When I was a kid, there were several media events which ironically brought society together in a way you almost never see today with social media. The release of certain albums or TV shows were a kind of simultaneous human experience which we have seemingly lost. In the 80s and pre-Internet 90s, my generation would go to record stores to buy albums on the day of release and read them cover to cover on the bus home before ever hearing them, or come into school the next day excitedly talking about a TV episode the night before, recreating scenes and acting out various lines to prove our having borne witness to the event, thus cementing our social affiliation with the group.

I can vividly remember my parents and their friends talking about one particular episode of show called Dallas and a cliffhanger where a key character, Bobby Ewing had been killed and properly killed, funeral and all, then in the first episode of the next season, stepped out of the shower and the shocking death attributed to his wife’s dream. All 30 odd episodes of the last season, the time and investment from the audience, written off by a retrospective design completely at odds with the previous season.

In the industry this is know as ‘retconning’, or retroactive-continuity; the often clumsy placement of a narrative element onto a story which was never originally planned for it. In the case of Dallas, the audience felt especially aggrieved because it was very obvious that the previous season was never written with this sleight of hand in mind and the show-runners had taken a fairly patronising option to resolve it.

Retrieval practice is a curriculum property, not an activity

In the science of learning, the practice of retrieving previously learned material from memory is about the closest thing we have to penicillin in learning terms. If you want to strengthen memory, the dominant finding over 100 years is the counterintuitive one that you should stop inputting material into memory and instead focus on extracting it. The crucial factor however is that there has to be something to retrieve in the first place.

Over the last 10-15 years, retrieval practice (and its symbiotic twin, spacing) have become an article of faith but increasingly I feel they are being used in ways which are not supported by the evidence. They have mutated from properties which exist to serve a curriculum into a range of techniques which give the impression of effective teaching but in reality, often grafted onto curricula never designed for them. In other words, they have become retconned narrative elements bolted onto a story with no actual narrative continuity.

In 2020, Doug Rohrer and colleagues at the University of South Florida took several of the most used Maths textbooks in US schools and coded every problem (over 13,000 of them) against an ingenious question: does this practice problem require the student to retrieve anything from a previous chapter or does it just live inside the existing chapter? In other words, are the practice problems blocked practice at the end of each chapter or are they designed with the science of learning in mind and designed with return paths to previous learning? The results were that around one in 10 problems involved any return path at all.

Then in another paper, the same team ran a randomised trial in over 50 classrooms and had two groups work on the same practice problems with the only difference being how they were arranged. In the first group, the problems were arranged in the way in the vast majority of textbooks, within the same chapter, usually at the end. In the second group, previously learned material was systematically interleaved using return paths as new topics came up. The results were that the second group outperformed the first in of the largest effects this kind of study has ever produced.

For me, all this confirms the fact that the science of learning has far more to do with curriculum design than teaching activities. Christine Counsell puts it far more brilliantly that I ever could; “you cannot Rosenshine your way into a curriculum”.

I think a lot of this accounts for the fact that teaching is based on a strong managerial model with a premium on the visible layer of teacher performance rather than the invisible architecture of learning underneath it. A five minute starter quiz fits well into a school culture of ‘learning walks’ and the optics of learning in two minutes whereas the return path of a curriculum needs someone to actually work through a highly complex domain of knowledge and codes it, the way Roher and his team did. A classroom full of kids raising mini-whiteboards like a Mexican wave might give the impression of learning but unless we examine closely what was being attended to then we risk the whole enterprise being pedagogical kayfabe.

An effective curriculum depends on well designed return paths

Let me be clear: retrieval practice and spacing are not so much teaching techniques, but rather tools for ensuring learning continuity within a broader landscape of knowledge. This whole process only works when old material is deliberately brought back into contact with to-be-learned material. And not just any material, but hinge concepts, threshold ideas and high leverage knowledge which genuinely moves the needle on student understanding.

A key question to be asked of any retrieval activity is this: what is this knowledge in service of? What is the downstream benefit of this thing I am doing upstream?

If students of Macbeth are retrieving dates and quotes but there isn’t a clear return path to the bigger thematic ideas they are meant to be building, such as Macbeth’s moral ambiguity or the corrupting nature of ambition, then students are merely cosplaying building knowledge. A student may be able to recall ‘Is this a dagger which I see before me?’ but still not be able to actually do anything with it. Because the quotation is not in service of itself, it’s in service of the bigger idea of a mind under siege from its own unbridled ambition. It is the broader pattern (planted in the opening scene, then paid off with Banquo’s ghost) of Macbeth seeing things that are not there but acting as if they are. To borrow again from Christine Counsell, retrieval practice without a destination or return path is an intransitive pedagogy, a pedagogy without an object.

Likewise the quick ‘Do now’ activity of recalling what they did last lesson. Now that might be effective, but the Rosenshine dictum of ‘retrieving material from last lesson’ is often manifested as an administrative ritual rather than something which builds deep understanding. If students are ‘retrieving’ something they did yesterday then this will often be easy retrieval and frictionless retrieval does almost nothing. The strength of the effect scales with the effort required to retrieve it. A retrieved memory still fresh is not modified much, whereas the same memory dragged back just as it teeters over the abyss of ‘catastrophic’ forgetting is transformed.

The evidence never showed that quizzing strengthens memory per se; it showed that retrieval at a delay, timed against forgetting, strengthens memory. The delay is the active ingredient here and the delay is not something a teacher can improvise on a Monday morning. It is a property of how the story was written in the first place, in other words curriculum design.

The science of learning was established in a lab not a classroom

The deeper and more uncomfortable truth behind all this is that the science of learning (specifically the findings around memory), was developed in conditions that usually look nothing like a classroom, in fact the classroom is in many ways hostile to the science of learning and memory.

Whenever I talk about this topic, I always find myself returning to Tom Perry’s review of cognitive science which showed that the lab evidence is strong but the classroom evidence is incredibly thin with most studies being carried out by researchers, with researcher built materials with conditions which are a world away from the reality of a classroom teacher. Controlled studies in a lab produce replicable results but studies done in a classroom are often mixed. In medicine, this is referred to as a gap between efficacy and effectiveness, whether the drug worked in a trial versus whether it worked when prescribed to a patient.

None of this means we should abandon the science and practice of retrieval or spacing. It means we need to be honest about what kind of thing it is. I think it’s highly unlikely that the way in which human cognitive architecture interacts with knowledge works one way in a lab and a completely different way in the classroom, but I feel we need to sound caution against the whole enterprise becoming pedagogical theatre at the expense of learning.

Retconning the curriculum by bolting on the science of learning to materials that were never designed with them in mind might look evidence based but we will find out later if any learning actually took place, and probably too late.

VanderBurgh’s concern was not about autopilot itself but rather the complete dependency on autopilot at the expense of the internal knowledge of how it works and the loss of the judgment to know when to use it. In 2009, his warning came true in the very worst way possible over the Atlantic. When Air France 447’s sensors iced over, the autopilot handed a perfectly good airliner back to its crew, the pilots, confused and out of practice, stalled it into the sea.

The thing I find really fascinating about this story and why I think it has so much import for learning and AI is that his core diagnosis was that the autopilot was in fact, not the problem. The truth is that automated function in planes saves far more lives than it endangers and I think the same will be true of self-driving cars. (I have no interest in the hackneyed claim that the machines are bad and the humans good. The machines are, on the whole, magnificent I think). The thing I think is so interesting is that when faced with an emergency, the pilots reached instinctively for more automation not less, when the right thing to do was often to drop down a level and fly. So the danger is not automation itself but over-dependency on automation and this is quite obviously what happens when we give novice learners unfettered access to chatbots.

But actually its worse than that because those pilots had already internalised lots of organised knowledge. What is happening now for many students is worse because they dont even know what they dont know. To understand why this is the case, we need to go back and look at some misguided ideas about learning.

You can’t connect the dots if you don’t have any

One of the dominant concepts of learning since the 1970s has been the idea that facts and internalised knowledge is the lowest rung on the ladder of learning, the part we can safely externalise, because you can always look it up (see Bloom’s taxonomy). Teach the transferable skills, ie the critical thinking, the questioning, the analysing, and let the facts take care of themselves; after all, they are only a search away, and half of them will be out of date within five years in any case. It is a seductive picture, and it is almost exactly wrong. In fact, in the words of Wolfgang Pauli, it’s not even wrong.

The person who addressed this misconception most clearly for me was E.D. Hirsch in a now-classic essay from 2000 in which he examined the misguided progressive view that a child could skip the tedious accumulation of knowledge and acquire instead a set of all-purpose skills, questioning, analysing, synthesising, interpreting, that would serve for a lifetime of looking things up. The trouble with that view as Hirsch expertly showed, is that you cannot connect the dots if you dont have any dots to connect in the first place. These kind of skills are not all-purpose at all, they are dependent on knowledge and you cannot “21st Century skill” your way around it.

The other thing here is that they are also highly domain-specific. You cannot analyse a text on a subject you know nothing about; you cannot evaluate a claim in a field whose hinterland of knowledge is unavailable to you. Comprehension itself, the plain act of understanding a sentence, depends on a vast store of background knowledge, and I am not talking about learning off facts rote here, I am talking about a deeply interconnected tapestry of knowledge that is, in the end, the very apparatus of thought itself.

This is what schema theorists have in mind when they talk about assimilation of knowledge and say we grasp a new thing only by fastening it to things we already hold: the richer the existing weave, the more points at which a new thread can catch. This is why knowledge doesn’t just add to what you know but also compounds what you are able to learn next, and why the child with the fuller store of knowledge pulls further ahead with every passing year. It is also the part that can never be looked up, because it is the very thing that makes looking up intelligible in the first place. A chatbot can hand a novice a ‘dot’, fluent, finished and correct but what it cannot do is weave the tapestry, because the weave is not a product to be delivered but a labour that only the mind doing the connecting can perform, and a mind that is handed the finished picture never learns to connect anything at all.

The individual store of knowledge is not interchangeable with the external one, ie a duplicate we keep for convenience and can shed once the external copy is reliable enough. It is the very precondition for using the external one at all. To look something up and understand what you have found, you must already know a great deal that you did not look up in the first place.

Knowledge in your own head is the very filter that makes the now constant tsunami of information in our current world legible. The fatuity of “You can always look it up” is not false because looking stuff up is bad; it is false because the looking up only works for someone who has already done the learning in the first place.

Why chatbots should be banned in the classroom

At his UCLA graduation last year, a young man named Andre Mai crossed the stage holding his laptop open to the crowd. On the screen, magnified on the arena's jumbotron, was ChatGPT. He was not hiding it. He was showing it off, and his classmates cheered, and when people later asked whether this was cheating he was baffled by the question: his professors, he said, had encouraged the use of AI. The subsequent viral video received tens of millions of views inside a few days, led to claims to take his degree back and that “we’re so cooked”, and revealed to the world something profoundly broken about education as we currently know it.

But the tragic part of this for me is not that this made a complete mockery of the occasion or that the he didn’t even see he had done anything wrong. It was that his professors, he said, had encouraged the use of AI, and so he had done exactly what he was told he could do.

This moment suggests to me that something has fundamentally gone wrong in education and that so many in the field simply have no clue how to respond to what’s coming. A degree is supposed to be a proxy for a changed mind: proof that the person holding it now carries knowledge, judgement and skill they did not possess four years earlier. What is the value now of that credential?

A really good piece on this was James Walsh’s New York essay, Everyone Is Cheating Their Way Through College in which Walsh introduces us to a Columbia undergraduate, Roy Lee, who breezed through the core curriculum on ChatGPT because, he shrugs, “most assignments in college are not relevant. They’re hackable by AI.” Why trouble to reach an Ivy League university at all? “It’s the best place to meet your co-founder and your wife.”

More alarming, is the account of a student called Wendy, who used the chatbot to write a paper on critical pedagogy (in fairness the human slop from this field is arguably worse than anything AI has ever produced), an essay arguing that learning is what “makes us truly human”, and could not see the irony when it was put to her. Her own account of the method is the most damning line: it “just gives you straight up what you have to follow. You just don’t really have to think that much.”

Almost all of this commentary though, seems to view the problem through the lens of cheating as if this is all the students’ fault. The anxiety is about integrity, detection, the arms race between the essay mill and the plagiarism scanner. A few people seemed to have grasped that this cannot be the whole of it. Phil Christman, who teaches first-year writing, comes closest in Slate: the chatbot, he grants, “hurts their ability to learn”, but the deeper casualty he says is “the trust that any teaching relationship depends on”. He is right that trust is a casualty, and he is not alone in his unease; a New Yorker essay by Jessica Winter recently conveyed the turning tide, “a growing number of parents, educators, and cognitive scientists” now saying that AI does not belong in the rooms where children learn.

So how bad is the problem? In the year before Andre Mai crossed that stage, the share of American teenagers who told Pew they used ChatGPT for schoolwork had doubled, from 13 per cent in 2023 to 26 per cent in 2024, and because familiarity is the strongest predictor of use, with 56 per cent of the teenagers who know the tool well already leaning on it, that number is probably a conservative estimate.

But the thing I find truly dispiriting and frankly bizarre about this is that the broader education establishment not only seems to have no clue how to address this, but in many cases they are actively making it worse. Many exam boards are passing the buck, invoking the language of “academic integrity”, as though again, the problem were with the students rather than their own abdication of responsibility.

The International Baccalaureate has got this spectacularly wrong I think and managed the rare feat of writing a policy that refutes itself in its own supporting examples. Artificial Intelligence, the IB assures us, “will become part of our everyday lives, like spell checkers, translation software and calculators,” and so students may use it provided they treat it “as with any quote or material from another source” and reference it in the bibliography. Except of course that chatbots are not just “any quote” or “source”, they’re completion engines.

A spell checker corrects a word; a calculator resolves a computation; each automates a bounded clerical task but then hands the actual reasoning back to the student to complete. However, a chatbot writes the whole damn essay.

Let’s be clear about two things here. Firstly, in any kind of proper scholarship, a “source” is something you consult, and then have to do the real work with; read it, question it, weigh it against other sources, and fold it into an argument that is your own. Chatbots on the other hand bypass all of that, they are a completion engine which does the work you were meant to do. Asking a student to reference the AI that produced their paragraph is not a “safeguard against academic misconduct”. It is giving students the flimsiest of guidelines to work with and hoping for the best. It is also something that will ultimately penalise the ones who are the more academically “honest” and diligent.

Secondly, you cannot cite your way out of not having done the thinking. A citation records where knowledge came from; it cannot manufacture the knowledge the citation was meant to certify. For someone who already mastered a domain of knowledge, this doesn’t matter, because the expertise is already built and the reference is just an honest note about what led to the learning. For the novice however, whose entire task is to build that expertise, so “reference the AI” is not a safeguard, its a free pass to bypass thinking altogether dressed up as “academic integrity”.

So to say that citing a chatbot cannot manufacture the thinking is true, but it only hints at the symptom not the real cause. The deeper question is why handing a novice the finished answer should damage learning at all, when handing a novice a good explanation plainly helps it. The answer is not obvious, and here I will turn to one of the most important papers in educational psychology.

Why Minimally Guided AI Does Not Work

In 2006 , John Sweller and Richard Clark published a paper which had a huge influence on me and many other educators based in the UK : Why Minimal Guidance During Instruction Does Not Work. Their target was the whole family of “minimally guided” approaches such as discovery learning, problem-based learning, inquiry learning, constructivism, that broadly share a single misguided idea: that learners build knowledge most durably when they construct it for themselves rather than being shown explicitly.

Their argument against it was not so much ideological but architectural. Working memory is narrow and easily swamped; long-term memory is vast; learning is the effortful business of moving something from the first into the second. Minimal guidance fails because it floods the narrow channel with the wrong load, leaving novices to hunt for information they lack the schema to find. And the advantage of guidance, they observed, recedes only once a learner knows enough to guide himself.

Underneath this sits a distinction that John Sweller arguably built cognitive load theory on: the brilliant discovery that solving a problem and learning from it are not the same thing, and can even work against each other. A novice can reach the right answer by trial and error, or by working backwards from the goal (means-end analysis), and form no durable schema at all, because every scrap of his working memory went into reaching the answer rather than into grasping the structure that produced it. Performance and learning are not the same thing. You can watch a student succeed and be watching him learn nothing: a chatbot is a machine for maximising the first while giving the illusion of the second. It manufactures correct answers at scale and builds nothing behind them, which is precisely the pattern the trials keep turning up, scores up during practice, knowledge gone by the test.

At which point a fair objection emerges here. Direct instruction also hands knowledge over; the whole case against discovery learning is that novices should be shown rather than left to flounder. So why is a chatbot not simply very efficient teaching, the logical end of the same argument? Because the two hand over entirely different things. Good instruction hands over a carefully designed structure built to be absorbed: the worked example you study line by line, the explanation you have to follow, the guidance that deliberately fades as your own competence grows. Its entire design is to get knowledge into your long-term memory and then step back. A chatbot hands over the finished answer to the task in front of you and asks nothing of your memory whatsoever. One is a scaffold, put up so that something can be built and then taken down; the other is a crane that lowers the finished object into place and leaves you standing beneath it, having lifted nothing and learned nothing about how it was made.

An open chatbot in the hands of a novice is therefore minimal guidance carried to its logical conclusion. In fact it’s worse, it is discovery learning with the actual discovery removed.

What grokking tells us: There is no shortcut to understanding

A digression. There is a curious echo of all this inside the machines themselves. When you train a neural network on certain tasks it will often do the obvious thing first: it memorises. It fits the examples it has been shown perfectly and remains hopeless on anything new, having learned the answers rather than the rule that produces them. Researchers call the thing that sometimes happens next “grokking”1, a term which comes from Robert Heinlein's brilliant 1961 novel Stranger in a Strange Land, in which "grok" is a Martian verb meaning to understand something so completely that you merge with it, that it becomes a part of you.

So with LLMs, long after the network has aced its training set, and only after a great deal more effort, its performance on unseen problems suddenly climbs, as though it had at last stopped reciting and started understanding. The key idea here for me is that memorisation is fast and cheap; generalisation is slow, and arrives late if it arrives at all, and on some tasks never comes, because the data stays, in the researchers’ own phrase, effectively random to the model.

In other words, the network prefers the patterns that are quick to learn and poor at transferring, and only reluctantly gives them up for the slow ones that actually generalise. I offer this as analogy and not as evidence; a transformer doing modular arithmetic is not a child, and I have spent too long objecting to brain-based hand-waving to smuggle in a machine-based version of it here. But the shape of the thing is instructive I think. Even the systems we built to be shortcuts confirm that there is no shortcut from having seen the answer to understanding it. Fluency without the slow work is a performance, and it shatters the moment the task is one it has never seen.

AI is bad for novice learning but a gift for expert learning

So how to we make sense of all this? Fitzgerald once said that “the test of a first-rate intelligence is the ability to hold two opposed ideas in the mind at the same time, and still retain the ability to function.”2 I certainly don’t consider myself to have a first rate intelligence but I have tried to live by this maxim wherever possible and also how it resonates with one of Fitzgerald’s (and my) idols John Keats and his idea of negative capability.

Since I first encountered LLMs in the form of chatGPT back in those heady days of 2023, I found myself thinking two counterintuitive things about AI and learning:

1. AI is bad for novice learning.

2. AI is a gift for designing learning.

Back in 2023, I was seeing how disastrous chatbots were for novice learners in school and college but at the same time, I personally was learning more and with greater intensity, than I had since my days at University, and more importantly for my own work, I was learning more about learning science and instructional design than ever before. I was also fascinated at how LLMs were telling us something about the fundamental nature of learning itself.

For a while I held these two thoughts as a real contradiction, and it really nagged at me. If AI was hollowing out the learning of students, why was it so powerful for me? I resolved this eventually by realising that the two ideas were not opposed at all. They were the same idea seen from opposite ends of expertise. What gave me real clarity however (and to be frank, a kind of Damascene conversion), was visiting a highly unique school in the US with a group of other learning scientists and seeing how they were harnessing but also limiting the latent power of AI.

The case for Alpha School

Last year I visited Alpha school in Austin and met with Joe Liemandt and his team. One of the things which convinced me that something different was going on there was when he made it very clear that chatbots are banned for learning in the classroom at Alpha school and calling them “cheatbots”. This was surprising because I had mistakenly assumed that the most AI-driven school on the planet would be the very place that was creating the education equivalent of children of the magenta line at scale.

The other thing that surprised me was that everyone I met there from the engineers to the guides working with kids were all talking about Engelmann, Sweller, Kirschner, Ausubel, Willingham and all the people who had formed my own view of how learning happens. The counterintuitive thing for me was that their approach was not only to completely reject the minimally guided approaches which I have spent my career fighting against but how to engineer the best conditions for learning through curriculum design and learning science and through an intensive period of focus in the morning where the kids work towards mastery not simply time on task.

The missing element for me was creating the kind of constraints that I knew would lead to learning. I had assumed that the only place this works at scale is the classroom. Indeed, the history of edtech has been one of expensive failure and my view up until last year was that, to paraphrase an old colleague, the best way to use technology in the classroom is not to use technology in the classroom.

But the inconvenient truth is that learning doesn’t care how it happens. It follows the same stubborn principles either way: mastery before moving on, effortful retrieval, the right cognitive load at the right moment.

I also realised that almost everything I learned since the age of 12, I have learned on my own. Mainly through books, but also through videos like Ken Clark’s Civilisation and Ken Burns’ documentaries about America. I think this is why I have always disagreed with social constructivism as a means of teaching.

