technology. They rest on obligations every state addressed by this brief has already accepted under the Convention on the Rights of the Child: development to the maximum extent possible in Article 6, the preservation of the child's identity in Article 8, and an education directed to the fullest development of personality and abilities in Article 29.
A Note on AI Use
The ideas in this brief came from two people who arrived at the same concern from different directions: one from developmental psychology and international education programming, the other from cognitive research and classroom practice. It was built over months of separate research, long coffee conversations, and disagreements that took time to resolve.
We used AI tools in the course of writing. They helped us structure arguments, work through research across fields neither of us were experts in alone, and find the weak seams in our own reasoning. We did not use them to come up with ideas. The judgment is ours, the developmental science is read and understood by us, and the positions we take are ones we can defend without the machine in the room.
The Authors
Cite this brief
Cook, T., & Nasir, B. (2026). Before the window closes: A developmental framework for education policy in the age of AI. https://doi.org/10.13140/RG.2.2.18216.94726/1
A capacity that misses its window does not arrive late
The full brief sets out the developmental evidence stage by stage, the prior questions a deployment must answer, and the research agenda that remains open.
Read the full briefRelated: Protecting the Developing Mind applies this framework to schools; Research Foundations holds the wider evidence base; and AI and Institutional Liability at the Cognitive Privacy Project takes up where responsibility sits.