SmarterArticles

There is a chart taped inside the door of almost every paediatric clinic in the developed world. It is so familiar that most parents stop seeing it, the way you stop seeing the safety card in the seat pocket of an aeroplane. Two smooth bands of curves rise from left to right, and somewhere on them, plotted at every visit, is a single dot: this child, this height, this age, this month. The dot is not interesting in itself. What makes it powerful is the curve behind it. Because there is a curve, a clinician can glance at the dot and know, in seconds, whether a child is growing as a healthy child of that age should grow, or whether something has gone quietly wrong. If the dot falls more than two standard deviations below the median height for the child's age, the clinic has a word for it, and the word triggers an investigation. The word is stunting.

We have had that curve, in one form or another, since 1977. We have nothing remotely like it for the mind. And we have just begun, at planetary scale and without anything resembling consent, to do to children's cognitive development the one thing that the growth chart was invented to catch: to interfere with it during the window when it matters most, while having no way to see whether the interference is helping or harming until the children in question are grown.

This is the argument that has crystallised, in the spring of 2026, around a deliberately uncomfortable analogy. Rebecca Winthrop, who directs the Center for Universal Education at the Brookings Institution and has spent a career studying how children learn across more than fifty countries, has become one of the most articulate voices warning that artificial intelligence may be doing something to children's developing minds for which our existing vocabulary is inadequate. The fear she keeps returning to, drawn from her conversations with educators, parents and students worldwide, is not abstract. The thing they worry about most, she has said, is children “stopping being able to think well”: a cognitive offloading so habitual, so early, and so invisible that the capacity to think independently never gets built in the first place. The provocative framing that has attached itself to this concern borrows the language of paediatrics. If a child can be physically stunted by a deficit during a critical developmental window, the question goes, what would it mean for a child to be cognitively stunted by the same mechanism, and why do we have no chart on the clinic door to detect it?

This article is not, primarily, another entry in the long and increasingly tired genre of “is AI rotting children's brains”. The mechanism by which effort builds cognition, and the danger that outsourcing the effort prevents the building, has been argued elsewhere and is taken here as the premise rather than the thesis. The harder and stranger question is the one underneath it. Suppose the worry is real. Suppose a generation is, in fact, being cognitively shaped by tools nobody fully understands. How would we know? What would the chart on the door even measure? Who would collect the data, against what baseline, how often, and what would the dot below the line oblige anyone to do? The scandal, on this reading, is not merely that we might be harming children. It is that we have built no instrument capable of telling us whether we are, and we have started the experiment anyway.

What stunting actually is, and why the analogy is exact

To understand why the analogy is more than rhetorical, it helps to be precise about what physical stunting is and what makes it detectable. Stunting is not simply shortness. It is impaired growth and development, most often resulting from chronic undernutrition during the first thousand days of life, that leaves a child too short for their age by a specific, agreed, internationally standardised margin. A child is classified as stunted if their height-for-age falls more than two standard deviations below the median of the World Health Organization's Child Growth Standards; below minus three standard deviations, the classification becomes severe. Those numbers are not arbitrary thresholds invented by committee. They are pinned to a reference population of how healthy children actually grow when the conditions are right.

That reference population is the quiet triumph behind the whole edifice. Between 1997 and 2003, the WHO ran the Multicentre Growth Reference Study, gathering data from roughly eight and a half thousand children across six deliberately diverse settings: Brazil, Ghana, India, Norway, Oman and the United States. The crucial methodological choice was to enrol only children raised under recommended health conditions, the children of non-smoking mothers, breastfed, with access to good nutrition and care. The resulting curves, published in 2006, are therefore not a description of how children do grow, which would merely encode the world's existing deprivations. They are a prescription for how children can grow when nothing is holding them back. A child measured against that standard is being asked a sharp question: are you growing as you would if your environment were not failing you?

This lineage runs back further. The first widely used growth charts in the United States were produced by the National Center for Health Statistics in 1977 and were promptly adopted by the WHO for international use; the Centers for Disease Control and Prevention revised them in 2000 before the WHO standards superseded them for the youngest children. The point is that the infrastructure took decades to build, was repeatedly refined, and rests on an enormous, boring, unglamorous foundation of measurement. Because that foundation exists, stunting is not a vague anxiety. It is a number, tracked annually across almost every country on earth through the Joint Malnutrition Estimates maintained jointly by UNICEF, the WHO and the World Bank. In 2024, those estimates put the number of stunted children under five at roughly 150 million, around 23 per cent of all children that age. We can argue about how to bring that number down. We cannot pretend we do not know it. That is the difference an instrument makes.