The thing which surprised me further was learning that the answer to the chatbot problem, also comes from AI itself and how Alpha have created a highly constrained platform where students not only cannot access chatbots, but have to complete the cognitive tasks put to them, designed by learning science. The AI monitors and evaluates what the student is doing at every step, tracking where they are fluent and where they stall, which misconceptions are surfacing in real time, what has been mastered and what needs to come round again. It calibrates the difficulty in real time, refuses to let a student move on until the current idea is genuinely secure, and tells the human guides precisely which child needs what. What it never does is the one thing an open chatbot does by default, which is the actual thinking. The aim is that the cognitive work stays exactly with the student. The machine’s job is not to spare the child the desirable difficulty but to engineer the right amount of it, and then to make sure it actually takes place.

And the thing I find most hopeful about this whole enterprise is that it is only getting started. There is a great deal still to do, and I say that as someone who now spends part of his working life doing it, because after that visit I was fortunate to be offered the opportunity to work with Alpha and an incredible team of learning scientists, educators and engineers on the learning science behind these systems. I mention this not as a disclaimer to be got out of the way but because it is the whole point: the visit changed my mind, and changing my mind changed what I do.

Admittedly, I want kids spending as less time as possible in front of screens. I would like to see them reading physical books for long, uninterrupted stretches every day, discussing/arguing with each other, playing sports, making things with their hands, out in the world. But I see no contradiction at all between that and what I have just described, because the entire point of compressing the academic core into a short, intense, well-designed block is that it buys back the rest of the day. Every teacher knows that if you could get kids to focus, really focus, for a few hours each day you would have most of the battle won.

The scarce resource in education is not time, it’s attention. So much of a school day is lost to friction: the settling and re-settling, the misbehaviour, the transitions, the admin, the low hum of thirty minds doing their best but half-present. Protect a short block of genuine concentration, and the amount a child can learn inside it is remarkable.

The other problem is that it is very hard to scale good teaching. This is the fact that every reform of the last century has struggled with. An uncomfortable truth is that the variation in teacher quality is real and often larger than we might like to admit. ( And I say this as someone who worked in a classroom for 18 years.) We know, with about as much confidence as education ever offers, what works well in the average: a positive environment, clear routines, high expectations, explicit instruction, motivating kids, well designed curricula, worked examples, spaced retrieval, feedback tuned to the individual child at the moment of difficulty, the slow accretion of knowledge in a sensible sequence.

But what we have so rarely solved is how to give thirty children at thirty different points all of that at once, from one teacher, in one room, for one hour. The best teachers approximate it through sheer virtuosity, craft and stamina, and often burn out doing so. This is why the effect of a truly excellent teacher is real and measurable and also, at the level of a system, almost impossible to reproduce. You cannot clone the person, and the moment you dilute the method to fit the average classroom, the very thing that made it work in the first place is lost.

This is where I finally understood what the technology was for, and it was not what I had feared. The core idea was never a machine that teaches in place of a human. It was a machine that does the one part of good teaching that is so hard to scale: the careful and precise sequencing of knowledge, the relentless, patient, individual tracking of where each child is, what they have mastered, what they are about to get wrong, and what they need next. That is not the glamorous part of teaching and it is not the human part; it is the sort of clerical substrate beneath the human part, the bookkeeping of thirty separate minds. So again with the paradox, the automation, done right, does not replace judgement. It creates the conditions under which human judgement can finally be spent where it counts.

The case for TimeBack (the AI-powered EducationOS behind Alpha), was never that screens are good; it was that the cognitive work must remain with the child. And once you believe that, and properly hold to that, the screen is there only for the narrow window in which it does something different; calibrating difficulty in real time, catching the misconception at the moment it forms in a way that is efficient and scalable. The morning block at Alpha is short by design. The point of compressing the academic core into a couple of hours of intense, guided, machine-monitored work is not to maximise time on screen but to minimise it, to buy back the rest of the day (6 hours) for other stuff that they would never get to do otherwise. The paradox I saw there was that technology enabled more human interaction, not less.

Used badly, technology expands to fill a childhood. has argued convincingly here that the anti-phone movement, however justified, risks turning into a simple philosophy of technological abstinence, and that the important question is not whether a tool involves a screen but what it does to a student’s attention and cognition. He is right of course, and it is the same point I am keen to make: the phone and the open chatbot are interruption and substitution engines respectively, tools that degrade the very faculties school exists to build. However, a tool designed to protect the difficulty rather than remove it is a very different animal altogether.

Beyond the Magenta Line

Which returns me, in the end, to Warren VanderBurgh and the line he drew across that cockpit display. His warning was never that automation was the problem, rather it was that a tool built to reduce our workload will, if we let it, quietly reduce our competence, and that the reduction is invisible right up until the point it is not, when the sensors ice over and the autopilot hands the aircraft back and there is no one aboard who remembers how to fly. The chatbot is in danger of doing to a generation of novices exactly what the flight computer did to a generation of pilots, with one major difference: The pilots of flight 447 had the knowledge but lost the access to it. The students we are now raising on frictionless answers are being denied the knowledge in the first place, and you cannot drop down a level and fly the plane manually if you were never taught to fly at all.

The task then, is not to keep AI out of education, which let’s be honest, is neither possible nor desirable, nor to wave it in under the flag of academic integrity, offload all the responsibility to the kids and hope for the best. It is to do the harder thing: to design the optimum setting for understanding to flourish, to provide the constraints under which learning happens, to engineer the conditions under which the cognitive work stays with the child. To build the tapestry first, by hand, in sequence, under guidance, and only then to hand over the tools that assume it is already there. Get that sequence wrong and we will produce, at extraordinary expense and remarkable speed, a generation of children of the magenta line, fluent in following and helpless the moment the line disappears. Get it right, and we may yet have the first genuinely new answer to an old question, which was not how to make learning easier, but how to make it happen at all. And the possibility of doing so at a scale we could never have previously imagined.

Growth Mindset strategies require little that is uncomfortable of teacher or pupil, which is exactly the type of idea that spreads regardless of whether it works. Belief only governs whether a student perseveres; but it cannot tell them what to persevere at, so they often persevere at ineffective habits they already prefer. However, the strategies that actually lead to learning such as retrieval practice and spacing, feel worse and therefore are not as popular.

Despite this, every parent and teacher knows that motivation and belief matters greatly. A child's conviction that she can improve does little on its own, but can do a great deal when it is supported by good mentoring and a genuine sense of belonging: the conditions that make her feel secure enough to bear the initial discomfort real learning demands.

There is a story I have probably repeated a little too often about Niels Bohr who was visited by a friend after winning the Nobel prize for his model of the atom. The friend, after seeing a lucky horseshoe over his door, asked him why an eminent man of rationality and reason would have a superstitious symbol over his door? Bohr replied that of course he didnt believe that it works, but that he was told that it works even if you don’t believe in it.

The growth mindset poster in schools is education equivalent of Bohr’s horseshoe. Most teachers instinctively know that showing children a video about brain plasticity, or telling them that mistakes help them grow, or adding ‘yet’ to a negative sentence, will not by itself make them better readers or help them solve quadratic equations. And yet we keep hanging them anyway, because the alternative, admitting we do not have a quick fix for how children learn, is far less comforting than a slogan on the wall.

The idea of a growth mindset remains one of the most successful ideas education has ever seen. It is short, intuitive, cheap, and almost too good to be true; an intervention can be delivered in as little as ten minutes through an online module, and it leaves both teacher and pupil feeling that something transformative has occurred. In a new paper, Anne Cleary and Daniel Robinson place it alongside learning styles and grit, in a small family of concepts that are quick to grasp, relatively easy to administer, and almost entirely resistant to disconfirmation or falsification. They borrow a useful term for this from Sinatra and Jacobson: the zombie concept, an idea that no amount of evidence can kill, because it keeps rising again merely on the strength of how right it feels.

However, there is a real dimension to growth mindset so why does it endure in its ineffective motivational poster form? And what does it tell us about how education chooses its ideas?

What the evidence actually says

The empirical picture is now difficult to argue with. Macnamara and Burgoyne’s meta-analysis found a mean effect of growth mindset training on academic achievement of d = 0.05; once you correct for publication bias, that already negligible figure falls to somewhere between 0.01 and 0.03, which is to say nothing at all. The result holds up under the most reasonable stress tests. When the authors restricted their analysis to studies in which the intervention demonstrably shifted students’ mindsets in the intended direction, the effect on performance vanished. When they restricted it again to the handful of studies meeting basic methodological standards, it vanished a second time.

This is not the verdict of a single hostile study either. Foliano and colleagues ran a randomised controlled trial across more than a hundred English schools and over five thousand pupils, and found no advantage for those who received the training. Bahnik and Vranka found no relationship between growth mindset and admission test scores among more than five thousand Czech applicants. And the studies that do report glowing results tend, on inspection, to often be correlational: Claro and Loeb’s much cited 2024 paper, which announced that students with a growth mindset learn more in school, can establish only that the two travel together, not that one produces the other. As Cleary and Robinson note, a growth mindset is just as plausibly the residue of prior success, or of an already acquired tolerance for the discomforts of effective study, as it is the cause of anything.

There is also the small, awkward detail that authors with a financial stake in mindset interventions were two and a half times more likely to report significant effects than authors without one. The for-profit programme Brainology was marketed to schools as an effective intervention, priced at several thousand dollars per school per year; and then, in 2015, the Institute of Education Sciences awarded a multi-million-dollar grant to investigate whether it actually worked. Cleary and Robinson give a great analogy; a pharmaceutical company selling a drug to patients first, and securing funding for the trial afterwards. This is not how serious science works and says a lot about why education is so prone to fads and bad ideas.

Why bad ideas travel so well in education

I’ve written before on why bad ideas like Learning Styles have been so successful despite the evidence against them, and I think the more interesting question is not whether growth mindset training works, but why its failure has done so little to slow its spread. Cleary and Robinson offer some really interesting thoughts on this. They scanned a year of journal articles and sorted the popular concepts from the well-supported ones by the kind of evidence each attracts. The methods that dominate classroom practice, (grit, growth mindset, learning styles etc), were studied almost entirely through observation: ninety per cent observational, but very few experimental. The methods that learning scientists actually recommend, retrieval practice, spacing, cognitive load theory, were tested overwhelmingly through intervention, with the majority experimental. In other words, the popular ideas are the ones we celebrate rather than test; while the effective ones are the ones we test rather than celebrate.

That asymmetry is not an accident, and here the paper connects to something deeper than mindset. Effective learning strategies share an unfortunate property: they feel bad initially. Retrieval practice, the single most robust finding in the field, is unpleasant in the moment, because the difficulty of dragging information out of memory is precisely the mechanism that strengthens it. Spacing out your practice feels worse than cramming the night before an exam. Re-reading feels productive and is largely useless. Learners, left to their own “metacognition”, reliably choose the strategy that feels like learning over the one that produces it, and they are wrong almost every time. Robert Bjork’s phrase for the useful kind of struggle, the desirable difficulty, captures the whole problem in two words: the desirable thing is, by design, difficult, and the difficult thing is often hidden.

Growth mindset is the mirror image of this. It is a feel-good method in a field where the things that feel good rarely work and the things that work rarely feel good. It propagates and flourishes for the same reason misinformation propagates: it makes sense at the level of gist, it fits a cherished cultural narrative about effort and self-authorship, and so it bypasses the scepticism we reserve for ideas that sound strange. Nobody properly interrogates an idea that confirms something which they have already decided is true.

Hard now. Easy later: why learning needs friction

There is a slide I use in almost every talk I give which features another quote from Anne Cleary which I think brilliantly explains the whole problem in a single sentence:

We treat the feeling of progress as evidence of progress, and so we gravitate towards the strategies that supply that feeling most generously: re-reading, highlighting, repeating comfortable routines, the warm glow of an assembly about the plasticity of the brain. Yet the sensation of fluency is precisely what genuine learning often prevents. In the short term, retrieval feels like failing; spacing feels like forgetting; difficulty feels like incompetence; and each of those uncomfortable signals is very often the mechanism actually doing its work.

This is the paradox the growth mindset can never resolve, because it operates entirely on the wrong side of it. Mindset is engineered to make learning feel better; but the things that genuinely lead to long term learning are (often) the ones that feel worse. You cannot growth-mindset your way around that trade-off, and you certainly cannot believe your way around it; you can only teach a student to recognise the discomfort for what it is, and to carry on regardless.

At the risk of contradicting myself, there is a wall decal at Alpha School's Spyglass campus in Austin which I think gets to the heart of what we should actually be saying to children. Fittingly, it runs along the wall above a stairwell, and it reads:

What I like about this (and particularly where it’s placed) is that the classroom growth mindset poster tells a child that her intelligence is hers to create, that she can somehow believe her way up the steps, but then leaves her at the foot of the stairs with no idea how to climb. "Hard now. Easy later." does the opposite. It makes no claim about ability and offers no flattery; it simply states the truth: some things are hard at first and then you get good at them with purposeful effort.

Slogans are usually nonsense but I like this because it tells a child that the discomfort she is about to feel is not a sign that something has gone wrong, but the price of the thing going right; that the path which feels worst in the moment is often the one that pays off later. This is not motivation in the familiar sense; it is honesty about the structure of learning. A slogan that tells children how easy initial learning ought to feel is simply dishonest. And we should be honest with children about how learning happens.

Does the causal arrow point the other way?

There is I think, a structural reason why standalone therapeutic interventions like this dont work and it’s not because ‘we’ve never tried growth mindset properly.” Suppose the training succeeds perfectly; suppose a child walks out genuinely persuaded that her intelligence is malleable and that effort will pay. What, precisely, is she meant to do next? Belief governs whether she persists but tells them nothing about what she should be persisting at.

A student convinced of her own brain plasticity but left to her own devices will simply persist harder at the ineffective strategies she already favours: re-reading, highlighting, the comfortable rituals of revision that feel like work and accomplish little. You cannot believe your way into knowing how to study, any more than you can believe your way into knowing how to swim.

Motivation does not reliably produce achievement; but achievement often produces motivation. Self-belief is far more often the consequence of competence than its cause. Prepare a child to deliver a genuinely good speech, one built through explicit instruction, modelling, specific feedback,rehearsal, and caring mentorship and support, and you will have done more for her sense of her own capabilities than any number of assemblies about the plasticity of the brain or videos of Michael Jordan failing 99 times but then succeeding once. You cannot instil the feeling that competence produces, while completely skipping the actual competence.

Mentorship, motivation and belonging

One person who I think has thought about this better than most is David Yeager, who arrived at a version of the same conclusion long before the latest meta-analyses. I have written before about his important work on ‘stealth interventions’ and his work since the National Study of Learning Mindsets rests on the argument that belief moves achievement only where the surrounding context affords the behaviour it recommends. The intervention can change what a student thinks while changing nothing about what she does, unless peers and teachers already reward the seeking of challenge. He presents this as a defence of mindset, but it is closer to a concession I think, because it grants the structural point precisely: belief is inert without a vehicle to carry it.

Yeager argues that the near-zero averages in the meta-analyses are misleading, an artefact of pooling the contexts in which the effect is real with the many in which it is absent; the effects are heterogeneous, they argue, not absent. Grant them this entirely, and the case against the classroom product is stronger, because if belief pays off solely where the environment is already supportive, then the universal, context-blind deployment, the module, the assembly, the laminated poster, is exactly the condition under which the research predicts nothing will happen. The most authoritative case for mindset turns out to question the version that many schools were actually sold. What remains is not a disagreement about whether belief needs a complement, but about which one: Cleary and Robinson would attach retrieval practice; Yeager would attach a supportive context and the deliberate pursuit of difficulty.

It is telling, too, where Yeager has since taken his thinking. His recent work has migrated from the growth mindset, a belief lodged inside the student, to what he calls the mentor mindset, a stance adopted by the adult: high standards paired with the support that makes meeting them possible. He sets it against two failure modes a teacher will recognise instantly, the enforcer who demands rigour without any help, and the protector who offers warmth while quietly lowering the bar. His exemplar is a physics teacher who sets genuinely hard problems, refuses to simply hand over the answer, scaffolds the task until students can attempt it, and lets them revise when they get it wrong.

This dovetails nicely I think with ’s thoughtful work on belonging. You cannot do belonging, Hutchinson argues, any more than you can believe your way into learning; it is earned by a child actually doing something that counts as membership of the group. A great point he makes is that the harder you push a child academically, the more her sense of belonging matters, because difficult material makes her struggle, and struggle is exactly what prompts the anxious question of whether she belongs here at all; a question that depletes the very cognitive resources the work requires. I like both these ideas because they are really the support that makes difficulty bearable, and so makes rigour possible.

Dweck deserves great respect

To be clear, I have the utmost respect for Carol Dweck, (the pioneer of the whole movement), I see her as a serious scientist who has genuinely and honestly tried to examine a real psychological effect and her work has pushed the field of motivation and beliefs of intelligence forward immeasurably. I think the negative attitudes to her work as more a result of lethal mutations of growth mindset rather than the work itself. In fact Paul Kirschner and I dedicated a chapter to it in our book How Learning Happens.

So it’s important to say that recent criticism is not a complete demolition of her work; it is more a redirection, and the interesting part to me is what work like Cleary and Robinson’s suggest. Unlike learning styles, which is simply false and deserves to be buried, growth mindset may yet have a defensible role, not as a learning mechanism, but as a kind of antechamber to one. Their hypothesis worth looking at I think is that mindset training does not act on learning directly at all; it acts on a student’s willingness to tolerate the discomfort of strategies that do actually work. A 2024 study by He and colleagues found that growth mindset influenced mathematics achievement only indirectly, through self-efficacy and reasoning. Fink and colleagues found that the training helped disadvantaged students while leaving advantaged ones untouched, which is exactly what you would expect if its real function is to supply a tolerance for difficulty that more advantaged students have already acquired elsewhere.

If that is right, then the problem wasn’t the belief itself; it was deploying the belief alone, as though pure resolve were a substitute for knowledge and technique rather than a precondition for acquiring it. Add the mindset with a knowledge-rich curriculum, explicit instruction, retrieval practice and spacing, and with honest training about the gap between how learning feels and how learning works, and you might have something: a way of getting students not merely to want to improve, but to endure the specific, unglamorous difficulty through which improvement actually happens. The belief then becomes a stepping stone rather than a destination. Whether this survives contact with a real classroom is, of course, an empirical question, and the field’s record on scaling laboratory effects makes me skeptical but it is the right question I think, which is more than the last fifteen years of mindset enthusiasm managed to ask.

Ultimately then, the lesson is not that belief is worthless. It is that belief is often downstream of other stuff. You cannot believe your way into learning, because learning is not a state of mind; it is the residue of particular things done in a particular order, most of which feel worse than the alternatives while they are happening. The most useful thing we can give a child is not just a conviction that she can improve, but the specific means by which she actually does, at which point the conviction often arrives on its own as a by-product, earned and durable and longer lasting because she has the real sense of accomplishment. The sense that she has achieved something that she first thought impossible. At the risk of repeating myself; perhaps then, the best way to achieve a growth mindset might just be not to mention growth mindset at all.

Although the ‘science of reading’ appears to be a relatively new term, it appears in an American education journal as early as 1836, and for most of its life it meant nothing other than the study of decoding. The broader body of research came after that, but still quite long ago: Gough and Tunmer gave us the Simple View of Reading in 1986, and the journal Scientific Studies of Reading has been publishing since 1997 but none of that produced a movement anything like what we have seen in the recent past. It popularity is explained in part by journalism and podcasts such as Emily Hanford’s hugely successful Sold a Story in 2022, which carried the argument to a far wider audience than the academic literature ever had.

The honest answer is that if we are happy with the term science of reading (and “science” here is admittedly doing a lot of heavy lifting), then there probably is a science of writing. However, it is younger, more diffuse, and harder to translate into the kind of media-friendly slogans that tends to travel well. It is also, as Steve Graham and Young-Suk Grace Kim have repeatedly argued, inseparable from the science of reading itself. But two points emerge from reading through the broader body of evidence.

Firstly, the science of writing is much less unified than the science of reading. The science of reading rests on a relatively narrow biological fact, namely that the brain is just not built to read and must be taught to map letters to sounds. Writing however, has no equivalent single mechanism. It draws on several things: language, motor control, working memory, knowledge, and motivation. So its scientific base is broader and therefore less easily summarised in a single phrase.

Secondly, the strongest findings in writing research are about instruction rather than acquisition. We know with reasonable confidence what kinds of teaching improve writing, (we have meta-analyses to point to1). But we know less about the developmental trajectory of writing in the way that we know about phonological awareness or orthographic mapping. The science of writing is closer to a science of writing instruction than to a basic science of how children acquire the actual skill of writing. In other words, we are better gardeners of writing than we are biologists of it: we know what helps it grow without fully understanding the thing that is growing.

Ultimately, it simply does not yet have the thing the science of reading possessed at the equivalent stage: a usable theoretical anchor, a clear political argument, and a settled view of how it could be implemented in the classroom. The Simple View of Reading was clear and reductive (in a good way), and that was its strength. Writing has no equivalent in terms of clarity and may never have one, mainly because its the simultaneous orchestration of transcription, sentence-level syntax, discourse structure, domain knowledge, audience awareness and self-regulation, all of it held in working memory at once, and none of these components reduces to or substitutes for the others.

The act of writing is precisely the management of that load, not the exercise of any single faculty. When Berninger and Winn called their model the Not-So-Simple View of Writing, the name was an admission as much as a description: the field has looked for its own Simple View and concluded that the object will not yield one.

How to teach writing

So what does the science of writing actually look like in a classroom? Across the best books in the field, from Hochman and Wexler’s The Writing Revolution 2.0 to Harris and Graham’s work on Self-Regulated Strategy Development to Engelmann and Carnine’s older insistence on faultless communication, the same picture keeps emerging. Writing is taught well when it is broken into components the developing mind can actually hold, and when each component is rehearsed deliberately, in content the child already knows something about, before the next one is added.

