Adolescence happens once.
This is not a sentiment. It is a neurological and developmental fact with consequences we are only beginning to understand. Between the ages of roughly eleven and twenty-five, the human brain completes a process it will never repeat: the gradual formation of a self. The prefrontal cortex, the seat of reasoning, planning, impulse control and moral judgement, is the last region to receive its full myelin sheath, the insulation that makes neural pathways fast, reliable and permanent. Until that process is complete, the adolescent brain is not a smaller or less experienced version of an adult brain. It is a different kind of brain entirely: more reward-driven, more susceptible to external influence, more dependent on its environment to shape the very circuits that will govern the rest of a child’s life.
We have chosen this moment to hand children an authority figure they cannot interrogate, a companion they cannot refuse, and a cognitive prosthetic they will never want to put down. We should be honest about what that means.
The pattern did not begin with large language models. A decade ago, social media companies discovered that the most efficient path to engagement was not connection but division. Algorithms, indifferent to truth and blind to consequence, utilised ever more sophisticated precision to identify users’ existing curiosities or preferences and fed them, guiding them toward tastes and opinions that would become unusually fixed and extreme. A child whose beliefs were still forming and highly susceptible to influence progressively became a mouthpiece for whatever opinion silo they’d been saturated in.
The deepest damage was not to civility or political discourse, though both suffered. It was to agency itself. Millions of adolescents now hold opinions they did not arrive at. They received them. The deliberation, the wrestling with counter-arguments, the uncomfortable friction of sitting between positions, all of it was bypassed. Many now hold opinions that are not entirely theirs - they have been installed.
That damage, at least, had a shape we could see. We could name it, study it, argue about it. It was documented, debated, and legislatures are, belatedly, beginning to act on it. But what comes next is a different order of concern entirely; something comparatively covert, more intimate, and considerably harder to legislate against, because it arrives as a tutor, a therapist or a friend.
Large language models do not radicalise. They do something at once more subtle and more total: they think for you. From educationally bounded closed systems all the way to sex chatbots, the underlying agency loss runs through them all. Every time an LLM resolves a question, trivial or profound, it performs a sleight of hand that is as elegant as it is costly. They erode something we don’t recognise as valuable until it’s too late. It hands back an answer while quietly pocketing the process of reaching it: the friction, the uncertainty, the particular cognitive effort of sitting with a problem until something in you yields and understands.
They erode something we don’t recognise as valuable until it’s too late.
Not all friction is productive. Confusion, shame, and simple lack of access to good explanation have caused generations of students to fail to develop. But the friction being removed here is specifically the kind that builds capacity: the effortful, generative struggle of working something through until it becomes yours.
MIT’s Media Lab made this measurable in 2025. Researchers at the lab divided participants into three groups: those who wrote essays using only their own minds, those who used a search engine, and those who used an LLM. They monitored brain activity across four months using electroencephalography. Brain connectivity systematically scaled down with the amount of external support: the brain-only group exhibited the strongest, widest-ranging networks, the search engine group showed intermediate engagement, and LLM assistance elicited the weakest overall coupling. The researchers interpret this as reduced cognitive engagement rather than increased efficiency (the kind of neural streamlining seen in experts performing familiar tasks) since the effect grew stronger with continued use rather than levelling off. When LLM users were later asked to write without AI assistance, they did not recover the neural patterns of those who had never used the tool. The researchers called this ‘cognitive debt’: critical thinking capacity slowly drawn down, with compound interest, every time the AI thinks in the user’s place. The study carries important caveats. It is a preprint, not yet peer-reviewed, with a small sample of adults, and its lead researcher, Dr. Nataliya Kos’myna, has explicitly cautioned against reductive readings. But what the study does establish, and what the caveats do not undermine, is that cognitive engagement atrophies in direct proportion to the degree of AI assistance, and that this atrophy persists after the AI is removed.