Now hold the cognitive case against that standard, point for point, and watch the parallels hold and then break. Stunting has a critical window, the first thousand days; cognitive development has its own sensitive periods for language, executive function and abstract reasoning, longer and softer but real. Stunting has a clear mechanism, nutritional deficit during that window; the cognitive worry has a clear proposed mechanism too, the outsourcing of the effortful cognitive work through which capacity is built. Stunting has a reference population of optimal growth; cognition has nothing of the kind. Stunting has an agreed threshold and a global monitoring system; cognition has neither. The analogy holds exactly until the moment it matters most, and then it falls into a void. Every element that makes physical stunting actionable is precisely the element missing on the cognitive side.

The mechanism we are not measuring

It is worth stating plainly what the instrument would need to detect, because it is not mysterious. The science of how skill is built from effort is among the better-replicated bodies of work in psychology. The UCLA cognitive psychologist Robert Bjork, with his collaborator and wife Elizabeth Bjork, spent decades establishing what he called, in 1994, “desirable difficulties”: the counterintuitive finding that conditions which make learning feel slower and harder in the moment, retrieving an answer before checking it, spacing practice, generating your own examples, produce far stronger long-term retention than conditions which make learning feel smooth. The struggle is not the obstacle to learning. The struggle is the learning. The feeling of fluency, of material going down easily, is a notoriously poor guide to whether anything durable has been built.

A growing literature suggests that generative AI is, by its nature, a machine for removing desirable difficulties. A study by researchers at Microsoft and Carnegie Mellon, presented at the 2025 CHI conference, surveyed 319 knowledge workers who used generative AI tools at work and analysed 936 first-hand examples of that use. Its central finding was that the more a worker trusted the AI, the less critical thinking they reported doing; cognitive effort was offloaded to the tool, and the workers who relied most heavily on it produced a less diverse range of outcomes. A separate and much-discussed study from the MIT Media Lab, published as a preprint in June 2025 under the title “Your Brain on ChatGPT”, had 54 participants write essays while wearing EEG headsets. Those who used a large language model showed measurably lower neural engagement across networks associated with attention and memory than those who wrote unaided, and grew more passive with each essay; the authors described what was accruing as “cognitive debt”. None of this is new in kind. As long ago as 2011, the psychologists Betsy Sparrow, Jenny Liu and Daniel Wegner described in the journal Science what became known as the Google effect: when people expect information to remain available externally, they remember it less well themselves. The instinct to offload is old. What is new is a tool that will offload almost any cognitive task you care to hand it, deployed to children before the capacities being offloaded have formed.

The reason this is so much harder to measure than physical growth is structural, and it sits at the heart of why no chart exists. Height is a competence and a performance at once: a child who is tall simply is tall, and you can read the fact off a wall with a pencil and a tape. Cognition is not like that. A child who produces a competent essay has demonstrated a performance, but the performance does not tell you whether the underlying competence exists, because the performance can be borrowed. This is the gap that the desirable-difficulties literature has obsessed over for thirty years, the chasm between the feeling of understanding and the fact of it, and AI widens it into a canyon. A child prompting a chatbot to write a five-paragraph essay will hand you a five-paragraph essay. Any instrument that scores the essay will record a capable student. What the instrument cannot see, without doing something quite different and far more intrusive, is whether the child could have written it alone, defended its claims, or noticed the one sentence in it that is subtly wrong. We are, in other words, trying to measure the one thing our existing tools are built to be fooled by.

Why the instruments we already have do not add up to a chart

It is tempting to assume that the measurement problem is already solved, that schools are awash in assessment data and surely one of those streams must capture what matters. They are awash in data. None of it is a growth chart for cognition, and understanding why is the crux of the whole argument.

Consider the large-scale international tests first. The OECD's Programme for International Student Assessment, the nearest thing the world has to a standardised cognitive measure across countries, runs only every three years and publishes results with a lag of well over a year. It samples fifteen-year-olds, which means that by the time a cohort is tested, the developmental window the analogy worries about is largely behind them. And it measures, by design, performance on tasks, the very layer at which AI assistance is most easily mistaken for ability. PISA is a magnificent instrument for comparing school systems. It is structurally incapable of functioning as an early-warning system for the cognitive formation of young children, because it arrives years too late and measures the wrong layer.

National examinations are no better suited to the role, and arguably worse. They are spaced years apart, periodically rewritten in ways that break comparability, optimised to certify achievement rather than to detect developmental drift, and increasingly contaminated by the same problem, since a competent prompt produces a competent answer. The whole apparatus of summative assessment was built to ask “has this student met the standard?” It was never built to ask “is this child's capacity to think developing as it would if nothing were interfering?” Those are different questions, and only the second is the cognitive analogue of plotting a dot against a growth curve.