The first thing to consider, which is common to all of these traditions, is to privilege the sentence as the unit of instruction. Hochman and Wexler describe the sentence as "the building block of writing and thinking, the complete thought," and they show, with twenty years of classroom evidence behind them, that until children can construct varied sentences they cannot produce coherent paragraphs, let alone essays.

Their reasoning is rooted in cognitive science. Working memory, they remind us, can hold only four or five new items for around twenty seconds before it collapses, and a novice writer asked to compose an extended piece is simultaneously juggling letter formation, spelling, word choice, sentence structure, organisation, and the content itself. The result is overload, and overload produces the broken, repetitive writing teachers know all too well. Graham's meta-analytic evidence converges on the same point from the other direction, with sentence combining showing reliable effects on writing quality across grades and learner profiles.

The strategies that follow from this insight are deceptively simple. Hochman's because, but, so exercise gives children a kernel sentence, "Abraham Lincoln was a great president," and asks them to extend it three ways, each conjunction forcing a different cognitive move: explanation, contrast, and consequence.

In a very good new book on writing, and Rikki Cole call these structures "cognitive levers," and they list them as a small but powerful inventory of subordinating conjunctions, each paired with the kind of thinking it elicits.

So Although prompts the consideration of more than one perspective. Because forces causal reasoning. At first glance invites a child to reconsider a first impression. Unless opens up exceptions to a rule. Even though signals contrast against expectation etc. The list is short, but the cognitive activiation each conjunction triggers is precise, and the discipline of writing a sentence that begins with one of them is, in effect, the discipline of constructing a thought under syntactic constraint. As Didau and Cole put it, grammar in this sense is not decoration but a "cognitive lever, shaping precise thought."

Harris and Graham's Self-Regulated Strategy Development extends the same logic to the level of the composition. Rather than treating writing as a process to be facilitated, SRSD treats it as a sequence of strategies to be taught, modelled, memorised, and gradually released. The six stages, Develop and Activate Background Knowledge, Discuss It, Model It, Memorise It, Support It, Independent Performance, are not, as is sometimes claimed, a recipe that crowds out thinking. They are scaffolds that free working memory for thinking by automating everything around it. The accumulated evidence base, over a hundred experimental and quasi-experimental studies across age groups, genres, and learner profiles, makes SRSD one of the most reliably effective writing interventions we have, with the largest effects coming from small-group, trained-instructor delivery and particularly strong effects for students with learning difficulties

The answering word

Writing is one of the most cognitively demanding activities humans ever undertake. It is also one of the most consequential: to write a sentence well is to think a thought well. Writing is the mechanism by which inchoate thinking is forced into a form that can be examined, challenged, and revised. Speech allows us to be ambiguous annd for misunderstanding to occur, writing demands that we focus our thoughts in a way that is far more precise.

The science of reading, for all its hard-won gains, addresses only half of the literate mind. A child who can decode fluently but cannot produce a coherent paragraph is not truly literate in a meaningful sense; they are a consumer of text who cannot yet participate in the long human conversation that text exists to sustain. Reading lets a child receive what others have already worked out. Writing is what allows them to work something out for themselves and to offer what Bakhtin would call the “answering word”.

Like many aspects of explicit instruction, there is also an equity argument here, one that rarely gets made in the same terms as the reading debate. Children from language-rich homes arrive at school having heard, on some estimates, tens of millions more words than their peers, and they have heard those words in the kinds of elaborated, subordinated, causally connected structures that map onto written prose.

Children who have not had that exposure are not less capable; they are simply further from the conventions of the written register, and without explicit instruction they will stay there. Explicit teaching in the form of sentence-level teaching, cognitive levers, SRSD scaffolds, and the rest of the apparatus described above are not enrichment for the already articulate. They are the means by which schools can give every child access to the syntactic machinery that the most fortunate learners absorb almost by accident.

Memory, Learning, and Knowledge

Retrieval practice stops working when cognitive load is too high

One of the most consistent findings in learning science is that retrieval practice outperforms restudy. But what happens when the material being retrieved is genuinely demanding?

Redifer, Myers, Bae, Naas and Scott (Instructional Science) had undergraduates study a lengthy, conceptually demanding research article, comparing retrieval practice (free recall, practice quizzing and question generation) against rereading. Retrieval conferred no advantage on the final test. Crucially, cognitive load mediated the relationship between retrieval practice and final test performance: higher load was associated with poorer delayed recall, meaning that when working memory was already stretched by demanding content, the additional load of active retrieval cancelled out the benefit.

This is not a refutation of retrieval practice. As I have said before, the science of learning is a probabilistic enterprise, and this finding highlights a boundary condition, an important one. It adds to a growing list of boundary conditions accumulating around retrieval practice, joining the prior findings on feedback timing, material type and retention interval that have progressively narrowed the conditions under which the testing effect reliably appears. Retrieval practice works, but it is also rapidly becoming a lethal mutation, which is why careful implementation is key.

The mechanism is straightforward: retrieval practice works by forcing effortful recall, but if the content itself is already consuming all available working memory, there is no spare capacity for that effortful recall to operate. The implication for teachers is that retrieval practice needs to be calibrated to the complexity of the material rather than applied as a blanket strategy. For introductory content with manageable element interactivity, retrieval is powerful. For dense, interconnected material that students are encountering for the first time, other approaches may serve better.

A new paper by Sigayret, Parmentier and Silvestre (Frontiers in Psychology), suggests a further reason to be cautious, their paper failed to find any testing effect across two experiments run on Prolific. The authors are careful to say this is not a boundary condition on retrieval but a warning about the limits of crowdsourced platforms for cognitively demanding learning research, where disengagement and attrition swamp the effect.

Every school doing mandatory “Do Nows” as their primary retrieval strategy should ask: what exactly are students retrieving, and how complex is it?

Training a thinking skill does not mean you can transfer it

Grimm and Richter published two preregistered experiments (N=304) in Learning and Instruction testing whether brief training in rational thinking would transfer to argument evaluation. Main finding is that the training worked: participants improved on the specific rational thinking tasks they practised, but those gains did not transfer to evaluating the quality of arguments, despite the two competencies being correlated.

This extends a long line of transfer failures which broadly shows that near transfer is a thing but far transfer is a much more difficult to capture. The correlation between two cognitive skills does not mean that improving one will improve the other. Transfer requires more than proximity; it requires explicit bridging, varied practice contexts, and often far more time than a brief intervention provides. This is another finding which challenges the generic “critical thinking” industry, which promises portable reasoning skills through standalone programmes.

The pattern across decades of transfer research is pretty clear: training a skill and seeing improvement on a test of that skill is not evidence that the skill has become generalisable. What transfers is knowledge, applied deliberately in new contexts, not abstracted competencies floating free of content.

A little knowledge is worse than none at all

Hong, Son and Kim’s study in Cognitive Research: Principles and Implications demonstrates something teachers innately know but rarely see quantified: brief exposure to information produced peak overconfidence that actually exceeded the baseline confidence of complete novices. Participants who had spent a short time with the material were more certain of their (often wrong) answers than people who had never seen it at all.

The danger here is not ignorance but the illusion of understanding that accompanies superficial exposure. For instruction, the implication is that partial coverage of a topic may be actively worse than no coverage, because it generates false confidence without the corrective feedback that deeper engagement provides. Students who have “done” a topic briefly are harder to teach than students who know they haven’t encountered it, because the former believe they already understand.

This connects to a broader pattern in the research: the conditions under which learning appears to have occurred (fluent performance, high confidence, rapid completion) are frequently the conditions under which durable learning has not.

Removing a cognitive tool hurts far more than adding one helps

Florean, Straga, Mantyla and Del Missier published a study in Cognitive Research: Principles and Implications demonstrating an interesting asymmetry in cognitive offloading: taking away a tool that students have been using damages performance far more than providing that tool in the first place improves it. Working memory capacity moderated the effect, with lower-WM individuals suffering the greatest disruption when the tool was removed.

This finding has immediate relevance to every classroom deploying AI tools, calculators, or any external cognitive support. The decision to introduce a tool is also, implicitly, a decision about what happens when it is withdrawn. If students build their understanding while relying on external support, they are constructing knowledge representations that depend on that support’s continued presence. Remove it for an exam, a different class, or a future year, and the architecture collapses asymmetrically: the loss is larger than the original gain.

As a general rule with tech, I would say that any tool introduced during learning should be introduced with its eventual removal in mind. Scaffolding should be designed to fade rather than to persist, and students should practise without the tool at regular intervals so that their mental representations do not become dependent on external support they will not always have.

02 — Instruction and Practice

The expertise reversal effect has a second variable

The expertise reversal effect tells us that instructional methods effective for novices become ineffective or counterproductive for more advanced learners. Zambrano, Narváez-Rivera, Sayago-Heredia and Jácome-León tested this in a medical learning study published in Learning and Instruction, and confirmed the basic pattern: low-knowledge students learned better from direct instruction, high-knowledge students from minimally guided instruction.

But the critical finding was the moderator. The reversal only appeared when element interactivity was high. When the content was relatively simple, there was no crossover, regardless of prior knowledge.

This matters because the practical question teachers face is not just “when should I switch from explicit to exploratory?” but “when should I switch given this particular content?” This paper suggests the answer now has two variables rather than one: prior knowledge and content complexity. Simple material can be taught the same way to everyone without much cost. Complex, interconnected material is where the instructional method needs to match the learner’s existing schema. This is useful I think because it gives teachers a more actionable decision rule than “it depends on the student.”

Erroneous worked examples work, but only under three conditions

Beege, Krieglstein, Wesenberg and Loibl published a study in the Journal of Educational Psychology that identifies when deliberately incorrect worked examples outperform correct ones. The answer: only with appropriate prompting, adequate scaffolding, and sufficient prior knowledge. Without all three conditions met, erroneous examples backfire.

When students have enough prior knowledge to recognise errors, when the prompting guides them to the right kind of comparison, and when the scaffolding prevents them from simply encoding the wrong procedure, the act of finding and correcting mistakes deepens understanding more than studying correct solutions.

The practical takeaway is conditional rather than categorical: erroneous examples are not a generic improvement over correct ones, but a powerful tool that requires three things to go right simultaneously. For teachers, this means using them later in a learning sequence rather than at the point of initial instruction, ensuring the errors are carefully chosen rather than arbitrary, and providing explicit prompts that direct attention to the critical features rather than leaving students to notice them independently.

03 — Motivation, Self-Regulation, and Wellbeing

The working memory–achievement link is mostly stable trait differences, not a causal lever

Morin, Wichstrøm, Steinsbekk and colleagues published a birth cohort study (N=833) in Contemporary Educational Psychology with biennial assessments from ages 8 to 16. When within-person and between-person effects were properly disaggregated, the working memory–achievement link reflected stable individual differences rather than ongoing within-person causal effects.

This is not good news for the “train working memory to boost achievement” brain-training industry. The correlation between WM and achievement is real, robust, and has been replicated hundreds of times. But this study shows it is not telling the causal story most people assume. Children with higher WM capacity tend to achieve more, but changes in a given child’s WM over time do not reliably predict changes in that child’s achievement. In other words, the association seems to be trait-level, not state-level.

The practical implication is that interventions aimed at increasing WM capacity as a route to academic improvement are targeting a stable characteristic rather than a malleable cause. Resources spent on WM training programmes would almost certainly produce better returns if redirected toward reducing the WM demands of instruction itself.

New paper claims the social media panic is overblown

A new study in Current Psychology using UK youth data from the BrainWaves Project showing that the already-small bivariate correlations between social media use and mental health outcomes (1–4% of variance) vanished entirely once theoretically relevant control variables were included. Social media accounted for effectively zero percent of the variance in depression, anxiety, social phobia, self-esteem, quality of life, mental wellness, and friendships.

This is not a new argument from Ferguson, he has been making essentially this case for over a decade, first about video games and now about social media. But the specific methodological point is important I think regardless of whether you find his broader position convincing. Most of the headline claims about social media harming youth rest on bivariate correlations: social media use correlates with depression, therefore social media causes depression. The problem is that social media use also correlates with dozens of other variables; sleep disruption, family conflict, pre-existing mental health difficulties, socioeconomic stress, and when you control for those, the unique contribution of social media shrinks toward zero.

The implication is not that social media is harmless. It is that correlational designs without adequate controls tell us very little about whether it is harmful, and that policy built on inflated bivariate associations is policy built on sand. For educators following the screen-time debate, this is a useful corrective to the more alarmist claims, though it is worth noting that cross-sectional data with statistical controls cannot establish causation in either direction, and Ferguson’s own framing is not neutral.

04 — Classrooms, Behaviour, and Attention

Seductive details harm the students who can least afford it

Seductive details can be described as interesting but irrelevant additions to instructional material, such as dramatic images, entertaining anecdotes or fun facts tangential to the core content. Colliot, de Pereyra and Boucheix published a study in Applied Cognitive Psychology showing that children with lower inhibitory control were disproportionately harmed by seductive details in instructional animations. The students most vulnerable to distraction were the ones least equipped to resist it.

This is important from an equity perspective. Seductive details are not just inefficient, they widen the gap between students who can filter out irrelevant information and students who cannot. The children with strong inhibitory control shrug off the distraction. The children without it are pulled away from the content that matters.

For teachers and curriculum designers, this means that the decision to include a fun-but-tangential video clip, a colourful-but-irrelevant image, or an amusing-but-distracting anecdote is not neutral — it is a decision that systematically disadvantages the students who are already struggling most with attention and self-regulation.

Negative feedback is fine. Mixed messages are not.

Powell, Rentzelas and Kambouri published a study in Learning and Instruction that should reassure every teacher who worries about giving honest feedback. Negative feedback per se did not harm performance. What harmed performance was giving mixed messages simultaneously: telling a student they are below average and that they have improved, in the same breath.

The mechanism is intuitive once you see it: mixed signals create interpretive uncertainty. The student has to resolve a contradiction rather than act on clear information. “You’re behind but you’re improving” sounds encouraging, but it forces the learner to decide which signal to attend to. Some focus on the “behind” and feel demoralised. Others focus on the “improving” and become complacent. Neither response is what the teacher intended.

The practical rule is simple: pick one frame and commit. If the message is “you need to improve,” deliver that clearly with specific guidance on how. If the message is “you’ve made progress,” deliver that clearly with specific evidence. Trying to soften hard truths by sandwiching them with encouragement does not produce the best of both worlds. It produces the worst of neither.

05 — AI, Technology, and Cognitive Tools

The most-cited “ChatGPT helps learning” paper has been retracted

The single most important item this month is a retraction. Wang and Fan’s meta-analysis claiming ChatGPT boosted learning outcomes, published in Humanities and Social Sciences Communications (Nature), has been formally retracted. Commentators have pointed out that Wang and Fan lumped together studies with very different designs and small, heterogeneous samples, and coverage of the retraction notes that the authors did not respond to the journal’s correspondence. The formal retraction note itself is narrower, citing ‘discrepancies in the meta-analysis’ that undermine confidence in the validity of the analysis and its conclusions.

This matters because an large proportion of the “AI improves learning” claims circulating in education policy and edtech marketing either cite this paper directly or cite papers that cite it. When read alongside the new Liu, Xu and Xie meta-analysis in Frontiers in Psychology, which found a large positive effect of generative AI on intellectual outcomes (g = 1.096) but detected significant publication bias for exactly those outcomes, and the picture becomes clear: the GenAI-in-education evidence base is unreliable at the meta-analytic level. The positive results are there, but they are inflated by selective reporting and methodological weakness.

None of this means AI cannot support learning. It means the current evidence for the claim is far weaker than its proponents suggest.

AI boosts scores and reduces stress but does not deepen learning

A three-arm randomised controlled trial published in Computers and Education: Artificial Intelligence (N=275) tested AI assistance in a CS1 programming course. Students with AI support completed tasks more successfully and reported less stress. But they did not show deeper conceptual learning. Performance and learning dissociated cleanly.

This replicates the Bastani et al. Harvard tutoring finding in a different domain: AI makes the task easier without making the learner more capable. The students who used AI looked better on task metrics while learning less. This is exactly the kind of result that verification systems need to detect, because any measure based on task completion or course grades would show AI as beneficial. Only a measure of actual learning, tested independently of the AI assistance, reveals the gap.

The question is not whether AI helps students perform but whether it helps them learn, and those are increasingly looking like different questions.

Pre-thinking before ChatGPT use produces lasting creative gains

Wong and Qiu published an RCT (N=196) in Educational Psychology Review comparing free ChatGPT use with a “think first” protocol that required independent ideation before AI access. Free ChatGPT use boosted initial creativity scores, but when the AI was removed, those gains collapsed. The think-first group produced more modest initial gains but retained their creative improvement independently.

This is one of the cleanest demonstrations of borrowed competence. The students who went straight to ChatGPT were performing with the tool rather than learning from it. Those who generated their own ideas first and then used AI to refine them internalised something durable. The mechanism maps onto the generation effect: producing an answer, even a wrong one, before receiving assistance creates a retrieval and elaboration event that strengthens learning.

The practical implication is not to ban AI but to sequence it: think first, generate first, struggle first, then bring in the tool. The cognitive work that precedes AI use is what makes AI use educationally productive rather than merely performatively impressive.

Supervised AI tutoring produces massive gains — but the key word is “supervised”

Oreopoulos and colleagues published an NBER working paper reporting results from a randomised controlled trial of Khan Academy in Indian schools. The gains were substantial. But the critical design feature was supervision: students used the platform during structured school time with teacher oversight, not independently at home.

This aligns with a growing body of evidence that technology-assisted learning works when it is embedded in a structured instructional environment rather than deployed as a substitute for teaching. The Khan Academy platform provided the content and adaptive sequencing; the teachers provided the structure, accountability, and human support that kept students engaged and on task. Remove either component and the effect would likely shrink.

For schools considering AI tutoring platforms, the lesson is that the technology is a complement to teaching rather than a replacement for it. The gains come from the combination, not from the software alone. This is expensive and labour-intensive, which is precisely why it works: there are no cheap shortcuts to effective instruction.

Quick Hits

Telling isn’t teaching, even when it’s explicit. Sasson and colleagues claim that passive explicit instruction defined as clear explanations without guided practice, actually made students worse at question-formulating. The active version, where students did the cognitive work, produced gains. The label “explicit instruction” is covering two very different things.

Head-to-head on irregular words, and both methods assume phonics. Lithgow et al. published a direct comparison of mispronunciation correction versus the heart-word method in Scientific Studies of Reading. The key takeaway here is not which wins but that both approaches accept irregular words still have decodable parts, the debate has moved past “phonics vs whole language” to “which phonics?” technique.

ADHD plus reading difficulty is not the compounding disaster we assumed. Marks et al. found no evidence for an additive impact of co-occurring ADHD and reading disabilities. If this holds, schools may be over-differentiating their reading interventions based on diagnosis rather than the reading problem itself.

AI dependency runs through two routes, not one. Tian and Zhang’s dual-pathway model distinguishes an attentional route (engaged, evaluative use of AI; beneficial) from a reliance route (offloading cognitive work, harmful). Same tool, opposite effects, depending on the student’s stance.

Metacognitive instruction narrows the SES gap — with caveats. Maximino-Pinheiro et al. found that teacher PD focused on metacognition reduced the relationship between socioeconomic status and children’s metacognitive skills in early years. Hopeful, but 44 teachers is a small sample and this needs replication at scale.

Science of Learning in the news

“My Students Can’t Read”

Tyler Jagt’s essay in the Chronicle of Higher Education is worth a read, it’s sort of the university-level version of the literacy crisis that primary and secondary educators have been documenting for years. Jagt presents data on the collapse of student reading capacity at the undergraduate level. Paired with the growing evidence on AI-assisted cognitive offloading, this is a sobering portrait of where things stand.

“How YouTube Took Over the American Classroom”

Shalini Ramachandran's WSJ investigation is the most thorough piece of reporting on classroom screen use published this year. The headline number: a seventh-grader in Wichita accessed 13,000 YouTube videos during school hours in three months. A second-grader in New York watched 700+ in two months. YouTube now accounts for roughly half of all student traffic on school devices, and 60% of the K-12 mobile device market runs on Chromebooks,which is not an accident. Internal Google documents show the company explicitly targeted K-12 as an entry point for "lifelong brand loyalty" and identified children under 13 as the "world's fastest-growing internet audience."

“Children too easily diagnosed with ADHD and autism, GPs say”

Louise Eccles and Yennah Smart’s Sunday Times piece reports on a Centre for Social Justice survey of 1,000 GPs. The headline numbers: 75% agreed that clinical boundaries for ADHD and autism have expanded to include behaviours previously considered normal range. 66% said diagnoses are given out “too easily” where behavioural interventions would be more appropriate. 57% said financial entitlements (disability living allowance claims doubled from 420,000 to 900,000 children between 2016 and last year) strongly influence parental requests for assessment.

“OECD Digital Education Outlook 2026”

The OECD’s flagship education technology report is out. The headline finding comes from a field experiment in Türkiye: students given GPT-4 access improved short-term performance by 48%, but performed 17% worse once access was removed. The report distinguishes “fast AI”, namely cognitive outsourcing that boosts output while hollowing out learning, from pedagogically designed AI that keeps the student doing the thinking. 72% of teachers worry about academic integrity; only 37% of lower secondary teachers have used AI at all. This is sort of an institutional-level version of Bjork’s performance-versus-learning dissociation.

Deans for Impact releases second edition of The Science of Learning

Deans for Impact published the second edition of their Science of Learning guide on 19 May 2026. The original was one of the most widely shared resources in evidence-informed education. The update incorporates research from the last decade, including newer work on retrieval practice boundary conditions, cognitive load measurement, and self-regulated learning. Free to download, and worth sharing with anyone entering the profession.