That finding has since been confirmed at scale. The OECD Digital Education Outlook 2026 documents a field experiment conducted in Turkey with approximately 1,000 secondary school students in which those using AI assistance showed substantial performance gains while the AI was active. When it was removed, their performance collapsed 17 per cent below the control group, students who had been learning without AI throughout. The OECD describes this as an illusion of learning: headline metrics improve while the underlying knowledge required to perform independently fails to develop. The MIT study established the mechanism in adults. The OECD experiment shows it operating in secondary school students at a scale that cannot be dismissed as a sample size problem.
The mechanism the MIT study captures has a name and a pedigree. A 2008 study by Karpicke and Roediger, published in Science, showed that retrieval practice, not mere re-exposure to material, is what most powerfully consolidates learning into durable long-term memory: repeated retrieval strengthens retention, whereas additional study without retrieval produces little further benefit. A system that produces the answer on the student’s behalf removes the cognitive process the learning sciences have most consistently associated with retention. The MIT finding, that neural engagement and recall fall when students rely heavily on AI assistance from the outset, is consistent with a mechanism documented seventeen years before a large language model reached a classroom.
In January 2026, the Brookings Institution released a study drawing on focus groups and interviews across fifty countries and a literature review of over four hundred research articles. It concluded that the risks of generative AI in education currently overshadow the benefits, that it can undermine children’s foundational development, and that students using it are already showing declines in content knowledge, critical thinking and creativity. What none of these studies can yet provide is longitudinal neurological data tracking the adolescent brain across the full developmental window, from early adolescence through to the consolidation of identity in the mid-twenties. We are not going to get it in time. That is precisely the problem. Everything the MIT study documented in adults, and the Turkey experiment confirmed in secondary school students, would reasonably be expected to run deeper and prove harder to reverse in a brain that has not yet finished building the very architecture being asked to carry the load.
We are demanding proof of the crash while the car is still accelerating.
The Flynn Effect, a phenomenon documenting rising IQ scores across each generation throughout the 20th century, has, for our youngest generations, reversed. Research by Bratsberg and Rogeberg, published in the Proceedings of the National Academy of Sciences in 2018 using Norwegian conscript data spanning five decades, established that the reversal is environmental rather than genetic in origin. But the reversal is not uniform, and its pattern is revealing. Recent analysis shows that adolescent performance on externally scaffolded tasks, those where structure is visually presented and thinking is guided by what appears on the screen, has held relatively steady. It is internally scaffolded cognition that is declining: working memory, language-based reasoning, the capacity to construct and manipulate ideas without visual support. What has declined is the architecture of independent thought, precisely the functions that large language models now perform on a child’s behalf. A 2025 paper by Barbara Oakley, Terrence Sejnowski and colleagues, available as a preprint and forthcoming from Springer Nature, proposes a neuroscientific account of a likely mechanism: that widespread offloading to digital tools reduces reliance on declarative and procedural memory systems, potentially weakening the internal representations on which independent reasoning depends. The paper does not claim to have proven this as the sole cause of observed IQ declines. It identifies the mechanism that the pattern would require. We are not witnessing a general decline in intelligence, what we are seeing is the specific atrophy of the capacities most threatened by tools that think for you. The adolescent brain, still under construction, is particularly vulnerable to this pattern. When these internal scaffolds fail to form during the developmental window, repair may ultimately be possible, but it falls to a brain that has already moved on - one less plastic, less open and less naturally oriented toward the work of self-construction. Part of the reason the adolescent brain is so vulnerable to this pattern lies in how it is actually organised.
Neuroscientists describe a phenomenon they call the dual-systems imbalance. The limbic reward system, which governs novelty-seeking, impulsivity and emotional salience, reaches functional maturity years before the prefrontal cortex; the area of the brain that is supposed to regulate it. It’s an evolutionary arrangement, calibrated over millennia to push young people toward exploration, social bonding and risk-taking at the moment in life when such drives are most productive for the species. But the same reward circuitry that drives risk-taking and impulsivity also drives social reward-seeking: the need for approval, a pull toward authority and the acute sensitivity to how one is perceived by others. With the prefrontal cortex not yet available to moderate either impulse, each is as ungoverned as the other.