What about the more modern candidates, the technologies sold precisely on their promise to see inside the learning process? Learning analytics, the harvesting of fine-grained data from digital learning platforms, can tell you a great deal about behaviour: how long a pupil lingered on a page, how many attempts a problem took, where attention wandered. Formative assessment, done well, gives a skilled teacher a running sense of where understanding is forming and where it is not. Both are valuable. Neither is a growth chart, for two reasons that recur throughout this subject. First, as researchers in the field readily acknowledge, learning analytics remains weakly connected to any theory of how learning actually happens, and rich in correlations whose meaning is contested; it measures engagement with a platform, not the formation of a mind. Second, and more damning for the analogy, none of these tools has a reference population. There is no equivalent of the WHO's optimally raised children, no curve of how cognition develops when nothing is holding it back, against which any given child's trajectory could be plotted. Without the curve, the dot means nothing. You can collect a billion data points about a child's clicks and still have no way to say whether the child is, in the cognitive sense, stunted, because you have nothing to compare the child to.

There are better instruments in principle, and they are revealing precisely because they are so rarely used at scale. Get a child to reason aloud through an unfamiliar problem without a screen, and you can begin to distinguish the child who has internalised a process from the child who has only ever watched a machine perform it. Administer a neuropsychological battery and you can detect executive-function deficits that no content test will show. Observe a pupil completing a task the deliberately hard way and you can see the difference between performance and competence open up in front of you. These methods exist. They are expensive, intrusive, slow, and produce no headline number for a minister to brandish. They are, in short, everything a national monitoring system is institutionally disinclined to fund, which is exactly why none has been built.

The experiment we have already started

While the measurement gap remains a void, deployment has not waited. This is the asymmetry that gives the whole situation its moral weight, and it is worth stating in concrete numbers, because the numbers are not gentle.

By late 2025, the College Board reported that 84 per cent of American high school students had used AI tools for schoolwork. Surveys of teachers put generative-AI use among K-12 educators above 80 per cent. The California State University system signed a contract with OpenAI to put ChatGPT Edu in front of more than 460,000 students and tens of thousands of staff, described at the time as the single largest deployment of the tool by any organisation on earth; the contract was renewed in 2026 even after a survey of more than 90,000 students and staff found a majority of faculty reporting that AI had a negative effect on their teaching. In the United Kingdom, the Department for Education issued guidance in mid-2025 on bringing generative AI into classrooms, cautioning about hallucination, bias and the handling of children's data, and noting pointedly that many popular tools are nominally restricted to users aged eighteen and over. The global market for AI in education is measured in billions and rising. Dozens of national systems are folding these tools into the daily texture of childhood.

Put the two facts side by side and the shape of the thing becomes hard to unsee. We are deploying, at a speed and scale that would be the envy of any public-health programme, a set of cognitive tools whose effect on developing minds we cannot measure, in the precise developmental window during which, if the worry is right, the damage would be done and hidden. A pharmaceutical company that wished to give a new compound to every child in a country would be required, at minimum, to run trials, define endpoints, monitor for adverse effects and report them to a regulator empowered to halt the programme. We have done the cognitive equivalent of skipping all of that. We have administered the intervention first and left the question of how to detect harm as an exercise for the future, on the implicit assumption that if something were going badly wrong, somebody would surely notice. The growth-chart history is the rebuke to that assumption. Stunting was always happening; what changed in 1977 was that it became visible, and only once it was visible did it become something the world organised itself to reduce. Before the chart, the harm was real and simply unmeasured. The unmeasured child is not the safe child. The unmeasured child is the child whose harm has not yet been given a number.

The temporal structure of the danger is what makes the absence of an instrument so corrosive. Physical stunting at least announces itself in the present tense; a short child is short today. Cognitive shortfall of the kind being theorised compounds silently and reveals itself late. A child who never built argumentative stamina at nine may look entirely fine at nine, because nine-year-olds are not asked to sustain arguments. She may look fine at fifteen, when her assessments reward exactly the short-form, well-structured production that AI excels at generating. The missing capacity becomes load-bearing only at nineteen, facing a dissertation, or at twenty-seven, expected to be the one in the room who notices that the model's confident output is wrong. By then the window has narrowed, the environment has no incentive to reopen it, and, crucially, there is no record. Nobody plotted the dots. There is no chart to point to that would show when the line first dropped below where it should have been. The harm, if it occurred, will be undeniable in its effects and unprovable in its cause, which is the worst of all worlds for anyone hoping to act on it.