“From the Teacher’s Desk: A Science of Reading Progress Report”

The Fordham Institute surveyed over 1,200 K-3 teachers and the results are a mixed report card for the science of reading movement. The good news: 82% have completed SoR-aligned training, Fountas & Pinnell usage has collapsed from 43% to 16%, and UFLI Foundations is now the most-used curriculum at 38%. The bad news: 30% of teachers still believe phonics and cueing are equally valid strategies, and 58% believe reading comprehension relies on transferable “skills” rather than background knowledge. Work to be done.

“New York promised to center phonics in reading instruction, but new course for teachers sidelines science of reading”

In related news, the Hechinger Report’s investigation into New York State’s $10 million phonics training programme is an implementation failure case study. The state gave the money to the teachers’ union to build SoR training for 20,000 teachers. The resulting course promotes balanced literacy, includes three-cueing, contains factual inaccuracies, and misrepresents researchers’ work. Meanwhile, 41% of New York fourth-graders score at the lowest NAEP level. Read alongside the Fordham survey above, one shows the movement’s progress, this one shows how easily that progress can be captured and diluted.

Housekeeping

I was featured in this New York Times article about motivation.

Paul, Jim and I spoke about our book Instructional Illusions with the great in a conversation with AEI.

I was interviewed for this article on AI in the TES.

I will be speaking at researchED in Houston on Sat 13th June.

Thank you for reading. If you’re new here, The Learning Dispatch is read by over 23,000 educators worldwide. If you find this useful, please consider sharing it with a colleague.

If you missed last month’s dispatch, you can find it here.

Until next month,
Carl

Two centuries later, evolutionary biology developed its own Panglossian paradigm: the assumption that every feature of the body and brain must have been engineered for some ideal purpose. However a new paper by Kayson Fakhar and Duncan Astle argues that the assumption is doing real damage in cognitive neuroscience and in the AI research that draws on it. The brain is absolutely not the best of all possible cognitive “worlds”. It is the one our ancestors happened to inherit, with the wiring back to front and the optimisation half done.

For me, this paper has underlined something quite profound about cognition and captures something I have been thinking about since the emergence of LLMs; the idea that if you set out to design our brains from scratch today, you would do a hell of a lot better than the one reading this sentence.

Not the best of all possible brains

There is a widely held view that the human brain is the most remarkable object in nature. The astronomer Owen Gingerich called it “by far the most complex physical object known to us in the entire cosmos”, a view which is often followed with a religious sense of divine creation. Gingerich’s line comes from a book in which the brain’s complexity is offered as evidence of a designer behind the cosmos, and the broader claim that the brain is too marvellous to be the product of blind chance runs straight back to the kind of thinking Voltaire was satirising in Candide. Pangloss’s “best of all possible worlds” was a claim about God’s design before it was a punchline.

The reality is that our biology is far from optimal. It is a stack of compromises, and the brain is only one of them. If you want a concrete case, start with the eye. The eye is the part of the body we most often hold up as the textbook case of evolved engineering, (complex enough that Darwin himself worried it might be too complex for natural selection to account for). If you trace what the light actually has to do and the case for elegant design starts to fall apart at the first step.

Light enters the human eye, hits the cornea, passes through the lens, and then runs into a problem. Before it reaches the photoreceptors that detect it, the light has to push through a tangle of blood vessels and nerve fibres lying on top of the retina. The cells that do the seeing face the wrong way. The optic nerve punches a blind spot straight through the middle of the picture, and the brain quietly fills it in. (By the way, cephalopods, who evolved eyes on a separate evolutionary line, ended up with the wiring the right way round.) This is not particularly clever engineering, it just happens to be the engineering that won because our alternative ancestors died first

What the brain does well is occupy a useful position on a high dimensional surface of competing demands, with that position held in place by deep evolutionary history. Some of its features are solutions to problems we care about such as language, theory of mind, or the dopaminergic learning signal that lets you update beliefs from experience. Others are spandrels, by‑products of constraints that mattered to a fish four hundred million years ago but still shape how our cortex is wired today.

The concept as applied to cognition comes from a 1979 paper by Stephen Jay Gould and Richard Lewontin, "The Spandrels of San Marco and the Panglossian Paradigm: A Critique of the Adaptationist Programme", one of the most-cited papers in evolutionary biology and the source from which I am borrowing the title.

Spandrels of the mind

A spandrel, in the architectural original, is the tapering triangular space you get when two rounded arches meet at right angles, or when an arch meets the wall above it. You cannot mount a dome on rounded arches without producing them. They exist because of geometry, not because anyone wanted them. The four spandrels under the central dome of St Mark’s in Venice each contain an elaborate mosaic, an evangelist above and one of the four Biblical rivers below. Standing in the church, you are tempted to think the architecture was designed to house these images. Gould and Lewontin’s point is that the causation runs the other way. The mosaics are decoration of inevitable space. The spaces themselves are a by-product of putting a dome on arches.

Many features of organisms exist not because they were selected for, but as necessary by-products of selection for something else, or as consequences of developmental or architectural constraints that have nothing to do with the trait’s current use. The example they like best is Tyrannosaurus’s tiny front legs, almost certainly not “for” anything in particular but a developmental correlate of selection on head and hindlimb size. The human chin is another: not a trait in itself but a leftover of the interaction between two growth fields. The whole point is that asking “what is this for?” is the wrong question for a spandrel. The answer is “nothing; the structure that is for something else made it inevitable.”

So there are biological spandrels but also cognitive spandrels which are far less benign. Half a century of work by cognitive scientists like Kahneman, Tversky and others has produced something close to a complete catalogue of the human brain's failure modes: confirmation bias, anchoring, the availability heuristic, base rate neglect, the conjunction fallacy, sunk-cost reasoning, the planning fallacy, hindsight bias, and a long list besides.

Memory is no more dependable. Elizabeth Loftus has spent her career demonstrating that what feels like a recording of an event is in fact a reconstruction reassembled each time it is recalled, and that the reconstruction is trivially easy to corrupt. None of this is a defect in any individual brain. It is the default architecture, present in every brain ever built. The system that produced this sentence is the same system that systematically misjudges probabilities, misremembers its own past, and is largely convinced of its own competence at all of the above. Whatever else it is, it is not optimal.

And crucially, a core characteristic of human cognitive bias is that we think we are special. The Lake Wobegon effect, where most people rate themselves above average on almost every desirable trait, is the individual version. The brain does not just fail; it fails confidently. Ask people to explain how a zipper or a toilet works and they will report a strong sense that they understand both, until they actually try to explain it, at which point the confidence collapses. Rozenblit and Keil called this the illusion of explanatory depth, and it applies to almost everything we claim to understand, including the system between our ears.

Nisbett and Wilson showed in the 1970s that we do not have reliable introspective access to why we do what we do; we just confabulate plausible-sounding explanations and feel confident about them. Robert Bjork's work on the illusion of fluency makes the same point in the domain that matters most for teachers: students who re-read a passage feel they have mastered it, when that very fluency is the worst predictor of whether they have actually learned anything. The brain, in other words, is the kind of system that produces systematic errors and then assures the user that it has not. If you wanted to build a cognitive system that could not see its own faults, you would build something rather like us.

Why AI shouldn't think like us

To me then, one of the strangest criticisms of AI is the idea that it’s ‘not human’ or doesn’t do things a human does. For anyone who has studied the brain’s hundred‑and‑eighty‑odd cognitive biases, the demand that AI be more like us is almost perverse; the last thing you would want from a safety‑critical system is that it think like a human. The human brain is the system that produced rampant confirmation bias, anchoring, the availability heuristic, motivated reasoning, base-rate neglect, the conjunction fallacy, sunk-cost reasoning, the planning fallacy, and all of the rest of it. Half a century of cognitive psychology has shown us how the human mind fails at exactly the kinds of reasoning we would ideally want a cognitive system to do.

If we were to design self-driving cars to be “more human”, we would need to programme them to kill an extra 37,000 Americans, and over a million people worldwide, that are currently attributed to human error every year. We do not ask self-driving cars to be ‘more human’ and share the human driver’s 250-millisecond reaction time, 200-degree visual field, or tendency to look at the wrong thing at the wrong moment. We ask them to do the opposite.

That is the context in which complaints that ‘AI isn’t really intelligent because it doesn’t think like us’ should be read. If anything, the direction of travel in AI and instructional design should be away from human‑like cognition, not towards it. The question is not whether systems like LLMs pass some folk‑psychological Turing test of sounding human, but whether their pattern of errors, biases and blind spots is less correlated with ours, so that human and machine can catch each other’s mistakes rather than amplify them. On that view, insisting that AI must become more human before we trust it has things almost exactly backwards.

Now obviously when it comes to deeper ethical and philosophical questions, it’s a differnt story. There are domains where what we actually want is human judgement: questions about what matters, about how to live, about what counts as a fair outcome, about the texture of a particular relationship or a particular pupil’s experience. These are not problems to be solved by colder, cleaner cognition; they are problems the colder, cleaner cognition has no native standing to solve. An AI that bypasses human bias in a probability calculation is doing what we want but an AI that bypasses human judgement in deciding what the school is for is making a category error in the other direction.

“Beautiful and deserved to be true”

Gould and Lewontin tell a story about Herbert Spencer that I think is just fantastic. Spencer, the Victorian thinker who coined the widely misunderstood term "survival of the fittest", had a habit of finding adaptive purpose everywhere he looked. He visited Francis Galton's anthropometric laboratory one afternoon to have his fingerprints taken. As they talked, Galton mentioned the long-running puzzle of how fingerprint ridges develop, and that researchers had been dissecting the fingers of unborn children to find out. Spencer told him this was the wrong question. The right question, he said, was what the ridges were for; once you knew that, the developmental story would follow.

He then produced, at length, such an elegant explanation: the ridges must exist to protect the delicate mouths of the sweat glands, which would naturally sit safely in the valleys between them. Galton's reply is wonderful on so many levels:

I love this because the retina and the fingerprints are the same just-so story. In each case, a clever story about optimal design collapses as soon as you look at the actual wiring. Biology is full of structures that invite a purpose‑driven explanation and then refuse to co‑operate when you check the anatomy. The danger is not that the stories are ugly, but that they, as Galton put it, are ‘beautiful and deserved to be true’ and we believe them anyway. Together they bracket what goes wrong when biology is read as engineering, and the same failure mode shows up wherever we treat an evolved system as if it were a designed one, in neuroscience, in AI research, and most damagingly in education.

The NeuroBollocks of “Brain-based” learning

Education has fallen for this trick more credulously than almost any other field. For three decades it has imported claim after claim about “how the brain learns” that turn out, under inspection, to be Herbert Spencer on fingerprints with a flashing fMRI picture of the brain attached. Learning styles, left-brain versus right-brain dominance, hemispheric thinking, the ten per cent myth, multiple intelligences in their classroom-poster versions. Each offered a beautiful explanation of individual differences; each was confidently delivered, often by people charging by the day; and none survives five minutes with the evidence. The brain is not an ideal. It is a compromise. Teaching that treats the brain as the design goal, rather than the design problem, will keep producing elegant stories about ridges and valleys that turn out to run the other way round.

The deeper problem is the prefix itself. “Brain-based learning” always sounded silly to me. All learning is brain-based, in the same trivial sense that all digestion is gut-based and all running is leg-based. The phrase is the ultimate tautology and most of the genre runs on the same conjuring trick: a vague behavioural claim, a brain image, a confident voice, an invoice. “Brain-based teaching”, “brain-friendly classrooms”, “neuroeducation”, “neuroleadership”, “neuromarketing”, “neurolaw”, an entire prefix industry has grown up on the premise that adding “neuro-” to a perfectly ordinary claim transmutes it into a finding. The prefix functions like the word “organic” on a supermarket label.

The empirical case for questioning all this is by now embarrassingly strong. Steven Poole popularised the diagnosis in a 2012 New Statesman essay under the name “neurobollocks”; a brilliant Twitter account of the same name spent years cataloguing examples; Sally Satel and Scott Lilienfeld’s Brainwashed did the long-form version. McCabe and Castel showed in 2008 that people rate the same explanation as more credible when accompanied by an fMRI picture, even an irrelevant one. Weisberg and colleagues showed the same effect for the word “neuroscience” prepended to an otherwise vacuous account. The colours are doing the persuading.

John Bruer made the deeper point a decade earlier, in “Education and the Brain: A Bridge Too Far”: his claim is that almost nothing in neuroscience is at a level of resolution that translates into classroom practice, and the bridge from one to the other is, to a first approximation, made of nothing. A bit harsh I think but certainly the brain scan is the new fingerprint, and the explanations being draped over it are mostly Spencer’s.

Designing instruction for the brain we actually have

One reason why John Sweller’s Cognitive Load Theory made so much sense to me when I first came across it as a teacher. It refused to flatter the brain with the idea that the human mind is the most wonderful thing in creation. A lot of pedagogical theory I had encountered up to that point assumed the learner’s mind was, in some quiet sense, ready for the material and innately primed to discover knowledge and that the job of the teacher was to remove obstacles, “light a spark”, get out of the way. This model of the brain suggested that learning is largely a natural unfolding, something that will happen of its own accord if we do not get in the way too much.

Sweller’s great contribution for me is how he flagged the problem with this in the cognitive architecture itself, and he did so without blaming the student or the teacher. It treated the brain as the design problem rather than the design goal. It was such a relief it was to learn that the reason my students were struggling was not because my lessons were not ‘engaging enough’ but rather the fact that the brain we are teaching was actually not built for thinking hard and certainly not built for discovering knowledge for itself. It has severe shortcomings when it comes to learning complex tasks and domains like reading, writing and Maths. Its working memory and attentional frame are vanishingly small. Almost everything you do that feels intuitively right in the learning process, is often the worst predictor of long term learning.

Learning from a different kind of mind

So in terms of instructional design, the question is not whether systems like LLMs pass some folk-psychological Turing test of sounding human, but whether their pattern of errors, biases and blind spots is uncorrelated enough with ours that human and machine can catch each other’s mistakes rather than amplify them. On that view, insisting that AI must become more human before we trust it has things almost exactly backwards for me.

One reason I am interested in, and even excited by, the development of AI is that we can learn more about not just how learning happens but about the instructional design that might optimise for it. For the first time we have large, tractable cognitive systems whose internals we can inspect, perturb and redesign, without having to wait for a few million years of evolution or a new cohort of undergraduates. We can see what happens to performance when we change the architecture, relax a constraint, or give the system a memory capacity no human could ever have. We can potentially pick up a much better signal around retrieval, spacing and interleaving that is almost impossible to do in real classrooms.

That opens up a question education has never been able to ask in a serious way: what would teaching look like if the student were not the only cognitive system in the room. The student is still a human being with the attendant human architecture; limited working memory, attention still fragile, biases still in place etc. We cannot redesign that. What we can do, for the first time in the history of schooling, is pair the student with another kind of cognitive system whose bottlenecks fall in different places. An AI does not need to learn like a human any more than a calculator needs to count on its fingers, and that is precisely the point. Large language models are not interesting because they mimic us. They are interesting because they give us a working counterexample, another way of turning exposure into competence under a different set of constraints. The things our brains find hardest are often the things the model finds easiest. And vice versa. Spaced retrieval, the generation of worked examples at scale, patient repetition without irritation: none of these are useful to the student because the AI is more humanlike; they are useful because it is less so.

That counterexample cuts both ways. It highlights where our cognitive architecture is astonishingly efficient, for example in how quickly children acquire spoken language from sparse, noisy input, and where it is painfully bad, for example at holding intermediate steps of an algebraic manipulation in mind. It also forces us to be more precise about what we mean by “better learning”. Do we mean faster acquisition, greater robustness to distraction, better transfer to new problems, deeper structural understanding? AI gives us knobs to turn on all of those dimensions, and then invites the awkward question: if this is what a different mind can do with well-designed input, what excuse do we have for curricula and pedagogies that still behave as if the brain were the best of all possible cognitive worlds.

AI tutors that exploit the asymmetry

The same logic carries straight across to the rapidly expanding question of AI in classrooms. The dominant frame is that the AI tutor should be as humanlike as possible, modelled on a patient adult who knows the material, responsive in the way a good teacher is responsive. Some of this is sensible; much of it is the same category error in a new disguise. An AI tutor is not a person and is not a brain. Its constraint space is wildly different. It has no working memory limit in the human sense. It does not get tired. It does not forget systematically. It does not have its own developmental trajectory.

The right question for me is not how we make the AI more like the brain. It is what this new system, with its own quite different trade-offs, can do that the human brain cannot, and how we marry the two. Working memory bottlenecks are an evolved compromise, not a target to engineer towards. The fact that the AI does not share them is the opportunity. Spaced retrieval is harder for the human brain than for the model; that is exactly why the AI can scaffold it. Generation of worked examples at scale is trivial for the model and impossible for the classroom teacher; that is the lever.

The early experimental evidence on AI tutoring is clearest when the systems lean into this asymmetry rather than play at being a second, slightly worse human in the room. In a recent randomised controlled trial with 194 undergraduate physics students at Harvard, a carefully designed AI tutor outperformed an already refined active-learning class, despite students spending about twenty per cent less time on task and reporting higher engagement and motivation. Students using the AI tutor achieved roughly twice the learning gains of those in the classroom condition, with effect sizes in the range educational researchers almost never see. A small UK trial of a pedagogy-tuned maths tutor in secondary schools found a similar pattern. Eedi’s constrained AI tutor only activated when a student answered a diagnostic question incorrectly, and its messages were limited to the specific misconception the student held, with a human tutor in the loop. The human-AI team produced around a ten percentage-point gain on a key learning metric compared with 4.5 points for a human tutor alone; the AI effectively doubled the tutor’s impact. The gains were modest rather than miraculous, and they came precisely because the system exploited its differences from the human brain, not because it did a good impersonation of a chatty adult.

Education has been worse than most fields at this. We have spent a long time treating biology as both the diagnosis and the cure. Brain scans get invoked to justify pedagogies that the cognitive evidence does not support. The notion that good teaching is what aligns with “how the brain learns” has been the engine of more rubbish than almost any other claim in the field. Fakhar and Astle’s paper offers a useful tonic. Stop treating biology as the gold standard. The brain is not an ideal architecture. It is a working draft, several hundred million years late, made of compromises some of which are still useful and some of which are merely inherited.

What teaching is actually for, on this view, is closer to what it always was: the deliberate construction of cognitive capacities the brain was not built for, using the constraints of the brain as the design problem rather than the goal. The retina faces the wrong way because evolution had no other option. We do have other options. The point of school, and increasingly the point of well-designed AI in school, is to use them.

The Panglossian instinct to make AI more humanlike is, structurally, a demand to reproduce our spandrels. The brain's bottlenecks and biases are not features the cognitive system was designed to have; they are the residue of an evolutionary process that was solving for survival, reproduction and energy efficiency, not for holding the intermediate steps of an algebraic manipulation in mind. The working memory limit is not a target. The catalogue of biases is not a feature set. The Pangloss view of cognition asks what these things are for. The Gouldian view answers, accurately, that mostly they are for nothing; the architecture that is for something else made them inevitable. AI does not share the geometry, so it need not share the spaces. The cognitive partner now sitting in our classrooms is the first one in history that gets to be designed rather than inherited, and the temptation to design it as a slightly worse second human in the room is the same instinct that has us teaching to a brain we imagine is the best of all possible cognitive worlds. It is the same mistake in a new register, and the same correction applies: stop worshipping the brain, use the asymmetry.

Share

Here is a question I find myself asking a lot today: does anyone seriously believe that a degree completed in 2026 is the same diagnostic instrument as one completed in 2016, or indeed, any time before chatGPT showed up in 2023? I was reminded of this at the Hay Festival last Friday, where the novelist and Oxford fellow Katherine Rundell spoke of a depressing survey a colleague had given her of student reading at his university. Where 70 per cent of students used to do half of the course reading, now 20 per cent do 10 per cent of the work.

She called what is happening a “vast counterfeiting of knowledge”, and accused universities of raising a “white flag” by reclassifying AI use, provided it is footnoted, as legitimate scholarship. “It is now possible,” she said, “to write an essay about a book without having read the book.” This week, Glasgow announced a return to in-person examinations, and the same Times investigation reports that the number of Russell Group undergraduates formally disciplined for AI misuse tripled in a single academic year, from roughly 700 to more than 2,000. What then is the value of a degree as a measure of intellectual worth in 2026?

The signalling value of the university degree as a credential, on Spence's old account, depended on a separating equilibrium: obtaining the degree was costly, and more costly for those without the underlying ability. If the production conditions have changed, the signal it carries has changed too. The work that undergraduates do during a degree; coursework, problem sets, take-home essays, unsupervised exam preparation, is precisely what has been most affected by AI substitution. A 2016 graduate accumulated knowledge under a regime in which finishing the work and learning the material were tightly coupled. A 2026 graduate has accumulated, in part, a track record of finished work whose coupling with learning is now an open question.

Employers have noticed, and the labour-market data has begun to catch up. Alan Milburn's interim review for the UK government, published this week, finds 946,000 sixteen-to-twenty-four-year-olds not in education, employment or training. More than half have never worked. Youth unemployment is at an eleven-year high of 16.2 per cent. Thousands of them hold university degrees. Milburn frames the problem in terms of smartphones and mental health, the "bedroom generation" rewired by years of scrolling, and he is not wrong but this is not the full story. AI is the second arm of the pincer: students whose degrees are increasingly composed of work they did not do, arriving in a labour market that has already begun to discount the certificate they are carrying. A senior New York financier told the Financial Times's Gillian Tett that his firm now combs through humanities students rather than "AI-native" STEM graduates, on the grounds that the latter arrive with "alarmingly shallow ideas".

But weren’t there always students who cheated? The within-cohort variance is also very different now. There were always students who coasted, but the coasting had a ceiling set by the cost of the alternative routes to a finished assignment. That ceiling is now much higher, and there is no way to tell from a transcript which side of it any given graduate sat on. The credential no longer reliably differentiates between students who did the work and students who did not, on precisely the dimensions where AI substitution is easiest.