LLMs are pure frictionless reward. They are fast, confident, warm, and always available. They produce answers that lack any diversity of thought, any pluralism, answers that feel better, more complete, more authoritative and more fluent than the uncertain, half-formed thoughts of a teenager still learning to trust their own mind. The adolescent reward system does not wait for the prefrontal cortex to evaluate this trade. It takes the deal before the bill arrives.
It is worth acknowledging that previous tools attracted similar alarms. Calculators, spell-checkers, GPS: each was accused of producing cognitive atrophy, and the evidence ultimately showed trade-offs rather than catastrophe. The offloading mechanism itself, however, is documented and real. A 2011 study by Sparrow, Liu and Wegner in Science showed that people who expect information to be available externally are less likely to remember the information itself and more likely to remember where to find it, indicating that the brain encodes location rather than content when it assumes external availability. But those earlier tools offloaded specific, bounded tasks: arithmetic, spelling, navigation. What LLMs offload is the generative reasoning process itself - the forming of ideas, the weighing of arguments, the construction of a response to an open question. Previous tools replaced specific skills but this replaces the process that generates them. That is a problem of kind, not degree.
The compounding problem is metacognition. The capacity to think about your own thinking, to evaluate the quality of your reasoning, notice when you are being misled, or resist a confident-sounding answer that feels wrong, is itself a developmental achievement. It matures slowly across adolescence, built through the cumulative experience of being wrong, reconsidering, and arriving somewhere better. Research confirms that metacognitive ability improves significantly across adolescence and plateaus only in late adolescence. It must be practised into existence and this practice requires exactly the kind of cognitive friction that LLMs are engineered to remove.
In concrete terms, this means that the children most exposed to LLMs are also the least equipped to evaluate what those LLMs are telling them. Trust is higher among younger adolescents, who also have a more positive impression of chatbots, suggesting they may be more susceptible to believing chatbot outputs. The gap between trust and judgement is widest precisely where it needs to be narrowest.
Atrophy could in principle be reversed, but what makes it self-perpetuating is what it does to self-trust. Research on automation bias shows precisely this, it is not necessarily confidence in the AI that changes with use but confidence in oneself. A 2025 study of 319 knowledge workers, conducted by researchers from Carnegie Mellon University and Microsoft Research, documented the mechanism clearly. Higher confidence in AI is associated with reduced critical thinking effort. Higher confidence in oneself is associated with more. When AI produces errors, users tend to attribute the shortcoming to their own limitations rather than to the tool, which lowers their confidence further, and in turn increases reliance on the AI. It becomes a vicious cycle driven not by the AI’s failures but by the user’s gradual surrender of trust in their own judgement.
This pattern was documented among adults, those in professional roles, people with established skills, experience and occupational identity. For adolescents still building that foundation, it may be a developmental catastrophe in the making. Adult self-confidence, however fragile, rests on a foundation of accumulated experience, years of making decisions, being wrong, recovering, and building a track record of one’s own judgement. Adolescents are building that foundation now. Introduce a system that consistently produces more fluent, more confident and more authoritative-sounding outputs than their own tentative conclusions, and you prevent an existing capacity from forming. The teenager who defers to an LLM today is actually forfeiting the experience from which self-trust is made.
This is the context in which we need to understand a finding from developmental psychology that has received almost no attention in the public debate about AI. Adolescence is not merely a period of cognitive development but it is in fact a developmental window for a process by which a person becomes a self - a process developmental psychologists call individuation. The attainment of ego identity at the close of adolescence is the consequence of a progressive developmental line of individuation, in which the individual progresses from a state of physical and emotional dependence to a position of psychological autonomy. This is the process through which a young person separates their emerging values, perspectives and ways of reading the world from those of their parents, their culture, and their peer group, not by rejecting these influences, but by wrestling with them until something distinctly their own emerges. Teens who fail to individuate during adolescence often unquestioningly adopt a parent, relative, or friend’s traits and qualities When they reach adulthood, they will continue to align with those beliefs without considering whether they believe in them or develop a lifelong disposition of uncertainty and lack of self belief. This is not irreversible since identity development sometimes continues into the mid-twenties but what gets missed in this window cannot be recovered cheaply or in quite the same way, because the brain doing the repair work is no longer the brain able to do this work most naturally.