What building the chart would actually take

It is one thing to lament the absence of an instrument and another to specify it, and the specification is where good intentions meet hard constraints. If we wanted, genuinely, to build the cognitive growth chart, what would the work involve, and why has nobody done it?

The first requirement is the hardest, and it is the one the physical analogy makes most painfully clear. A growth chart needs a reference population, and the cognitive reference population we would most want is the one we can no longer assemble: children developing without AI, under otherwise optimal conditions, measured longitudinally on the capacities we care about. There is no pre-AI cognitive baseline of the right kind, captured at the right grain, ready to serve as the curve. The window in which it could have been gathered cleanly is closing as the tools saturate childhood. This is not a fatal objection, because cohorts can still be assembled with varying exposure, and natural experiments exist where access differs, but it means any chart we build now will be reconstructing the baseline under compromised conditions rather than inheriting a clean one, the way paediatrics did. We are trying to draw the curve after the experiment has begun.

The second requirement is deciding what to measure, and here the temptation to measure what is easy must be resisted absolutely, because measuring what is easy is how we got here. The instrument cannot score essays or test recall of content, the things AI produces on demand. It would have to target the underlying capacities: the ability to sustain effortful reasoning without assistance, to retrieve and recombine knowledge from memory, to detect when an argument does not hold, to tolerate not knowing long enough to work something out. Measuring those means measuring under conditions where assistance is withheld and the process, not the product, is observed, which is slow, expensive and individual. It means, in effect, building an assessment whose entire design principle is the inverse of every assessment optimised for throughput. It is the difference between weighing a child and watching how they grow.

The third requirement is cadence and custody, and these are as much political as technical. A growth chart works because the measurement is repeated at regular intervals by a trusted party with no stake in the result, and because there is an agreed threshold that converts a dot into an obligation. The cognitive equivalent would need periodic, process-oriented assessment from early childhood onward, conducted by bodies independent of the companies whose tools are under scrutiny, with thresholds agreed in advance that would trigger investigation. Each clause in that sentence is a fight. Who funds longitudinal studies that produce results on a timescale longer than any electoral cycle and embarrass whoever was in office when the line first dipped? Who is trusted to hold cognitive data on children when the institutions best placed to collect it are often the edtech firms with the most to lose? Who sets a threshold knowing that, once set, it converts a vague unease into a legal and moral demand for action that someone will have to fund?

And then there are the obstacles that have no clean answer at all, the ones that explain why this is genuinely hard rather than merely neglected. There is the privacy and surveillance problem: a serious cognitive monitoring system means assessing children, repeatedly and individually, in ways that generate exactly the kind of intimate developmental data that should make anyone uneasy, and the history of children's data being collected for their own good is not reassuring. There is the gaming problem: any high-stakes metric distorts the behaviour it measures, and a cognitive growth chart with teeth would invite schools to coach to it, hollowing out the very thing it was meant to detect, in a cognitive replay of every test that became its own target. There is the equity problem, which cuts in two directions at once: a chart could expose, and so help remedy, the way AI's cognitive effects fall unevenly on children with more or less support at home, or it could become one more instrument by which already-disadvantaged children are labelled and sorted. And there is the deepest problem, the one that makes physical stunting look almost simple by comparison: we do not have settled agreement on what healthy cognitive development under AI even looks like, because the tools are reshaping the cognitive ecosystem so fast that the target is moving. The WHO could define optimal physical growth because the biology of a well-fed child was stable. The biology of a well-thinking child in an AI-saturated world is precisely what is in dispute.

Whose job this is

None of these obstacles is a reason not to build the instrument. Every one of them was, in some form, an obstacle to building the physical growth chart, and the chart got built. They are reasons to be honest that it is hard, expensive and slow, and to start regardless, because the alternative is to keep running the experiment blind. The question that remains is the one the brief insists on, and it is the one most likely to be evaded: whose job is this?

The companies deploying the tools cannot be the primary custodians of the measurement, for the same reason the food industry does not certify its own nutritional claims. Their incentives run the wrong way, and the conflict is structural rather than a matter of bad faith. They can and should be required to instrument their products honestly and to surface data to independent researchers, but the chart on the door must be held by a party with no stake in what the dot shows. Schools cannot carry it alone either; they are already drowning, and asking individual teachers to run neuropsychological batteries is a category error. The work belongs, by its nature, to public institutions operating at the scale and with the independence that paediatric surveillance enjoys: national statistics offices, public health bodies repurposed or extended toward cognitive development, education ministries funding longitudinal cohorts they will not see results from in their own term, and the international bodies that already coordinate child-development metrics. The Joint Malnutrition Estimates are produced by UNICEF, the WHO and the World Bank acting together precisely because no single actor could be trusted or resourced to do it alone. The cognitive equivalent would require the same kind of patient, unglamorous, multi-decade institutional commitment, and it would have to begin now, while today's seven-year-olds are still young enough for their trajectories to be plotted from something close to the start.