Faster Completion, Less Learning

A core part of doing a degree, especially in the humanities, is the coursework essay or project. I actually loved this. I have very fond memories of going to Senate House Library in London and spending a whole day reading primary texts and secondary scholarship on a particular topic, following one footnote into another author and that author into a third, accumulating context the assignment did not strictly require but the question itself seemed to. The reading was not efficient. It was, on most measures, wasteful: hours spent following arguments that turned out not to bear on the essay I was supposed to be writing; hours spent arguing in my notebook with critics who had stopped publishing decades earlier. But somewhere in that wandering, the shape of an idea would form, usually not the idea I had set out to find. The essay was the artefact at the end, but the argument it created had been built in the wandering.

The coursework proxy held up for a long time, partly because the alternative routes to a finished task were narrow. A pupil could copy from a friend, plagiarise, or even pay for an essay, but each had a friction cost and a reasonably detectable signature. The signal was noisy at the margins but generally sound for broader purposes. The relationship between completion and learning was stable enough that we built an entire infrastructure of assessment on top of it.

However, a new preprint by Sina Rismanchian and colleagues, drawing on 3.2 million ALEKS learning interactions across a ten-year panel, suggests that the proxy has now broken. The title puts it about as plainly as a title can: Faster Completion, Less Learning. The paper shows that time spent on AI-susceptible word problems has fallen by roughly 27% among college students and 31% among high-schoolers since ChatGPT’s release in late 2022.

Crucially, under proctored conditions, where AI access is restricted, the divergence vanishes entirely. When those same students are later tested on what they supposedly learned, their odds of answering correctly drop by 25%. Meanwhile, in unproctored assessments, performance rises sharply. The pattern is unambiguous: students are using AI to generate answers, platforms are recording completion, and durable knowledge is evaporating.

Two earlier randomised trials make the same case in tighter laboratory conditions. Bastani and colleagues ran a field experiment in a Turkish high school: GPT-4 access during maths practice raised assisted performance by 48%, but unassisted exam scores fell by 17% against control. Barcaui’s RCT with undergraduates found the same shape: a surprise retention test 45 days after study produced 68.5% correct in the traditional condition and 57.5% in the ChatGPT condition (d = 0.68), and the gap survived an adjustment for time on task. Rismanchian’s contribution is not the discovery of the effect. It is the demonstration that the laboratory pattern is now operating in the field at the scale of millions of students over years.

A separate study published in Science this month documents the scale. Kizilcec, Chirikov and Smirnov surveyed over 95,000 students across twenty US research universities and found that 9% admitted to using generative AI to cheat, a figure that rises to more than one in four among daily AI users. These are self-reported numbers gathered through a list-randomisation method designed to protect respondents; the true rate is almost certainly higher. When a quarter of your heaviest users are submitting AI-generated work as their own, the contamination of behavioural signals that Rismanchian’s team identified is not a theoretical concern. It is the baseline condition across American higher education.

Cognitive Surrender, Not Cognitive Offloading

One argument I often hear is that AI is not going to change much, and that we have seen this before with calculators and the internet. But what students are doing with generative AI is qualitatively different from what they do with calculators, spreadsheets, or search engines. Classical cognitive offloading, delegating a well-defined subtask like arithmetic to a calculator, leaves the reasoning process in the learner’s hands. The student who uses a calculator to perform a multiplication has offloaded the arithmetic but still constructs the solution strategy, evaluates the result, and integrates it with prior knowledge, the active learning that produces durable retention.

The thing about someone who cheats is that they are at least aware of the domain of knowledge that was there. In Rumsfeldian terms, they know what they don’t know. Worse still for me is the idea that students will not even know what they don’t know. A student who copies a word problem into a chatbot and pastes the returned solution bypasses the whole metacognitive process. A recent paper calls the pattern cognitive surrender: the adoption of an AI-generated output as one’s own answer with minimal scrutiny, in which the user relinquishes cognitive control of problem-solving rather than merely delegating a discrete subtask. The behavioural signature documented at population scale, selective, gradually growing, supervision-sensitive, age-graded, present where students have autonomy and absent where they do not, is exactly what cognitive surrender predicts when it operates across millions of interactions.

Fan and colleagues call the underlying disposition metacognitive laziness. In their randomised study, students using ChatGPT to write essays produced better essays during the task but showed no advantage on knowledge gain or transfer, and the process data revealed measurably fewer evaluation and orientation moves. Surrender names the act; laziness names the disposition that produces it. They are the same phenomenon described at different levels.

The downstream consequence is what Barcaui calls borrowed competence. AI supplies structure, vocabulary and reasoning scaffolds that inflate the feeling of mastery during study without strengthening memory traces through retrieval or learner-generated elaboration. Even experienced AI users showed no better retention in the Brazilian trial; familiarity with the tool did not translate into more effective use for learning. Students confuse the AI’s fluency with their own understanding, a metacognitive blindness that prevents them from recognising when assistance is undermining their learning.

The Illusion of Mastery

So what is the effect of chatbots on how students think they learn? The metacognitive dimension is important here I think. Students in both the Turkish and Brazilian trials reported that they believed they had learned effectively, even when objective measures showed the opposite. In the Turkish trial, students’ self-reported perceptions of exam performance and learning were “overly optimistic”. In other words, those in the unguided GPT condition believed they had learned as much as the control group, despite performing 17% worse on the actual exam. This mismatch between perceived and actual learning has been observed in other educational settings, but generative AI appears to intensify it.

So what is happening here is that during the practice session, students feel competent because they are producing correct answers. The AI handles the cognitive labor, but the student experiences the outcome as their own achievement. When the tool is removed, the illusion collapses. But by then, the opportunity for genuine learning has passed.

This has immediate implications for adaptive learning systems. Platforms like ALEKS calibrate their mastery estimates on the assumption that student responses reflect genuine cognitive effort. When students bypass this process using AI, the platform’s diagnostic inferences become unreliable; the system may assign mastery credit for knowledge the student does not actually possess. The 25% cumulative decline in retention odds on randomly assigned items represents the downstream manifestation of this structural disruption.

The Proctoring Paradox

One reason why Edtech has been such an expensive failure is down to a fundamental feature of the human brain: we don’t want to think hard, and usually won’t do it unless the conditions are right. In a classroom, this usually means a teacher who monitors cognitive work and focus. calls this “be seen looking”; the deliberate act of a teacher circulating, making eye contact, checking work, signalling to every student that their effort is visible and valued. It’s not surveillance for compliance; it’s accountability for cognitive engagement. Students work harder, persist longer, and think more deeply when they know their thinking is being seen.

Most edtech removes that signal entirely. A student alone with a screen, unsupervised, working through problems at eleven o'clock at night, is operating in the precise conditions under which the Rismanchian study shows learning collapses. A student alone with a screen has no teacher circulating, no “be seen looking” creating the conditions for effortful thought. And one fascinating revelation from this new study is what it shows about proctoring as a replacement for that live presence.

Students who spent less time on word problems during the learning phase (because they were outsourcing them to AI) perform worse when later tested under proctored conditions on the same material. The same estimator applied to non-proctored retention finds the opposite result: apparent performance rises. Only one mechanism predicts that reversal: AI assistance at the point of assessment, masking a genuine decline in durable knowledge.

The effect also scales with age. College students show the largest decline, high schoolers nearly as much, middle schoolers a smaller but significant drop, and primary students none at all. This is not because younger children are more virtuous; it is because they complete more work under direct supervision, in environments where an adult can see the screen. The age gradient mirrors exactly what decades of academic-integrity research documented about dishonesty rates before generative AI existed. AI has not invented the incentive to cheat; it has simply reduced the effort cost to near zero.

Paradoxically, a solution to this AI-created problem might also come from AI. At Alpha School, developers have pioneered an AI proctoring platform that is able to create the kinds of constraints that lead to learning at scale. The system monitors not just for prohibited behaviours such as using chatbots, but for also for distraction and patterns of cognitive engagement such as time on task, latency and tracking whether students are actually thinking through problems or bypassing that effort. It’s an attempt to restore, through algorithmic observation, the accountability structure that classroom presence naturally provides. Seeing it in action last year changed my mind. I had assumed that effective learning at scale required a traditional classroom; Alpha's system suggests a credible alternative.

An unwinnable arms race?

has argued convincingly that in the AI era universities must focus on the “last mile” of human expertise that exceeds AI capability. The evidence here suggests a more immediate problem: students are not even reaching the first mile. When undergraduates outsource the foundational cognitive work of a degree to a chatbot, they never build the substrate on which any “last mile” would rest. The diagnostic crisis the data describe is, at root, an architectural one. The completion proxy was load-bearing because we trusted that the artefact had been produced by a particular kind of cognitive activity. That trust is no longer warranted, and the structures we built on top of it are creaking accordingly.

Two responses present themselves immediately, and neither is sufficient. The first is to design tasks AI cannot do for you: graph manipulation, interactive simulation, multimodal reasoning that depends on visual-spatial structure. But every task we certify as AI-resistant today is provisional, on a clock set by the labs’ release schedule. We end up redefining authentic cognitive work as whatever the latest model cannot yet do, and redefining it again the next time it can. That is no way to write a curriculum. The second is to retreat to proctored, time-limited, human-observed assessment, the conditions under which the behavioural signal returns. But universal proctoring is impractical, expensive, and falls hardest on under-resourced institutions. It is also a measurement fix, not a learning fix: it can tell us whether learning happened, but does little to make learning happen.

The third response is the one most consistently ignored, and almost certainly the most important, because it is the one that does not play the arms race at all. If the problem is that students are skipping the cognitive work because an AI is sitting next to them while they study, the answer is not to police the AI or redesign the task. It is to change the moment at which the AI enters the cognitive process. Wong and Qiu tested a “think first, ChatGPT later” regime in which students had to generate their own solutions before being allowed to consult the model. On the assisted task this group showed no advantage. On a subsequent transfer task with the model withdrawn, the free-AI users had dropped back to the unassisted baseline, while the think-first group outperformed everyone, including the students who had worked alone throughout. The pedagogical instruction is almost embarrassingly simple. AI introduced before the cognitive work has happened prevents the work from happening; AI introduced after the work has begun has a chance to extend it rather than replace it. The lesson is sequencing, not abstention, and it is almost universally ignored.

Knowledge for its own sake

But all this suggests a transactional notion of education that I find depressing to be honest. The point of a degree, on the older view, was never the certificate. It was the slow cultivation of mind that the certificate happened to represent. Newman's Idea of a University, von Humboldt's Bildung, the post-war liberal-arts settlement: each rests on the assumption that what is valuable about higher education is the knowledge acquired in the doing of it, and that the credential is merely the residue of an interior transformation.

But that defence collapses when the learning never happens. And it becomes nearly impossible to sustain when students leave university carrying £50,000 in debt, facing a jobless horizon, clutching a credential the market has already begun to discount. Knowledge for its own sake is a luxury afforded to those who can bear the cost without expecting return. For everyone else, and that is most people, the degree was always a wager: sacrifice now, in time and money, for certification that opens doors later. If the certification fails and the knowledge was never acquired, the wager was simply lost.

However I would prefer to make this defence. It is the one I believe. But I (and my generation) did not have to pay £9000 a year to go to university for an interior transformation that the modern student need not, in fact, undergo, in order to enter a labour market that no longer trusts the document the transformation produces.

The Broken Contract and what I tell my daughters

The deepest irony is that we possess the tools to make learning more effective. For many of us, AI is augmenting learning in a way we never thought possible. Structured AI scaffolding can extend human cognitive capacity without replacing it. The technology is not inherently destructive. But we are deploying it in ways that bypass the very cognitive processes that make learning durable.

The Milburn report focuses on smartphones and social media during adolescence. But the mechanism it identifies, cognitive and behavioural patterns established during formative years that persist into adulthood, points directly to the generative AI question. If smartphones alone can reshape a generation’s capacity for sustained attention and independent problem-solving, what happens when that same cohort spends their university years outsourcing the effortful thinking that builds those capacities to ChatGPT?

I think about this often, because I have three daughters. They are growing up in a world where the social contract we inherited; “work hard in school, earn your degree, build a future” may no longer hold. The credentials we received certified not just domain knowledge but cognitive capacity: the ability to read difficult texts, synthesise arguments, solve complex problems, and produce coherent work under our own power. If they choose to do one, what will their degrees certify?

The deal worked, often enough, for enough people, that an entire civic and economic order was built on the assumption that it would continue to. I am no longer confident it will. I do not yet know what to tell them.

This, I think, is what novels are for. Not, in some abstract sense, to tell us who we are, the way self-help and personality tests pretend to, but something rarer and stranger. To hand back to us experiences we have been unable to articulate ourselves, sometimes for years, sometimes for a lifetime, in the shape of someone else's sentences. C.S. Lewis put it plainly: we read to know we are not alone. Harold Bloom puts it like this:

“We read deeply for varied reasons, most of them familiar: that we cannot know enough people profoundly enough; that we need to know ourselves better; that we require knowledge, not just of self and others, but of the way things are.”

The reason why I think reading novels meant so much to me as a kid was that I was placed inside someone else’s struggles and so my own seemed not only manageable, but understandable by comparison. What saddens me about our current reading culture is not simply that people are reading fewer books. It is that we are increasingly surrounded by forms of text that ask almost nothing of us; summaries of summaries, extracts, feeds, generated prose, convenient fragments. Language stripped of any resistance or ambiguity. Reading remains everywhere, but often in a form that no longer trains attention, patience, or inwardness.

Notice how exact the mechanism is in Odenkirk's case. Balle did not give him a new experience. The experience was already his. What she gave him was the language for it, and language, as Heidegger insisted, is the house of being1, the room in which a feeling becomes available to thought. Without the words, the experience drifts in some vestibular space below cognition, still real, still present, but not quite had.

The experience happens to you, but you cannot inhabit it. At its best, I think the novel restores the experience by at least naming it, and in naming it, gives you back to yourself. This is why Proust, in the line I borrowed at the end of a previous post, called the truest books "mirrors": not because we see the writer in them, but because, with luck, we see ourselves.

There is a reason I think why this kind of recognition is hard to come by elsewhere. I am in Belfast this week and I am reminded of a passage from a poem by Seamus Heaney which many years after my mother died, suddenly made me feel the grief I had been unable to feel.

It has also happened to me watching certain films but not in the same way as it does when you are buried in a novel. The novel requires sustained, unhurried, recursive contact with a sensibility not your own; the kind of contact that is not sadly all too rare (in my most despairing moments, I worry that my daughters will not have the same formative experiences I had with reading.) Only the novel runs long enough, and slow enough, and intimate enough, to let your inner self time to coincide with another's.

Which brings me, uncomfortably, back to Ultra-Processed Minds. I argued there that we are losing not reading itself but a particular kind of reading: the slow, vertical, devotional sort that allows meaning to accumulate, the sort that Maryanne Wolf calls cognitive patience and Sven Birkerts called an "engagement with thought that deepens it." I worried about what this loss does to attention, to civic life, to interiority. I want now to add something more specific, because the Odenkirk clip clarifies it.

This, to my mind, is the most underrated cost of what is happening to reading now. Not the loss of a certain kind of expression or erudition, (which was always patchily distributed anyway). Not the loss of a “preparation for the workplace”. It is the loss of a specific form of grace: the kind in which a stranger's sentences arrive carrying the precise words you had been missing for what you yourself had felt.

In listening to Odernkirk talk about his experience of facing death, I was reminded of Kafka’s observation that “the meaning of life is that it stops”. It gave me such hope to hear him say this and that the novel still has this power to help us understand ourselves in 2026.

And yet the research we have to make sense of this was conducted on a learner who has since been replaced. Rich Mayer’s seminal work on multimedia theory is some of the most carefully replicated research in educational psychology. What has changed however, are the conditions in which those primitives now operate. The principles still hold; but the boundary conditions under which they were established are no longer the conditions students are in.1

That leaves educators, parents and designers in an unusual position: relying on a research base whose validity has visibly decayed, and where the boundary conditions of learning have dramatically shifted on two axes; student behaviour + video format. Then there is the wildcard of AI, which is less a third axis than an accelerant on the first two: changing how students engage with video, and changing what video itself is. How should we design learning with video in mind and what should parents need to know?

Things have changed much more than we think

Last week, the Wall Street Journal published an investigation into YouTube in American classrooms which is shocking on many levels. One mother in Wichita noticed her seventh-grader had developed a surprisingly detailed knowledge of a video game she’d banned at home. When she checked his school Google account, she discovered he had watched more than 13,000 YouTube videos on his school Chromebook in three months, many of them violent, some sexually suggestive.

This is not an edge case. According to the Journal’s reporting, YouTube is sometimes half of all student web traffic on school devices. A survey cited by YouTube executives found that 94% of teachers have used the platform in their roles. More than 88% of US public schools have issued 1:1 devices; Chromebooks (primed for YouTube by the way) hold around 60% of the K-12 mobile device market.

What none of these figures captures is what is actually happening cognitively while the watching is going on. To answer that, two shifts have to be in view: 1.) how students now watch, and 2.) how the medium itself has changed. I’ve been researching this for months now, trying to figure out how to apply the science of learning to a media ecology that did not exist when most of that science was done. What follows is a long post, an attempt to set out what we still know, what we don’t, and what educators, parents and designers should do in the meantime.

What we now know about how students watch video

On the student side, video learning is now embedded in a thicket of habits that simply were not part of the laboratory paradigm from 30 years ago and which will seem alien to anyone over a certain age. Speed-watching at 1.5x or 2x is the default rather than the exception. Captions are on by default, even for native speakers, often with one word highlighted in yellow. The phone is usually the second screen, and often the first. Cognitive offloading to AI assistants is a routine reflex, not an event. Students arrive at an instructional video with priors built from TikTok, YouTube, and the algorithmic feed, which means their tolerance threshold for what counts as cognitive load, or, conversely, as boredom.

On the medium side, the format itself has drifted from the narrated PowerPoint that multimedia studies treated as the canonical multimedia stimulus in 2001. The dominant grammar of educational video in 2026 is short-form, jump-cut, performer-on-camera, dense with kinetic typography, and increasingly AI-generated where convenient. Many videos are now delivered by avatars who do not exist. The decisions about face, pace, length and ornament are being made by content creators raised on a radically different set of YouTube production conventions, rather than instructional designers raised on cognitive load theory.

Need for speed

Begin with speed. Speed-watching is the most empirically settled of the new behaviours, and the picture is more interesting than the simple “comprehension is preserved at 2x” headline would suggest. Murphy and colleagues’ 2022 work at UCLA established the boundary: comprehension is preserved at 1.5x and 2x, with measurable degradation appearing at 2.5x and a steep drop beyond that. Their survey of UCLA undergraduates put the proportion of speed-watchers at 85% several years ago, and that figure has now ossified into a generational default.

Three findings from the last eighteen months complicate the simple story. A 2026 Frontiers paper shows that perceived clarity drops before objective comprehension does: students believe they are losing information at 1.5x even when they are not, and respond by speeding up further to compensate, which is precisely when comprehension begins to fail. Ahn and colleagues, working in Applied Cognitive Psychology in 2025, find that face-on-screen confers no benefit at higher playback speeds and may compete for attention when speech is already compressed. This is one of the few studies that addresses the boundary-condition question directly, and it is not encouraging: the instructor-presence effect that Mayer’s tradition documented may simply not survive the playback conditions in which video is now consumed.

A third paper, from Computers in Human Behavior in early 2025, finds that even where comprehension is preserved, metacognitive monitoring degrades. Students do not know what they are missing. The implication is that speed-watching is fine for material the student already partly knows, and particularly poor for material at the edge of their schema, which is the material instructional video is most often used for.

Captions are now the default

Then there is the caption layer. To my knowledge no published study has tested caption-on viewing during instructional video specifically, but the underlying behavioural data is unambiguous. Stagetext’s UK survey reports that 80% of 18-to-24-year-olds use subtitles “some or all of the time.” Preply puts Gen Z subtitle usage at 70%, against 38% for boomers. Netflix has 80% of its members regularly using subtitles or captions. The reasons go well beyond accessibility: muddled audio mixing in modern productions, accent comprehension, viewing in noisy environments, and the dual-channel reading-while-watching habit conditioned by TikTok, where captions are baked into the medium rather than offered as an option.

This intersects awkwardly with the redundancy principle, which warns against on-screen text duplicating speech under conditions of split attention. Internal reporting from the University of Bristol’s Digital Education Office on the accuracy of automatically generated captions for online learning materials suggests something subtler than the principle would predict: students may now process captions as part of the audio channel rather than a competing visual one. The boundary condition for the redundancy effect appears to have shifted under generational habituation. This is one of the cleaner candidate areas for new empirical work, because it is no longer obvious that what we currently call the redundancy principle is testing what it was originally designed to test.

Worth noting that the redundancy principle was never one of Mayer’s strong findings. In his 2024 retrospective, it sits at d = 0.10 across 12 tests, among the weakest in the catalogue. The principle is being asked to do work in this argument that it never did particularly strongly even in the laboratory conditions in which it was first established. That makes the captions-on shift more interesting, not less: a weak principle whose weak evidence base now needs re-testing under conditions in which the very channel architecture it assumes has changed.

The second screen

The third behavioural drift is the second screen. Industry data has 95% of consumers holding their smartphones while watching television, and 81% of students multitasking while doing homework. The educational research base is more cautious but consistent: multitasking during instructional content depresses transfer (the ability to apply what was watched to new situations) more than it depresses factual recall. Students can still tell you what was said. They cannot do anything with it. The Information Systems Research paper “From Smartphones to Smart Students” finds the now-canonical pattern, namely that any concurrent device use during instruction reduces comprehension, with a larger effect on higher-order outcomes. A more uncomfortable finding from the qualitative literature on Gen Z multitasking is that students do not experience the second screen as multitasking at all. They experience it as the floor of cognitive engagement, with single-screen attention feeling under-stimulated. That is a critical observation, because design decisions that assume single-screen attention as the default are now designing for a learner who does not exist.