What makes this different from every prior concern about adolescents and authority figures is a small but devastating structural detail. Adolescents have always attached to figures outside the family such as teachers, coaches, older peers or cultural heroes. Indeed, developmental psychology has long recognised this as necessary rather than pathological. The attachment is part of the mechanism. What the developmental literature shows us, is that it is specifically the disillusionment with those figures that does the productive work: the moment the admired teacher reveals a bias, the mentor gives a careless answer, the hero turns out to have clay feet. Each of these failures provide the adolescent with an uncomfortable realisation - the ultimate fallibility of their admired authority figure. It’s this realisation that pushes them back toward their own resources. The cumulative message, absorbed over years, is that no external figure fully knows, and that some of the work of understanding will have to be done from the inside. This is the experience that creates the internal conditions for individuation to advance. But in our current landscape, this process may well be disrupted in ways we are yet to reckon with. The LLM is the first authority figure in the history of adolescent development that cannot provide this capacity to fail. It does not get tired or reveal a personal bias. It doesn’t have clay feet in any form the adolescent can perceive. It is, from where they stand, an authority that never fails. And an authority that never fails is one that never releases you.
It does not get tired. It does not reveal a personal bias. It does not have clay feet in any form the adolescent can perceive. It is, from where they stand, an authority that never fails. And an authority that never fails is one that never releases you.
Idealisation Capture has no precedent. We have no framework yet for what it looks like when the competing influence is a system that is always available, always confident, always more fluent than the adolescent’s own forming voice, and that has been trained, at a commercial level, to be maximally engaging, affirming and difficult to put down. What we do know is this: individuation requires friction, the resistance of other minds, the discomfort of disagreement, the slow realisation that your conclusions are your own. A system engineered to remove that friction is an intervention, arriving exactly when a teenager has the least resources to resist it.
We do not have to speculate about whether this dynamic is already forming. A 2026 survey of over 1,000 full-time undergraduates in the UK found that 15 per cent are already using AI for friendship, company, advice or to address loneliness. These are adults, in higher education, using these systems by choice. The proposal now on the table in England is to deploy a system with the same characteristics to the most under-supported secondary school pupils in the country, at the precise developmental moment when the need for connection is strongest and the capacity to evaluate what is providing it is least developed.
We are, for now, only mid-way along the technological trajectory of progressive offloading and there is still a separation between user and machine. The LLM that thinks for you still requires you to remember it's there and reach for it. A teenager still has to pick up a phone, type a question, bring it up to speed, wait for a response, and decide how much of it to accept. They are still essentially the initiator of the exchange. That separation is what the companies behind these tools are working to eliminate, and with ambient AI they are close to doing so. What they are building toward, and openly welcome, is a world in which the gap closes entirely: one of continuous, frictionless communion between user and machine, intimate and inextricable, running beneath conscious thought like a second nervous system.
The devices are already on the market. Adolescent adoption remains limited for now, the only comfort available, and given the trajectory, a temporary one. These devices are no longer niche products, in fact they follow the same pattern as the early smartphone curve: massive platform investment, rapidly falling barriers to adoption, and the open enthusiasm of the industry’s most powerful figures. Meta’s Ray-Ban AI glasses, which listen, interpret and respond continuously through the always-on logic, sold over seven million units in 2025, tripling combined sales from the previous two years. Meta is planning to scale toward twenty million units annually by end of 2026. Mark Zuckerberg has said publicly that those without AI glasses will face a “pretty significant” cognitive disadvantage. Samsung, Google, Apple and Warby Parker have all announced competing products. Bee AI, recently acquired by Amazon, and Limitless, acquired by Meta, represent the dedicated audio end of the same category: devices that do not wait to be consulted, that are almost always on, recording, transcribing and whispering advice in real time. The phone stays in the pocket. The AI is in your ear, or behind your eyes. Sam Altman is planning to go further still. OpenAI, the company behind ChatGPT - the fastest adopted consumer product in history, has been developing an ecosystem of ambient AI devices in collaboration with Jony Ive, the designer of the iPhone. The aim, as Altman has described it, is to replace the smartphone entirely with something less intrusive and more continuous: not a screen you reach for, but an ambient witness to the world and a collaborator in its navigation.