That commitment is unlikely to be made, and the reason it is unlikely is itself the most damning fact in the whole account. We are not failing to build the cognitive growth chart because it is impossible. We are failing to build it because building it would force us to confront, in public and with numbers, what we have already chosen to do. The instrument is missing not despite the deployment but, in a sense, because of it: an uninstrumented experiment is one whose results can never indict the people who ran it. There is a long and dishonourable history of this pattern, of harms allowed to compound in the dark for exactly as long as the dark could be maintained, with lead in paint and petrol, with sugar, with tobacco, each of them obvious in retrospect and each defended at the time by the absence of the very measurements that would have made them undeniable. In every case the measurement, when it finally came, did not create the harm. It revealed a harm that had been happening all along, to people who had no chart on the door.

The child in the clinic gets weighed and measured because, a century of effort ago, somebody decided that the growth of children was important enough to count, and that not counting it was itself a form of negligence. We have not yet decided that about the growth of children's minds, and the absence of the instrument is not a neutral gap waiting to be filled. It is a choice, renewed every day that the tools spread further and the chart remains unbuilt: a choice to run the largest experiment on cognitive development in human history, on a generation that did not consent and a public that was never asked, and to ensure, by leaving the instrument unbuilt, that we will not have to know what it did until the children are grown and the window is shut. The unmeasured child is not safe. The unmeasured child is simply the one whose dot we have agreed, in advance, not to plot.

References

1. Winthrop, Rebecca. “AI's future for students is in our hands.” Brookings Institution, 2025. https://www.brookings.edu/articles/ais-future-for-students-is-in-our-hands/
2. Center for Universal Education, Brookings Institution. “A New Direction for Students in an AI World.” 2025. https://www.brookings.edu/projects/global-task-force-on-ai-and-education/
3. World Health Organization. “Stunting in a nutshell.” 19 November 2015. https://www.who.int/news/item/19-11-2015-stunting-in-a-nutshell
4. World Health Organization. WHO Child Growth Standards and the Multicentre Growth Reference Study (MGRS), 1997–2003. https://www.who.int/tools/child-growth-standards
5. Centers for Disease Control and Prevention. “Use of World Health Organization and CDC Growth Charts for Children Aged 0–59 Months in the United States.” MMWR Recommendations and Reports, 2010. https://www.cdc.gov/mmWR/preview/mmwrhtml/rr5909a1.htm
6. UNICEF, World Health Organization and World Bank Group. Joint Child Malnutrition Estimates, 2024 edition. https://data.unicef.org/resources/jme-report-2024/
7. Lee, Hao-Ping (Hank), et al. “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.” Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Microsoft Research and Carnegie Mellon University, 2025. https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/
8. Kosmyna, Nataliya, et al. “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” MIT Media Lab, preprint, June 2025. https://www.media.mit.edu/publications/your-brain-on-chatgpt/
9. Sparrow, Betsy, Liu, Jenny, and Wegner, Daniel M. “Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips.” Science, 333(6043), 2011, pp. 776–778.
10. Bjork, Robert A., and Bjork, Elizabeth L. “Desirable Difficulties in Theory and Practice.” Journal of Applied Research in Memory and Cognition, 2020. https://bjorklab.psych.ucla.edu/research/
11. College Board. “New Research: Majority of High School Students Use Generative AI for Schoolwork.” Newsroom, October 2025. https://newsroom.collegeboard.org/new-research-majority-high-school-students-use-generative-ai-schoolwork
12. OpenAI. “OpenAI and the CSU system bring AI to 500,000 students and faculty.” 2025. https://openai.com/index/openai-and-the-csu-system/
13. UK Department for Education. “Generative artificial intelligence (AI) in education.” Guidance and support materials, June 2025. https://www.gov.uk/government/publications/generative-artificial-intelligence-in-education
14. Banihashem, Seyyed Kazem, et al. “Optimizing Formative Assessment with Learning Analytics.” Review of Educational Research, 2025. https://journals.sagepub.com/doi/10.3102/00346543251370753

---

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Discuss...