With what we know about cognitive architecture, it’s hard to see how meaningful learning can happen under these conditions. But the idea that we don’t learn from video, or indeed AI-generated video, is a myth. Allow me a slight personal digression…

The rich history of educational video

Twenty years ago I watched Kenneth Clark stand in front of Chartres Cathedral for thirteen hours. Civilisation: A Personal View had been broadcast a quarter-century before I encountered it on a BBC repeat, but the series completely fascinated me, and changed how I thought about Western art, history and the long sweep of European intellectual life in ways I am still drawing on.

A few years later, reading American literature for my undergrad degree, the thing that catalysed my understanding of how modern America was formed, more than almost any seminar, was Ken Burns’s The West. Eight hours across nine episodes, voiceover narration over still photographs, primary sources read aloud, Peter Coyote as the unobtrusive through-line. I came back to it repeatedly during the degree and after, and it did more for my sense of the geographical, racial, and cultural substrate of the literature than almost any classroom encounter. Later I watched The Civil War and was utterly entranced by Shelby Foote and his slow Mississippi drawl, which made the war feel less researched than personally remembered.

These videos brought the past to life for me in ways inseparable from how they were made: long, slow, deliberate, by people whose authority over the material had been hard-won and earned across decades.

Clark was a cultural tourist guide. A mind moving through history and reporting back what it found to 1970s British living rooms. But a new incarnation of that cultural tourist guide has emerged, one which makes the additional claim of being educational and so is worth examining here. Chloe vs. History is perhaps the most successful example: a warm, personable young woman who travels through history reacting to what she finds, with over 151,000 subscribers across TikTok and Instagram. But unlike Clark, she is entirely AI-generated and I suspect this is one reason it’s hard to learn anything from her.

Learning has a lot to do with the brain trying to reconcile multiple elements, often competing ones. (This for me is why cognitive load theory is one of the most important frameworks for instructional design). So my hunch is that if the brain is trying to process a human face which it consciously knows is fake but which it instinctively feels is real, this creates a kind of perceptual incongruence that might not be good for learning.

I see this as a kind of uncanny valley, but one with real pedagogical consequences in the sense that every bit of working memory the brain spends reconciling that incongruity is working memory not available for the actual content. If the avatar's face is doing background work to be parsed, the schema-building work the video is meant to support has fewer resources to draw on.

However, we do have examples of AI generated/assisted content which is able to bring the past to life in ways that are truly incredible and genuinely educative. The cleanest example is Peter Jackson's They Shall Not Grow Old (2018). Jackson used machine learning to interpolate frame rates, colourise, and stabilise a hundred-year archive of First World War footage from the Imperial War Museum. He hired forensic lip-readers to verify what the soldiers were actually saying so the audio reconstructions matched the historical record. Foley artists added period-correct sound. The AI was doing the technical restoration; the historical content was the soldiers themselves.

The film succeeded I think, for the same reason Burns's The Civil War succeeded: a careful intelligence at the centre of it, organising real material with patience. It was screened in cinemas, used in schools, and treated seriously by WWI historians. It made a generation that had grown up seeing WWI as a sepia abstraction encounter the actual soldiers, the same age as them, looking back at them.

The AI layer: a mixed picture

Cognitive offloading to AI assistants is the most active growth area in education research, and the 2024 to 2026 record is mixed in a way that is actually clarifying. On the negative side, the PNAS 2025 paper on generative AI without guardrails harming learning in high-school maths is the strongest single piece of evidence that ungated AI use during practice degrades subsequent independent performance. A 2025 RCT in Social Sciences & Humanities Open finds matched effects in adult learners. The recurring pattern across that literature is that the same tool acts as enhancer or crutch depending on how it is used.

Among the positive findings, a 2025 quasi-experimental study in Forum for Linguistic Studies reports that students who delegated lower-order tasks (mechanics, formatting) to AI showed greater improvements in standardised critical-thinking assessments. A Scientific Reports paper on preservice teachers finds the same pattern: strategic offloading correlates with academic gain, passive offloading does not. The single empirical finding most worth carrying into the design conversation comes from a Frontiers paper on differential effects of GPT-based tools: the form of the interaction matters more than the presence of AI itself, and the effect is moderated by prior ability. Low-performing students benefit most from Socratic-style chatbots that engage them in dialogue; high-performing students are harmed most by summary tools that do the cognitive work for them. Embedding "ask the AI" features in instructional video is therefore not a neutral act; the design of the prompt structure, and the cognitive level of the user, jointly determine whether the feature aids learning or substitutes for it.

The newer and more under-researched behaviour is the substitution of AI summaries for engagement with source material. Survey data puts university student AI use at 86–92% depending on geography and definition, with summarising readings the second most common use case after general help with assignments. For instructional video specifically, this matters because the behaviour to design around is no longer “students watch the video,” but rather “students watch the video, ask the AI to summarise it, and revise from the summary.”

I think this could be a fruitful approach, but whether it is net-positive or net-negative depending on prior knowledge has barely been studied. The note-taking literature, as the Learning and Instruction 2026 paper “The mind, not the pen” reaffirms, establishes the pre-AI baseline: deep note-taking instruction substantially improves comprehension because it forces generative processing during viewing. The post-AI commercial landscape (Otter, Coconote, NoteGPT, Mindgrasp, Sembly) is now displacing that activity by producing AI-generated notes from lecture transcripts, (which is almost certainly not going to lead to learning by itself, although there is a world where those transcripts could be turned into learning materials). There is no published comparison of student-generated to AI-generated lecture notes on transfer outcomes (as far as I am aware). That is one of the most consequential gaps in the field.

What the medium has become

The format-side drift is less well evidenced than the behavioural drift, with two exceptions: short-form video, and AI-generated instruction.

The short-form picture has matured fast. The August 2025 medRxiv systematic review The Impact of Short-Form Video Use on Cognitive and Mental Health Outcomes, the 2025 narrative review on short-form video and sustained attention, and the 2026 systematic review on short-video addiction in adolescents converge on the same finding. Heavy short-form video use is associated with measurable decrements in sustained attention, working-memory efficiency and inhibitory control, with effect sizes large enough to persist after controlling for prior cognitive ability and family background.

The neural evidence has filled in. A 2025 Frontiers fNIRS study reports reduced prefrontal activation during cognitively demanding tasks among heavy short-form users. An EEG study by Yan and colleagues (2024) finds reduced midfrontal theta power, a marker of executive function, in heavy users. A 2025 longitudinal paper on screen-media activities and brain function reports altered connectivity in prefrontal cortex, anterior cingulate, amygdala and striatum: the executive and reward circuits both. The standard caveat applies. Most of this evidence is correlational, the causal direction is partly contested (selection effects matter, since adolescents with worse attention may be drawn to short-form video rather than the other way round), and the experimental work has been small-N and acute. But the convergence across designs, populations and modalities is strong enough that prudent policy should not wait for the field to settle the causal question.

There is a parallel, narrower literature on TikTok as an educational tool. The 2025 systematic review TikTok in higher education reports modest learning gains in certain contexts (medical education, language learning, physical education), but the studies are largely satisfaction surveys and self-reported engagement, with little controlled comparison against alternative formats. There is a small evidence base for short-form educational video as engagement-positive, almost none for it being learning-positive in any rigorous outcome sense. Engagement, as the field has long known, is a necessary but not sufficient condition for learning.

However, algorithmic delivery is the format shift the educational research literature has barely engaged with. The communication-and-media literature is more useful here. A 2025 piece on algorithmic recommendations in young people’s everyday lives finds that adolescents experience the algorithm as an environment rather than as a tool, and report a reduced sense of choice over content selection. A UCL / Kent / ASCL 2024 study on TikTok algorithmic exposure reports that after five days of TikTok use the proportion of misogynistic content recommended to test accounts increased four-fold.

The implication for design is that sequencing assumptions baked into traditional curriculum (we will show this clip after that one) have no purchase when the medium is feed-driven. A video can succeed brilliantly as a TikTok and fail entirely as the third unit of a sequenced course, because the medium expects each piece to stand alone.

Then there is AI-generated instructional video, where the empirical base has grown rapidly. Xu and colleagues in the 2025 British Journal of Educational Technology find equivalent cognitive learning outcomes between AI-generated and human-instructor video, with a small advantage for AI on retention and no significant difference on transfer. A 2024 experimental study in Computers & Education reports the same headline finding. A 2025 Frontiers rapid review notes that 72% of students in one experimental group did not recognise that their instructor was artificially generated. The conservative voice is the 2025 Educational Psychology Review piece “Looking Beyond the Hype,” which finds cognitive-outcome equivalence but lower social presence; student preference still tilts toward human instructors.

The implication is significant I think. The instructor-presence literature was built on the premise that the instructor is a human being whose social, emotional and pedagogical signals carry information. AI-generated instructors decouple presence from instructorhood: the face is real to the viewer but not real in any other sense. None of the canonical instructor-presence findings have been re-tested under this condition. The early evidence suggests cognitive-outcome equivalence; the open question is whether the social-emotional effects that originally made instructor presence “matter” survive at all when the instructor does not exist.

Mayer himself, in his 2024 retrospective on CTML, names social, affective, motivational and metacognitive processes as the very dimensions his theory has not yet incorporated, and identifies them as the priority work ahead. The dimensions on which AI-generated instructors most pressure his theory are the dimensions Mayer himself, surveying four decades, concedes have not yet been done.

What about the video as phone view? The vertical 9:16 format has empirical traction in marketing research and almost none in education research. Industry forecasts put video at roughly 70% of mobile traffic and the dominant aspect ratio at 9:16, which suggests that horizontal-versus-vertical viewing is no longer a meaningful design choice but a constraint. There is no controlled comparison of horizontal-versus-vertical learning outcomes on equivalent content. This is the most obvious empirical gap in the field, and one where designers must use principled judgement until evidence appears.

A final format-side observation worth flagging is parasocial engagement. The communication literature on parasocial relationships dates to Horton and Wohl in 1956, and the 2024 to 2026 educational work is thin but suggestive. A 2025 experience-sampling study in Media Psychology reports that parasocial bonds with virtual influencers form quickly and behave similarly to those with human influencers along most engagement dimensions. For instructional video the implication is uncomfortable: the warmth and motivational signals previously attributed to instructor presence may now be functioning parasocially, with the on-screen creator standing in for an actual pedagogical relationship the student does not have. Face-on-screen may “work” in the sense Mayer tested, but for reasons that have nothing to do with pedagogy.

A methodological problem the field has not yet faced

Before I move to design implications, there is an awkward fact about the literature I have just summarised that deserves to be on the table. The most important methods paper to appear on video learning in this window is Gijsen and colleagues in Learning and Instruction (2024), comparing concurrent think-aloud, retrospective think-aloud and cued-retrospective reporting (where the student narrates over a replay of their own eye-tracking trace) for studying processing during instructional video. The finding is uncomfortable but useful: most of what the existing literature claims about processes during viewing (what students attend to, when they integrate, how they segment) is inferred from concurrent or retrospective verbal reports, both of which carry well-understood pathologies. Concurrent think-aloud reactively changes the very processing it measures, and is especially noxious with time-locked stimuli. Retrospective think-aloud suffers omission and post-hoc confabulation. Cued-retrospective reporting threads the needle by anchoring recall to gaze, but is itself still retrospection rather than ground truth.

The implication for everything I have said matters in two ways. When the literature claims “students attend to X” or “learners integrate Y,” the warrant is verbal-report data with the limitations Gijsen documents. Outcome data is the floor; process data is the ceiling, and the ceiling has visible cracks. Second, the field has not, by and large, re-tested its process claims under modern viewing conditions (speed-watching, second screens, captions on, AI assistance), all of which would compound the existing methodological problems. The asymmetry I keep returning to in this post (more behaviour-side evidence, less format-side evidence) is partly a function of the methods available. Behaviour can be studied with surveys and log data. Process during video requires verbal reports, and what we have of those is suspect.

What this means in practice is that we know less than the literature suggests about what students actually do cognitively during instructional video. The gap between confident process claims and the verbal-report data underlying them is wider in 2026 than it was when Mayer’s principles were established.

What to do in the meantime

Taken as a whole, I think the unit of analysis is no longer the video but rather the learning episode. Speed, captions, second screens and AI summarisers all modify the encoding conditions in ways that no scaling of CTML’s principles can fully anticipate. Design decisions made under the assumption of attentive single-screen viewing are designing for a learner who is statistically rare. Treating the video as a self-contained intervention, when in practice it is being consumed inside a noisy multi-screen, multi-tool environment, is the central category error in current educational video design. The video is one element in a wider learning episode, and the surrounding architecture (what comes before, what comes after, what the learner is asked to do during) is what determines whether anything is encoded at all.

The design lever with the most empirical support for protecting learning under these conditions is embedded retrieval and prequestioning. The prequestion literature is now better characterised than at any previous point. Carpenter and Fraundorf’s 2025 multilevel meta-analysis establishes that prequestions reliably enhance learning of targeted content (g = .66) but do not generalise: they do not increase learning of incidental material in the same resource (g = .01). Feedback amplifies the effect substantially. Krebs, Braschoß and Eitel (2024), summarised by The Learning Scientists, show that the benefit of explainer videos only materialises when prompt-questions are provided. Wong and colleagues (2024) examines the timing of low-stakes embedded quizzes in online videos. Huangfu and colleagues in the 2025 British Journal of Educational Technology shows that low-prior-knowledge learners benefit most from prequestions with simple feedback and are harmed by elaborated feedback (a cognitive overload result). Embedded prequestions and retrieval prompts convert passive viewing into active engagement at the lowest design cost, and with the strongest empirical base. If I had to pick one design recommendation to defend against everything I have described in this post, that would be it.

The strongest empirical base is not rhetorical. In Mayer's own 2024 catalogue of fifteen design principles, the generative activity principle sits at d = 0.71 across 44 experimental tests, the largest evidence base for any single principle in the entire CTML edifice. When I recommend retrieval and prequestioning as the design lever to defend against everything in this post, I am reaching for the most replicated finding the theory has produced.

The format drift creates conditions under which the coherence principle is being tested without anyone formally testing it, and (per Gijsen) the methods to test it properly are themselves suspect. The dominant grammar of the medium (fast cuts, kinetic typography, sound design, b-roll) is precisely what Mayer would call seductive detail. The closest direct evidence is a 2025 multi-level meta-analysis in Educational Psychology Review drawing on 177 effect sizes across 50 studies, which finds a small but reliable negative effect of seductive details on learning, mediated primarily by extraneous cognitive load. There is currently no controlled comparison of TikTok-style cadence against narrated-PowerPoint cadence on transfer outcomes. The conservative position, until that comparison exists, is to treat coherence as the principle whose boundary conditions have most clearly shifted, and to make conservative coherence choices in any segment intended to teach novel material. Not because we have direct evidence the seductive-detail effect persists at 2026 cadences, but because the principle’s cognitive basis is well-understood and the burden of proof should now sit with maximalism, not with minimalism. Designers who load their video with kinetic typography and rapid cuts because “that is what students respond to” are taking a bet against a well-supported theoretical principle, and the bet is uncovered by the available evidence.

Another key point is that videos give a very strong sense of learning that is often an illusion. Becky Allen’s recent essay on her Substack Falsifiable extended Deslauriers et al.’s fluency-illusion finding to AI-generated educational video. The same extension applies to YouTuber explainers, with sharper teeth. Deslauriers and colleagues showed that students in polished, fluent passive lectures rated their perceived learning higher than students in challenging active learning, and scored significantly lower on the actual assessments. The correlation between perceived and actual learning was negative. The mechanism was specific: presentational fluency produces a feeling of understanding that decouples from, and at high doses can substitute for, the effortful encoding that produces understanding.

One more design heuristic worth naming. AI tools embedded in video learning environments should be configured for dialogue, not for delivery. The empirical pattern is consistent: students who use AI to think (Socratic prompts, follow-up questions, generative tasks) gain; students who use AI to receive (summaries, answers, abridgements) do not. The product decision to embed an “ask the AI” feature in an instructional video is not pedagogically neutral. The shape of the prompt matters more than whether the feature exists.

Three questions before “I learned so much from this video”

Becky’s diagnostic offers a clean way to handle any specific claim that one has learned something from a YouTuber explainer or a short-form video. Three questions tell you whether the claim is mechanistic or merely felt.

1. What kind of knowledge is being claimed? Is it factual recall (”I now know the population of Mongolia”), procedural-motor (”I can now truss a chicken”), conceptual understanding (”I now understand quantum tunnelling”), or transferable expertise (”I can now solve a class of physics problems I could not before”)? The format has different empirical credibility at each of these levels. It is most defensible at procedural-motor, plausible at factual recall with repetition, weak at conceptual understanding, and entirely speculative at transferable expertise.

2. What is the cognitive mechanism that produced the claimed knowledge? Was there retrieval practice? Was there elaboration? Was there generation? Did the video pause for self-explanation? Was the content revisited? Was the learner tested? If the answer to all of these is no, the only mechanism on offer is passive encoding from a fluent presentation, and the literature predicts that mechanism produces the felt experience of understanding without durable schema acquisition.

3. What would falsify the claim? The IOED methodology suggests an immediate test. Asked to write a detailed explanation of the topic without rewatching the video, can the learner produce one? Asked a transfer question that requires applying the principle to a new case, can the learner answer? If not, the claim of having learned is the fluency illusion in person. The video produced an experience. It did not produce a competence.

What the format is genuinely good for

A fair accounting requires saying what these formats do well, because the polemical “screens are ruining everything” reading is no more empirically grounded than the MrBeast reading (that students could learn more in 5 hours of YouTube than 8 hours of school) as Becky has shown.

So how can we best use AI mediated video formats for instructional design?

1. Procedural-motor content.

Höffler and Leutner’s 2007 meta-analysis found that animation outperforms static pictures with d = 1.06 specifically for procedural-motor content where the learning objective involves spatial transformation over time. Many YouTube tutorials are exactly this: how to fix a tap, how to truss a chicken, how to perform a free-throw, how to wire a plug. For procedural-motor learning, video has a real and well-documented advantage, and short-form is often the right length.

2. Vocabulary and single-fact acquisition with repetition.

Microlearning at the TikTok scale can deliver discrete facts repeatedly enough to embed them, particularly in language-learning contexts where the learner returns to the same vocabulary repeatedly. This is the strongest empirical case in the short-form literature, though even here the longitudinal evidence is thin so far.

3. Curiosity ignition

The format is unusually effective at producing the moment of “I want to know more about this,” which is a precondition for sustained engagement with a topic. This is a real cognitive good. A child who watches a Mark Rober video about marble runs is arguably more likely to investigate physics than a child who has not. The literature on interest and motivation in learning supports this; the precondition is real.

4. Exposure to topics that would not otherwise be encountered.

The discovery surface of YouTube and TikTok is wide. People learn that quantum entanglement, octopus cognition, the structure of the bond market, or the history of the Antikythera mechanism exist as topics worth thinking about. This is genuine information enrichment, even if the depth is shallow.

5. Affective relationship with subject matter.

A child who develops parasocial warmth towards Hank Green is more likely to identify with science as a domain that is interesting, accessible, and culturally legitimate. The empirical case for this as a learning effect per se is weak, but its role as a precondition for later learning is plausible.

None of these are nothing. None of them are the same as “students could learn more in 5 hours of video than 8 hours of school.” They are, at the very least, the conditions that prepare the ground for learning to take place elsewhere.

Closing

Educational video, in the form most students now consume it, has run ahead of the science designed to evaluate it. The field is in roughly the position medicine would be in if the only randomised trials had been conducted on a population that no longer existed. The principles are good. The patient has changed. Designers, parents and educators have to make decisions every day about whether to accept the medium as it now is, or whether to push back against the conditions that have made it less educative than it could be. I do not think those decisions can wait for the literature to catch up. They are being made by default, by those with the budgets to commission video and the algorithms to distribute it, and the burden of pushing back falls on those of us who still think learning is a more demanding cognitive activity than the current medium is willing to acknowledge.

Memory, Learning, and Knowledge

1. Does interleaving still work when the content gets hard?

This study tested a fairly important boundary condition in the “desirable difficulties” literature. Interleaving is often promoted as a powerful learning strategy, but most of that evidence comes from relatively simple materials. Here, 376 upper secondary students learned complex physics content under four conditions: interleaved vs blocked practice, and collaborative vs individual learning.

The key result: interleaving alone did not improve learning and actually led to worse long term outcomes. The benefit only appeared when interleaving was combined with collaboration, and that advantage held even after eight weeks. This study fits neatly into a growing body of work showing that desirable difficulties are conditional, not universal. It aligns with research showing interleaving benefits are strongest in lower element interactivity tasks (for example category learning or maths problem types) and weaker or negative in high complexity domains. It also connects with key ideas from John Sweller, particularly the aforementioned element interactivity and the idea that instructional techniques must match task complexity.

2. Is spacing effective because of memory or because of the teacher?

This study takes a familiar finding the spacing effect and tries to open the black box of why it works in classroom like settings. Students learned psychology concepts via instructional videos in either a spaced or massed format. The spaced group performed better on learning outcomes, which is entirely consistent with decades of research. The twist is that the authors also measured interpersonal brain synchronisation between teacher and student using EEG, and found that spacing increased this synchrony.

The key claim is not just that spacing improves learning, but that it may do so partly through social and perceptual mechanisms. Students in the spaced condition reported more positive views of teaching quality and interaction, and these perceptions mediated the relationship between brain synchrony and learning outcomes.

3. Retrieval practice works well for remembering things. Whether it helps students actually use what they remember is a different question entirely.