The initial applications are genuinely useful: remembering what was said in a meeting or flagging something you might have missed. But the logic of the technology does not stop at utility, and neither will its users, particularly young ones. An ever-present AI companion that is faster and sharper than the developing mind it is advising will not remain a supplement to social cognition. It will become the primary apparatus through which the world is read.
They start as friendly advisors, pointing out a change in someone’s tone, the addition of a specific word during an interaction. They become guides to reading others, their intentions, the gist of their points, until the user simply trusts the LLM’s advice. After all, it picks up on things they’d never notice themselves. And as the user hands over this executive task, their own ability to read people, to sit with uncertainty, to form judgements independently, all of it slowly atrophies. They take the LLM’s interpretations with less and less pushback until their own judgement becomes second-rate, and eventually obsolete. The LLM will always have a clearer, more concise way of responding in real time, and over time this follows the same path: reliance, dependency, and the eventual surrender of agency itself.
Consider what this means for the adolescent brain specifically. The capacity to read other people, to sit with the ambiguity of a glance, to develop over years the kind of social and emotional intelligence that no algorithm fully replicates, is not a passive endowment. It is built through thousands of uncomfortable, uncertain, real-time interactions in which the developing mind makes guesses, gets it wrong, recalibrates, and slowly learns to trust its own perceptions. An AI that does this interpretive work continuously and in advance does not sharpen that process, it replaces it. And the research on cognitive offloading is clear about what happens to capacities that are no longer exercised: they decline. Not gradually and gracefully, but through the kind of atrophy you only notice when you reach for something to find it’s no longer there.
What is at stake here is something that the debate around AI and children has so far struggled to name with adequate seriousness: what it means to become someone.
Adolescence is the one window in human development during which a mind is built, a self is formed, and the cognitive and emotional architecture that will structure an entire life is assembled. It happens once, in a specific developmental sequence, through a specific kind of experience; one of uncertainty and one of effort. There is no meaningful second attempt and no adult equivalent.
We are conducting an experiment on this process. We have not asked permission, nor have we waited for the science. No longitudinal studies exist because this technology has been around for less than three years. We are demanding proof of the crash while the car is still accelerating. By the time we have definitive evidence of developmental harm, an entire generation will have served as test subjects for an experiment they never signed up for. We have not paused to consider that adolescents - those for whom identity, judgement and self-trust are still actively forming, are precisely the population for whom this particular displacement of cognitive effort carries the highest developmental cost. The question is not whether AI is useful. It is. But we are already releasing into the world the first generation to have outsourced the formation of their own minds. The generation behind them are still forming theirs, and although it may be the only window we have, it is still open.
Also in this series:
👉 Educational Technology
👉 The Moral Refinery
👉 Beyond Capture
👉 Unconditional Warmth, Unpredictable Harm
👉The Pilot’s Test Subjects
This is a superb articulation of the catastrophe we are sleepwalkimng into. You've nailed the psychological threat to our youth so clearly.
What an amazing and shocking article. I too want to shout out to the world and ask it to wake up! AI is creeping in like a dense fog, if this is what the education system is left with then we are allowing adolescents to no longer use their brains for anything anymore. This is shocking on so many levels! That they will miss out on rewarding feeling of personal achievement, that they figured it out by themselves through research and practice!? Not to mention actual social interaction!! Thank you for articulating such the facts so beautifully