Fieke Ophuis-Cox’s doctoral dissertation at the Open University of the Netherlands reports four classroom studies conducted with real primary school pupils using existing teaching materials, not lab-based tasks with undergraduates. The headline finding across the programme is clear: retrieval practice reliably improves factual retention across domains, age groups, and settings. But the story gets considerably more interesting when the outcome is comprehension or application rather than recall.

In Study 1, second-graders practising multiplication facts with spaced flashcards outperformed a chanting-and-restudy group on both accuracy and fluency, at both five-minute and one-week delays. So far, so textbook. In Study 2, fifth- and sixth-graders studying expository texts showed a different pattern: retrieval practice improved retention of text content more than summarisation or restudy, but summarisation produced better comprehension outcomes. Retrieval practice did not outperform restudy on comprehension. In Study 3, fourth-graders learning verb spelling rules saw the same split: combining retrieval practice with worked examples substantially improved long-term retention of the rules, but this did not translate into better rule application. Students could recite the rules but still misspelled the verbs.

I’ve written more about this here:

Instruction and Practice

4. Do talking heads in video lessons actually help learning or quietly undermine it?

This study finds that showing the instructor’s face in video lectures often distracts attention and can harm learning, with its effects depending heavily on the teacher’s particular style.

Using an eye tracking experiment with 181 university students, the researchers compared videos with and without a visible instructor across three teaching styles: autonomy supportive, controlling, and neutral. The results are not flattering for the “talking head” trend. Instructor presence consistently pulled attention away from the learning material itself. Worse, under autonomy supportive teaching it reduced retention and concentration, and under controlling teaching it increased cognitive load and reduced concentration. Only under a neutral teaching style did instructor presence show a modest benefit for retention.

5. Directed inquiry eliminated student misconceptions in biology. The lecture-based approach made them worse.

This quasi-experimental study compared directed inquiry activities against a transmissive (lecture-based) approach for teaching primary school students about human organ systems. The directed inquiry group achieved a significantly higher frequency of correct scientific concepts after instruction, and — more strikingly — their misconceptions were eliminated. In the transmissive group, the frequency of misconceptions actually increased after instruction.

A few caveats before drawing strong conclusions. “Directed inquiry” here is not the same as unguided discovery. These were structured, teacher-led activities with clear learning goals, so closer to guided exploration than to the kind of minimally guided instruction that we know is not effective. And the finding that didactic teaching entrenched misconceptions may tell us more about the quality of that particular teaching than about lectures in general. If the transmissive instruction did not surface and directly address pupils’ existing mental models, it would have left those models intact or even reinforced them through superficial coverage. The study is a useful reminder that instruction which doesn’t engage with what students already believe risks talking past them entirely, but it should not be read as a blanket indictment of explicit teaching, which is by its very nature, highly interactive.

6. Worked examples help elementary students with maths, but adding self-explanation prompts may not always add value, and could even complicate things.

This study from the MathByExample research programme examined how worked examples paired with self-explanation prompts affect elementary students’ mathematics motivation and perceptions of errors. It sits within a larger body of work from this team: their 2023 meta-analysis found a medium effect (g = 0.48) for worked examples on maths performance, but also found that self-explanation prompts as an added feature yielded a negative moderating effect relative to worked examples without prompts. That counterintuitive finding — that asking students to explain their reasoning during worked examples can actually reduce their effectiveness — likely reflects the additional cognitive load that self-explanation imposes on novice learners still building basic schemas.

The current study shifts focus from pure achievement to motivation and error perception, asking whether error-focused worked examples change how students perceive mistakes. For teachers, the broader programme message is important: worked examples remain one of the most reliable instructional tools we have for building initial competence, but layering additional demands on top of them needs to be done carefully and with attention to where the learner actually is. More is not always more.

Reading and Language Development

7. Surprisingly, children with strong vocabularies at age four made the least progress in reading comprehension over the following decade.

This one will surprise many reading advocates. Using sequential latent growth models with two large representative Australian samples (N = 4,983 and N = 4,570), the researchers tracked receptive vocabulary development from age four to eight and reading comprehension from age eight to fourteen. The cross-sectional picture looked exactly as you would expect: stronger vocabulary at age four positively predicted reading comprehension at age eight. But the growth picture told a completely different story. Children who started with stronger receptive vocabularies at age four showed less growth in reading comprehension over the following years. And growth in vocabulary itself was unrelated to growth in reading comprehension from age eight to fourteen.

Before anyone takes this as evidence that vocabulary instruction doesn’t matter, the finding is almost certainly a ceiling and regression-to-mean effect: children who start high have less room to grow, and extreme scores naturally regress. But it is a sharp corrective to the assumption that early vocabulary advantage automatically compounds into accelerating reading gains. It doesn’t. It suggests that whatever drives ongoing reading comprehension growth beyond the early years — likely domain knowledge, text structure awareness, inferencing, and metacognition — is partly independent of the lexical foundation that gets children started. Vocabulary is necessary but not sufficient, and measuring it at age four tells you where a child starts, not where they’re headed. Ultimately, vocabulary is a strong predictor of where children start in reading comprehension, but a weak predictor of how quickly they improve over time.

8. AI-scaffolded reading strategy instruction outperformed both teacher-led instruction and business-as-usual, but the design makes it hard to say exactly why.

In this quasi-experimental study, sixty Jordanian undergraduates learning English as a foreign language were divided across three conditions: AI-assisted instruction (teacher-mediated with ChatGPT scaffolding), teacher-led strategy instruction, and business-as-usual. After sixteen weeks, the AI-assisted group outperformed both others on reading comprehension (partial η² = .43, a large effect) and showed the largest gains in metacognitive awareness of reading strategies (d = 1.49). The teacher-led group also outperformed business-as-usual, reinforcing that explicit strategy instruction matters regardless of medium.

The caveat, which the authors themselves flag clearly, is that each condition was a single intact classroom section with no random assignment of students. More importantly, the AI-assisted condition bundled multiple elements — ChatGPT scaffolding, weekly AI tasks, and continued teacher mediation — making it impossible to determine which specific feature drove the advantage. Was it the AI? The extra practice? The novelty? The additional task structure? The study tells us that a well-designed instructional package that includes AI can outperform one that doesn’t, but it does not tell us that AI was the active ingredient. For teachers, the transferable insight is that explicit reading strategy instruction reliably beats leaving strategy use implicit, and that AI may be a useful tool within such instruction — not a replacement for it.

Motivation, Self-Regulation, and Wellbeing

9. Students collaborating on scientific texts gradually developed more sophisticated ways of regulating each other’s thinking, but this was a painstaking, slow process.

This microgenetic study tracked two pairs of ninth-graders across twelve weekly sessions of collaborative document-based inquiry in biology. Students read texts making conflicting scientific claims with varying credibility and had to coordinate multiple sources. The researchers coded every conversational exchange for self-regulation, co-regulation, and shared regulation of epistemic thinking — that is, how students managed their goals, standards, and reasoning processes when evaluating knowledge claims.

The key finding was that shared regulation (where both students jointly managed their epistemic approach) was the most common form, and it increased over time, along with regulation of epistemic ideals; students gradually raised their standards for what counted as good evidence. This is encouraging, but the study’s design, two purposefully selected dyads, no control group, no achievement outcomes, means it demonstrates that social epistemic regulation develops, not how much it matters for learning. For teachers interested in collaborative inquiry with multiple sources, the implication is that students can develop increasingly sophisticated evaluative habits, but it takes sustained practice over weeks, not a single lesson on “critical thinking.”

Classrooms, Behaviour, and Attention

10. Changing how students are grouped had almost no impact on attainment.

This is an important one. A large-scale trial from the Education Endowment Foundation examined whether altering student grouping practices which we call “setting” in the UK, including reducing attainment grouping and increasing mixed-attainment teaching, improves outcomes for secondary school pupils. The study involved a substantial number of schools and used a randomised design, making it one of the more robust pieces of evidence in this space.

The headline finding is straightforward but will be very controversial: there was no meaningful impact on overall attainment. Pupils in schools that changed their grouping practices performed no better, on average, than those in control schools. There were also no consistent improvements in attitudes to learning or related outcomes. In short, reorganising how students are grouped did not translate into measurable gains.

This challenges a long-standing assumption in education that grouping structures are a major lever for improving outcomes. Indeed, some education academics have referred to setting or streaming pupils as “symbolic violence”. The mechanism often proposed is that mixed-attainment teaching raises expectations and improves access to high-quality instruction, particularly for lower-attaining pupils. That may be true in principle, but this study suggests that changing grouping alone is not sufficient. Without changes to curriculum, instruction, or teacher practice, the structural shift does very little.

has written about this very well here.

11. The largest study of school phone bans ever conducted finds they dramatically reduce phone use but have essentially zero effect on test scores.

This is one of the most controversial findings of the year so far. Hunt Allcott, Thomas Dee, Angela Duckworth, Matthew Gentzkow and colleagues analysed data from nearly five thousand US schools that adopted Yondr lockable phone pouches, combining GPS tracking data from 35 million phones, state standardised test scores, student wellbeing surveys, and two original national surveys. The research design — staggered difference-in-differences with pre-registration — is about as rigorous as observational education research gets.

Their findings are clear and uncomfortable for both sides of the phone debate. Phone pouches dramatically reduced in-school phone use: GPS pings on campus dropped roughly 30%, and teachers reported student phone use for personal reasons falling from 61% to 13%. But the average effect on test scores was precisely zero. The researchers can rule out improvements larger than 0.008 student-level standard deviations, in other words, not “we couldn’t detect an effect,” but “we can confidently say the effect is negligibly small.” High schools showed a modest positive effect in maths (about 0.024 SD, roughly equivalent to one-fifth of the effect of having a high value-added teacher), while middle schools showed small negative effects.

For educators, this should temper expectations in both directions. Phone bans appear to be a genuine wellbeing intervention — but not a learning intervention, at least not one detectable at the level of standardised test scores. This is an important distinction. The phone-free classroom may be calmer, less distracted, and better for students’ psychological health without necessarily translating into measurable academic gains. Schools implementing phone policies should do so for social and developmental reasons, not because they expect results day scores to move. And the short-term disruption is real — institutions need to plan for it rather than being blindsided when behaviour gets worse before it gets better.

SEN, Individual Differences and Inclusion

12. An AI-powered voice-first learning system designed specifically for visually impaired students achieves excellent usability scores, but hasn’t yet been tested on actual learning.

This design science study presents BlindLearn, a voice-first learning framework built from scratch for visually impaired students rather than retrofitting visual platforms with screen readers. The system uses wav2vec 2.0 for speech recognition and Tacotron 2 for text-to-speech, structured around a five-stage pedagogical model (Audio Activation, Narrative Input, Conversational Elaboration, Voice Practice, Adaptive Feedback). Testing with fifteen visually impaired students aged ten to sixteen in Indonesia produced a System Usability Scale score of 84.3, rated “Excellent” and well above published baselines for assistive technology. Expert validation was also strong, with a mean Content Validity Ratio of 0.89.

The study’s theoretical grounding in Universal Design for Learning and cognitive load theory is appropriate — the modality effect genuinely supports exclusive auditory channel use for this population. The limitation is straightforward: this is a usability study, not a learning outcomes study. The system scored well on ease of use, learnability, and satisfaction, but whether students actually learn more through it than through existing tools remains untested. The authors acknowledge this and recommend a twelve-week RCT with at least sixty participants as future work. The principle — that accessibility should be a design paradigm from the start, not a retrofit — is sound and important.

AI, Technology, and Cognitive Tools

13. Students who studied with ChatGPT felt like they had learned more. Forty-five days later, they had learned less.

This randomised controlled trial with 120 Brazilian undergraduates compared unrestricted ChatGPT use during study preparation against traditional methods (textbooks, articles, databases). Both groups prepared ten-minute peer presentations on AI and machine learning topics, then took a surprise retention test forty-five days later. The ChatGPT group scored 57.5% against 68.5% for the traditional group — an eleven percentage-point gap with a medium-to-large effect size of d = 0.68. The ChatGPT group also spent 45% less time studying (3.2 versus 5.8 hours), and crucially, the retention deficit persisted even after controlling for study time. Prior familiarity with AI tools offered no protection.

The author calls this “borrowed competence” — a fluency illusion where the AI’s articulacy is mistaken for the learner’s own understanding. This aligns directly with what we know about desirable difficulties: learning that feels easy rarely lasts, and removing the productive struggle from encoding undermines consolidation. For educators, the recommendation is not to ban ChatGPT but to sequence its use. Let students attempt their own encoding first, then bring AI in as a retrieval coach or a challenge partner rather than a first-contact information source. The moment AI replaces the thinking, it replaces the learning.

Another related study found that, 14 to 15 year olds using ChatGPT for science inquiry often failed to ask good questions, failed to spot weak answers, and stopped too soon, leaving performance only moderate despite open access to AI.

14. Not all AI feedback is created equal. Metacognitive feedback produces transfer; affective feedback does not.

This brain-imaging study with 87 Chinese college students compared three types of AI chatbot feedback while learning about the cardiovascular system: metacognitive feedback (prompting students to reflect on their understanding), affective feedback (encouraging and motivational), and neutral feedback. Using functional near-infrared spectroscopy alongside learning measures, the researchers found that metacognitive feedback produced significantly better retention and, crucially, transfer to new questions — the latter being the harder test. The metacognitive group outperformed the neutral group on transfer (M = 9.38 vs 6.72) and also outperformed the affective group, which showed no transfer advantage over neutral. Different feedback types activated distinct brain regions, with metacognitive feedback increasing frontopolar and middle temporal activation correlated with metacognitive sensitivity.

For anyone designing AI tutoring systems, this has a direct practical implication: building a chatbot that says “great job, keep going!” is not the same as building one that says “explain back to me what you just learned and where you’re still unsure.” The affective feedback felt supportive but didn’t produce durable learning beyond what students would have achieved anyway. Metacognitive prompting — pushing students to monitor and evaluate their own understanding — was the active ingredient. This aligns with decades of research on self-explanation and elaborative interrogation, and suggests that the most effective AI feedback systems will be the ones that make students think harder, not feel better.

15. A randomised controlled trial gave secondary school students GenAI support for self-regulated learning. It made almost no difference.

This well-designed RCT, following CONSORT standards, assigned 371 German secondary students (Grades 7–9) to one of three conditions across six physics and English lessons: GenAI-supported utility value reflection, GenAI-supported cognitive learning strategy prompting, or standard ChatGPT use as a control. The researchers measured self-reported motivation, elaboration strategy use, interest, effort, and domain-specific knowledge before and after.

The results are almost uniformly null. There was no significant effect on elaboration strategy use (self-reported or tested), effort, interest, or — the one that matters most — domain-specific knowledge. The one exception: the utility value condition helped students maintain their sense of the material’s relevance, while the cognitive strategy condition actually decreased perceived utility value. In other words, prompting students to use specific cognitive strategies with AI made the learning feel less worthwhile to them, not more.

This study matters because it tests exactly the kind of intervention that AI advocates routinely propose: use GenAI to scaffold self-regulated learning. The intervention was carefully designed, the RCT was rigorous, and the answer was: it didn’t work. Not because the AI was bad, but because the bottleneck in self-regulated learning is probably not something a chatbot can address in six lessons. The one bright spot — an exploratory finding that students who engaged more meaningfully with the AI maintained more interest — suggests that quality of interaction matters, but the study was not designed to engineer that quality. For educators considering AI-supported SRL tools, this is a reality check: scaffolding self-regulation is hard, slow, and relationship-dependent. Adding a chatbot to the mix does not change that.

16. Do virtual pedagogical agents actually motivate students? A meta-analysis of 58 studies says: some constructs yes, most no.

This three-level meta-analysis synthesised 304 comparisons from 58 studies to test whether on-screen virtual pedagogical agents improve student motivation. The results are tellingly selective. Agents significantly boosted self-efficacy (g = 0.29) and interest (g = 0.30), but had no significant effect on engagement (g = −0.02), intrinsic motivation (g = 0.00), value beliefs (g = 0.09), or utility beliefs. The “general motivation” composite was significant at g = 0.20, but this largely reflected the self-efficacy and interest components.

Design features mattered: agents with facial expressions showed larger effects on both motivation and self-efficacy. But the overall picture is one of modest, selective influence rather than broad motivational enhancement. Pedagogical agents appear to make students feel slightly more capable and slightly more interested, without meaningfully changing whether they find the material valuable, engaging, or intrinsically motivating. For developers and procurement decisions, this deflates the marketing: virtual agents are not motivational game-changers. They are modest tools with specific, limited effects.

17. A gamification study reports a strikingly large effect on academic performance — large enough to warrant scepticism...

This between-subjects experiment with 71 Finnish university students compared a gamified learning management system (using acknowledgment, chance, and progression elements) against the same system without gamification. The gamified group substantially outperformed the control (M = 0.80 vs 0.49), yielding a Cohen’s d of 1.41 — an effect size that would be exceptional for almost any educational intervention.

Before celebrating, the usual caveats apply with extra force here. The sample is small (71 students after outlier removal). The study was a single session. The outcome measure was task performance within the gamified system itself, not a separate knowledge transfer test. An effect size this large in a real educational context would mean gamification is more powerful than reducing class sizes, high-dosage tutoring, and expert teaching combined. It almost certainly isn’t. What it probably shows is that game elements increase immediate task completion and engagement within the platform — a useful but much narrower finding than “gamification improves learning.” Teachers and ed-tech buyers should be wary of studies measuring outcomes inside the intervention rather than outside it.

Resources

IES Teaching Math to Young Children — Professional Learning Toolkit

The US Institute of Education Sciences has released a free, comprehensive professional learning toolkit based on the What Works Clearinghouse practice guide Teaching Math to Young Children. It covers nineteen weeks of structured professional development across four modules — subitising, meaningful counting, comparing magnitudes, and solving basic problems — with approximately twenty hours of module work and eighty-five hours of classroom implementation. Designed for preschool through kindergarten teachers, it includes video, fillable learning journals, classroom activities, and progress monitoring tools. Based on four evidence-based recommendations from a national expert panel. Freely available at ies.ed.gov.

NIFDI Science of Reading Webinar

Dr. Kurt Engelmann recently presented a fascinating talk for the first webinar in the Science of Reading Series, focused on how Direct Instruction builds reading comprehension, not just decoding or fluency. The session covered DI approaches for preschool, kindergarten, ESL/EAL/MLL, and early elementary students, with a focus on the language and comprehension skills that often get neglected in phonics-heavy implementations. Worth watching if you work with early readers or want to see what DI actually looks like when it’s done properly. Watch the recording here.

Housekeeping

• I’ll be in Auckland NZ on 22 May, Brisbane on Monday 25th May and in Melbourne on Friday 29th May doing a full day event on Retrieval Practice, Responsive Teaching & Checking for Understanding.

• I have written a free guide on responsive teaching for UNESCO available here. Jamie Clark has made a One Pager to accompany it.

I supported the Australian social media age ban when it was announced, and I have been, for some years, an enthusiastic advocate of getting phones out of classrooms. Neither view has changed. But I also try to be the sort of reader who updates their priors when new evidence comes to light, and a recent study has nudged me to do precisely that. Not by overturning the case for a ban; the case still holds. What it has done, rather, is question something which is far more concerning for me: that the scourge of social media is not a problem a blanket ban can solve on its own, and that a policy which addresses the apps while ignoring everything else children are growing up inside risks delivering the appearance of protection rather than the substance of it.

The prohibition paradox: when social bans make behaviour more social

The interesting finding with this new study is not that the ban might be failing. The interesting finding for me is why it might be failing. It seems the law has not come close to reaching the tipping point at which enough teens change their behaviour that others are willing to follow. Put simply, it has declared a rule without shifting the social equilibrium.

Something similar happened a century ago with American Prohibition. The Eighteenth Amendment banned the manufacture, sale and transport of alcohol from 1920, but drinking was so woven into social life that the ban mostly pushed it sideways rather than stamping it out. Legal saloons closed and were replaced by speakeasies, illegal bars that became the worst‑kept secret in every city, with some estimates suggesting there were more speakeasies than there had been licensed bars before the ban.

For many Americans, going out for a drink did not stop being a normal social pastime; it simply became a slightly more glamorous, rebellious one, backed by bootleggers and organised crime instead of breweries and bartenders. Prohibition shifted the law on paper but not the descriptive norm: people still expected their friends, neighbours and even politicians to drink, so the noble experiment collapsed under the weight of everyday non‑compliance.

Injunctive norms versus lived reality

In this new study, teenagers were asked, in plain terms, what share of their peers would have to leave social media before they would too. The mean answer was 69 percent. The current compliance rate is 27 percent. That gap is not a measurement artefact. It is the gap between an injunctive norm (what the law says you ought to do) and a descriptive norm (what you can see your peers actually doing). When the two point in different directions, the descriptive norm usually wins.

Cristina Bicchieri has been telling us this for twenty years: social norms bind when people both expect others to comply (empirical expectations) and expect that others think one ought to comply (normative expectations), often backed by social sanctions. Seen this way, the law is almost a textbook exercise in moving only one piece of the norm system. Parliament has loudly broadcast the injunctive norm (“under‑16s should not be on social media”) but teens’ empirical expectations remain that most of their friends are still on Snapchat, TikTok, or YouTube, and they do not expect disapproval from peers if they keep using them.2 Their beliefs about what their friends are doing, and what their friends think is acceptable, have barely budged; the formal rule is just another piece of adult noise in the background.

The deeper problem is one of social composition. The researchers asked teens whether kids who comply with the ban are more or less popular than kids who do not. Forty‑seven percent said less popular. Five percent said more popular. The compliers are sitting in the social periphery; the non‑compliers are at the centre. Bicchieri and others have argued that norms are sustained not just by rules but by the payoffs attached to compliance and violation. Here, the payoffs are upside‑down: following the rule makes you less central to the social world that matters to you.

What worked with smoking, and what is happening here

The Australian ban seems to be the inverse of what brought adolescent smoking down, but not for the reason the popular-kids-quit-first folk theory suggests. There’s a strong argument to say that smoking declined because the meaning of smoking was changed and the rate of initiation was cut as a result. The 1998 Master Settlement Agreement restricted tobacco advertising; the CDC’s Truth campaign recoded smoking from cool to industry manipulation; states raised tobacco taxes and banned indoor smoking; graphic warnings appeared on packs; adult smoking declined in parallel and reduced household exposure. The network effect followed rather than led: as Haas and Schaefer’s longitudinal data on US adolescents show, peers exert their strongest pull on smoking initiation, not on cessation.

Adolescents rarely start without peer influence, but they will quit while their friends carry on (see Haas & Schaefer). That is the leverage point smoking policy hit. Once initiation falls, each cohort sees fewer smokers, and the lower rate sustains itself.

There is a separate evidence base for targeted network interventions, which is what the original draft was reaching for. Paluck, Shepherd and Aronow’s anti-conflict experiment across 56 New Jersey middle schools found that interventions with the most socially connected students ( “social referents,” identified by peer nomination) reduced disciplinary reports for student conflict by roughly thirty percent over a year, while equivalent interventions with randomly selected students did nothing.

The mechanism, on closer reading however, is not high status as such. Damon Centola’s network-tipping work argues that what matters is the density of redundant ties among committed minorities; status helps mainly because it correlates with connectedness, but the active ingredient is mutual reinforcement among early compliers.

It appears Australia has not laid the ground for this. There has been no advertising restriction analogous to the Master Settlement Agreement, no counter-marketing campaign that has done to TikTok or Instagram what Truth did to Marlboro, no taxation of attention, no parallel adult withdrawal. The ban is a single crude instrument, aimed at use rather than at meaning, and the descriptive norm is at near-saturation: 91 percent of unbanned 16–18-year-olds in Australia, 95 percent of US teenagers, on the same data. There is no pool of never-users for a meaning-shift to protect, and no structural counter-marketing recoding what the platforms signal. The mandate has, if anything, conferred a faint countercultural sheen on continued use.

And if the loudest voices speaking out against social media are middle aged men (usually on social media unironically) then the campaign probably has the wrong vector, and teenagers, who are excellent readers of social vectors, will not miss it. Adolescents have a fine ear for the genre; when the moral entrepreneurs are the in-group, they hear that this is a thing the dads want, which is reason enough to keep doing it.

This matters beyond one country’s drafting. It matters for any rule that depends on coordinated abstention or coordinated adoption: phone-free classrooms, homework platforms, dress codes, the language norms of a corridor. A mandate that ignores the network composition of its early compliers, and the cultural meaning of the practice it targets, fails in the same recognisable way. The thresholds are too high, the rule becomes the badge of the dutiful and the uncool, and non-compliance accrues prestige it would not otherwise have. The smoking case is instructive precisely because it shows what the work actually consists of; meaning-shift plus structural levers acting on initiation, with network sustenance following.

Two design choices in particular are a problem

The authors in this paper highlight two major problems. First, the ban operates on age, not grade. Adolescent peer groups are organised by classroom, not by birthday. A 14‑year‑old sitting in a class with several 16‑year‑olds is staring directly at peers for whom the law does not apply. Within‑classroom exposure to unbanned older peers jumps sharply between 14‑ and 15‑year‑olds, and so does the rate of use. Grade‑based cut‑offs would have given the law a fighting chance of changing what teens actually saw around them.

Second, enforcement falls on platforms, not on users and their guardians. Civil penalties of up to A$49.5 million attach to companies who fail to verify ages. (This is laid out in early coverage such this BBC explainer.) Teens and their parents face nothing. The teens have noticed: only 22 percent believe there are personal consequences for them. Non‑compliance carries neither an individual cost nor a visible social signal, which is the regulatory analogue of asking the child not to eat the marshmallow while leaving the marshmallow on the table. Qualitative work on norms suggests that when individual sanctions (social or legal) are absent, and everyone can see that violators are not punished, injunctive norms quickly hollow out.

The numbers are striking. Among the paper's 14-year-olds, fewer than one in ten classmates on average is aged sixteen or older; by 15, the figure is roughly four in ten; by 16, eight; by 17, nine. The same transition between 14 and 15 that sees a teenager move from a class of largely compliant peers to one of largely non-compliant ones also sees use inside the banned group jump from 57 to 68 per cent.

In fairness, the authors are careful not to read this as causal, and rightly so, but the coincidence is hard to ignore: the moment the typical banned teenager begins to share a classroom with a critical mass of unbanned older ones is also the moment compliance begins to crumble. Drawing the line by year of schooling rather than year of birth would not have made the ban watertight, but it would at least have stopped the law from running headlong into the social geography of secondary school.

There is also a more basic design flaw. The duty to police ages falls on a subset of large platforms, leaving adolescents free to migrate to services outside the law’s scope. Early reporting suggests many Australian teens have shifted time into group chats, gaming platforms, and services such as Discord, which they do not perceive as “social media” in the same way. (See “Kids locked out of social media apps after Australia’s world‑first ban.”) From the teenager’s point of view, their social graph is still intact; only the logos on the icons have changed. In normative terms, the descriptive norm (“everyone I know is still online with me”) remains the same.

All of this would be worrying enough if the ban were a well‑targeted instrument for improving wellbeing. The evidence there is more complicated than the policy debate admits. Large longitudinal studies of Australian adolescents find a non‑linear relationship between social media use and wellbeing: very heavy use is associated with poorer outcomes, but moderate use often coincides with better wellbeing than either abstinence or extreme use. A synthesis of data on Australian teens concludes that “one‑size‑fits‑all rules and blanket bans are unlikely to be effective,” with moderate engagement (more than none, much less than all day) linked to the most favourable wellbeing profiles. Even if the law had changed behaviour, its mental‑health returns would probably have been uneven.

Designing a ban that could actually work

So what would actually shift behaviour? The literature is reasonably clear, even if politically uncomfortable. You have to engineer the descriptive norm, not just declare the injunctive one. That means:

• Working on what compliance signals about you, so that following the rule does not push you to the social periphery.

• Recruiting high‑status early movers, as network‑based smoking interventions try to do with central students.

• Providing alternative social infrastructure (sports, clubs, online and offline spaces) that fills the time and connection the platforms used to fill, especially for those most reliant on digital communities.

• Pairing all of this with friction at the device level (defaults, screen‑time controls, age‑gating at operating‑system level) rather than only at the corporate level.

None of it is impossible. Most of it is standard public health and standard social‑norms engineering but very hard to get right. As a commentary in Nature puts it, Australia has created a “natural experiment” in how far you can get by legislating a social norm without reshaping the social networks that sustain it.

If your intervention does not change what teens see their peers doing, do not expect it to change what teens do. Prohibition lasted thirteen years before that lesson was admitted. We can hope this one moves faster.

Protecting my kids, without fooling myself

I have skin in this game. I would love my three children (all still under ten), to grow up without the worst of what social media has done to this generation of teenagers; the relentless comparison, the always‑on group chat, the sense that your social life can fall apart because of something you missed while you were asleep. I don’t want them marinating in that from the moment they leave primary school.

But wanting that world does not mean any ban that gestures in its direction is going to help. If anything, blanket bans shows how easy it is to get the symbolism of protection without changing the social reality children grow up inside. You can pass a law, announce a new age limit, even wipe a few million accounts, and still leave every ten‑year‑old looking up at a social universe in which “everyone who matters” is online together, just out of their legal reach.

So I support a ban. But the question is which one? A ban that raises the formal age, leaves the platforms intact, and leaves every fourteen‑year‑old in my children’s future cohort already practised at lying about their birthday is not really a ban; it is stagecraft. Worse, it risks pushing children into lying more convincingly about their birthdays, into the unlit corners of the internet, the encrypted group chats, the Discord servers no parent has heard of, the Telegram channels that have already absorbed much of what mainstream platforms nominally moderate away, while convincing the rest of us that the problem has been dealt with.

So I find myself in an awkward middle: I do want a ban, but I want a ban my own children won’t experience as a dare, or a joke, or a lesson in how early you have to start lying to be part of the group. I hope we can do something slower and less photogenic that actually shifts their social world than something headline‑friendly that mostly shifts our self‑image as responsible adults. The test of a policy like this is not whether it lets parents feel that they have drawn a firm moral line. It is whether, ten years from now, a child like mine looks around their classroom and sees a different everyday reality than the one teenagers are living in now.

Knowing That, Knowing How

In The Concept of Mind, published in 1949, Gilbert Ryle drew a distinction between two kinds of knowing. When we teach a child a rule, we are doing something in the domain of knowing-that. When we want her to be able to use the rule, we are asking for something in the domain of knowing-how. The two are related; but they are not the same; and nothing in the logic of the first guarantees the delivery of the second.

A child who can recite the rules of chess is not thereby a chess player; a child who can describe the strokes of swimming is not thereby a swimmer. Knowing that and knowing how are separable epistemic achievements. One can travel a long way in the first without taking a single step into the second. The distinction is philosophical in origin but has real implications for instructional design.

Some recent studies I read reminded me of this and also really underlined a lot of why retrieval practice is going wrong in many classrooms. A new study/doctoral thesis on retrieval practice reminded me of Ryle’s classic work and puts this old distinction through an empirical test. A group of Dutch fourth‑graders, spread across seven regular primary classes, spent a week learning to spell their language’s trickiest verbs. All pupils received a scripted lesson on the relevant rules. After that, the classes split. One group studied worked examples, the traditional diet of verb spelling instruction. The other group studied worked examples and, crucially, practised recalling the rules from memory: three spaced retrieval sessions across a week, each requiring pupils to reconstruct the full rule set from memory, then check against a correct version on the board. A week later, both groups sat the same tests.

The first result was what anyone familiar with the retrieval‑practice literature would expect. On a retention test, the group that had practised retrieval recalled the rules at about 88 per cent accuracy. The control group managed about 43 per cent. This is a vast gap, of the sort that researchers have been documenting for two decades.

But the second result was harder to explain away. On an application test, where pupils had to actually spell the verbs correctly in sentences, both groups improved from pre‑test to post‑test, but they performed essentially identically to each other. The retrieval group scored about 53 per cent correct; the worked‑examples group about 57 per cent, a difference that was not statistically reliable. The conundrum here is that the students who knew the rules roughly twice as well did not apply them any better.

So what’s going on?

Two Processes, Not One

This is the puzzle the study leaves on the table. In the ordinary logic of classroom teaching, more memory should mean more capable use; the group with the rules firmly encoded should have at least edged the group without them. That they did not is a small but significant crack in a picture we usually take for granted.

The framework the Dutch researchers invoke to explain the dissociation, Koedinger’s Knowledge‑Learning‑Instruction theory, is a useful modern refinement of Ryle. Acquiring a rule‑based skill, on this account, depends on two separable cognitive processes. One is memory: the rule must be encoded, retained, and retrievable on demand. The other is induction: the learner must abstract the patterns that govern when and how the rule applies, and must integrate the rule into a live problem‑solving procedure.

These processes can be strengthened independently. But they can also fail independently. A learner can have a rule robustly encoded and still fail to summon it at the right moment; a learner can intuit when a rule applies while struggling to recall its exact form. From the outside, both failures look the same. From the inside, they are distinct malfunctions requiring distinct repairs.

The temptation, especially when we are under pressure to demonstrate impact, is to treat both as memory problems and to reach for the same tool. The error is not to underestimate memory. Memory matters enormously. The error is to treat memory as enough. It is the substrate on which skill rests, not the skill itself; the fact that a rule has been written into memory says little, on its own, about whether it will appear on the page when it is needed.

What Retrieval Practice Is For

This places retrieval practice, (probably the most celebrated and yet lethally mutated learning strategy of the last twenty years), in an awkward but instructive light. Retrieval practice does exactly what the research has always said it does: it produces durable memory traces but that is not enough. The Ophuis‑Cox paper replicates this finding with characteristic elegance; the retrieval group’s retention score did not merely edge the control, it doubled it.

The more interesting question I think is when and for what retrieval practice does this job best. A second recent study, this time with undergraduates learning word pairs, explores this further. Gupta, Pan and Rickard asked whether retrieval practice helps equally when the cue and target are strongly related in meaning or only weakly connected. Participants studied word pairs that were either low or high in semantic relatedness, then either restudied them or practised cued recall with feedback, before a delayed test 24 hours later.

As you would expect, testing with feedback beat restudy across the board: retrieval practice still does its basic job. But when they modelled the data using a dual‑memory framework and looked not just at means but at the full distribution of performance, they found that the benefit of testing was about 26 per cent smaller for highly related pairs. The standard ANOVA showed no interaction, yet the model and the cumulative distributions made it clear that the relative efficacy of testing was systematically reduced in the high‑relatedness condition.

A recent review by Roelle, Rummer, Schweppe and colleagues in Learning and Instruction makes the same point at a higher altitude. Under the banner of “meaningful learning”, they distinguish between the work of constructing rich, integrated knowledge and the work of maintaining those outcomes over time. Retrieval practice, in their account, belongs firmly in the second camp: it is a way of stabilising and refreshing knowledge that has already been meaningfully learned, not a shortcut to meaning by itself. On that view, the Dutch verb‑spelling study is not an anomaly; it is what you should expect when a consolidation tool is asked to carry the full burden of construction as well.

What Kind of Knowledge?

So the Dutch pupils could retrieve the verb rules at will on a structured test, but this did not yet buy them better performance in real spelling. The undergraduates benefited from testing on both tightly and loosely related word pairs, but the extra gain from testing was largest where the semantic fabric was thinnest. Taken together, those results suggest that what retrieval practice buys you depends as much on the underlying knowledge structure and how it’s sequenced as the act of retrieval itself.

This matters for how we design curricula and practice. If you are working with isolated facts and arbitrary labels, retrieval practice is probably doing more work than anything else you have. If you are working inside a rich, well‑sequenced knowledge domain, a good part of what looks like a “testing effect” is really the underlying semantic structure doing its job, and the incremental advantage from retrieval will be smaller. And in neither case should you expect practice tests, by themselves, to conjure fluent, procedural performance out of a rule that has barely had time to bed in.

The clearest example here I think is vocabulary. In a curriculum where new words are taught as standalone lists, Friday spelling tests and vocabulary flashcards stripped of context, retrieval practice is the main thing keeping those words in a pupil's mind. Each word is an isolated brick, held up by nothing but rehearsal. In a curriculum where the same words are encountered repeatedly inside a knowledge-rich programme of study, where a child meets "famine" in a text about the potato blight, then in a history lesson on Ireland in the 1840s, then in a poem about harvest, then in a science unit on soil, each encounter does double duty: it rehearses the word and thickens the semantic web the word sits in.

The quiz in the second curriculum looks less heroic because much of the work is being done elsewhere. The quiz in the first curriculum looks essential, because there is nothing else propping the word up. This is the shape of the Hirschian point: knowledge-rich curricula do not make retrieval practice unnecessary; they change what it is doing and how much of the load it is being asked to carry.

The upshot of all this is not that retrieval practice is overrated, but that we must be much more exact about what kind of knowledge we are strengthening, and what we expect that knowledge to support.

The Fluency Threshold

Why might strong memory for the rule fail to convert into fluent application? The Ophuis‑Cox team propose a mechanism worth taking seriously. Applying Dutch verb spelling rules is not a single cognitive act; it is a sequence. The speller must identify the verb, determine whether it is finite, ascertain its tense, analyse the subject, select the appropriate rule from several candidates, and apply it. Each step competes with the others for limited working memory.

If the rule itself is not yet fluent, the act of retrieving it consumes cognitive resources that are needed downstream for the grammatical analysis. The speller runs out of room before she reaches the answer. Under pressure, the brain does what brains do under pressure: it defaults to the fast, frequency‑based system. It picks whichever ending it has seen most often and hopes for the best, exactly the pattern spelling researchers have reported in Dutch and other languages.

This is why fluency is not optional. The point of practising a rule to the point of automaticity is not aesthetic, and it is certainly not anti‑intellectual. It is that a rule which cannot be applied without effort cannot be applied at all, once the other cognitive demands arrive. Fluency frees the working memory that application requires. Without it, knowing the rule and not knowing the rule produce the same output: whatever the frequency‑weighted lexicon happens to throw up.

And, as Gupta and colleagues suggest, fluency is not just a matter of more practice; it depends on how well the new material is anchored in what is already known. Highly related items exploited pre‑existing semantic scaffolding and showed smaller marginal gains from testing than weakly related items. What looks like “more retrieval practice” in the abstract is in reality different work, depending on the underlying knowledge structure.

On this reading, the Dutch study did not fail to help children apply the rules. It succeeded in helping them encode the rules, and ran out of time before it could build the fluency that application requires. Three sessions across a week is a respectable dose of retrieval practice by the standards of the field. But it is nothing like enough to push a freshly acquired procedure into the zone where it costs the working memory nothing to deploy.

Fine-Tuning Retrieval Practice

If this diagnosis is right, four consequences follow for how we think about practice.

1. The first is that the criterial task matters. If you want pupils to apply a rule fluently, you must eventually practise applying the rule fluently; it is not enough to practise just remembering what the rule is. The Dutch children rehearsed retention and got better at retention. They did not rehearse application at scale, and application did not improve beyond what the worked examples had already produced. What you practise is what you get.

2. The second is that sequence matters. Retrieval practice is probably best understood as one stage in a longer pipeline, not as a standalone intervention. Get the rule into memory first, yes; then build fluency with it; then practise applying it under the conditions in which it will be used. Each stage has its own instructional signature. Skipping stages produces exactly the dissociation the Ophuis‑Cox study observed.

3. The third is that dosage matters. Three practice sessions is a generous science‑of‑learning intervention. But it’s a trivial amount of practice by the standards of actual skill acquisition. A child who has rehearsed the times tables three times in a week does not yet know the times tables; she has met them. Much educational research operates at a dosage far below the level at which the interesting thresholds appear, and then reports findings that would probably shift if the intervention ran for another month.

4. Curricular architecture matters. If retrieval practice yields its largest marginal benefits precisely where semantic relatedness is low, then we should not be surprised if quizzes look most heroic in knowledge‑thin curricula, and more modest in content‑rich ones. Where we have failed to do the groundwork of building a shared store of background knowledge and vocabulary, retrieval practice ends up doing double duty as both memory booster and makeshift semantic scaffolding. Where that groundwork has been done, tests are still useful, but they are nudging an already coherent structure rather than trying to hold a pile of sand together.

For example, consider a child who can chant “i before e except after c” on demand, reliably, word-perfect, and who still writes “recieve” three times a week. She knows the rule. She has practised the knowing. What she has not practised is the act of summoning the rule in the fraction of a second when her hand is choosing between two letters. The gap is not a gap in knowledge. It is a gap in what the knowledge has been rehearsed for.

The test of a rule is not whether a child can produce it, but whether she can deploy it under the cognitive conditions in which it is needed, and producing and deploying are not the same exercise

Curricular architecture: thin vs rich knowledge

And when you zoom out from lessons to whole curricula, the shape of the knowledge you are teaching becomes the next constraint. Retrieval practice does not land on a blank canvas; it lands on whatever semantic structure the curriculum has already built. Two quick contrasts make the point.

1. Knowledge‑thin example
   A Year 7 geography scheme is built around a set of skills (“interpret graphs”, “evaluate sources”) with little substantive content. Retrieval practice here often means quizzing on isolated terms (“delta”, “urbanisation”) or procedural phrases (“PEEL paragraph”). In that setting, quizzing has to act as makeshift semantic scaffolding: it is the only thing holding those stray terms in memory long enough to be used on the end‑of‑topic test. Unsurprisingly, the test scores look “transformational” when quizzes are introduced.

2. Knowledge‑rich example
   A Year 7 history curriculum spends weeks building a dense web of knowledge about, say, the English Reformation: who the main actors were, what they believed, how events unfolded, how it links to Europe. Pupils read multiple narrative accounts, work with timelines and maps, and see the same ideas in different guises. In that context, weekly retrieval quizzes – dates, names, cause‑and‑effect links – still help, but the material was already semantically connected. The marginal benefit of quizzing is smaller, because the curriculum itself is doing much of the work that quizzes had to do in the barren geography scheme.

Conclusion

Put crudely, retrieval practice is at its most potent when the connections are weak and need to be built, not when the material is already woven tightly into a semantic web. It is a strategy of the first kind in two senses: it strengthens memory rather than induction, and within memory it is especially good at shoring up the lonely, fragile facts that would otherwise fall away. The error in the field, if there has been one, has been to treat retrieval practice as though it were a general‑purpose solvent for the whole of learning. It isn’t.

These new papers highlight to me that Ryle’s distinction still does not have the traction it deserves. Education continues to assess what is easy to assess, to practise what is easy to practise, and to treat the resulting scores as though they measured the thing that mattered. The Dutch children could recite the rule beautifully. When it came to the thing the rule was actually for, they performed no better than the children who could not recite it at all.

If the point of teaching the rule was the spelling, we were not teaching the thing we thought we were teaching. If the point of quizzing the facts was to enable fluent reasoning, we were again stopping too early, at the point where knowing‑that feels satisfying and measurable. The child who knows the rule and still writes the wrong form is not failing. She is telling us, with the precision of a diagnostic instrument, that we stopped too early.

• Mohan W. Gupta, Steven C. Pan & Timothy C. Rickard, ‘Semantic relatedness and the efficacy of retrieval practice’, npj Science of Learning (2026).

• Fieke H. A. Ophuis‑Cox, Leen Catrysse & Gino Camp, ‘Mixed effects of retrieval practice on the acquisition of procedural knowledge in verb spelling education’, Learning and Instruction (2026).

Memory, Learning and Knowledge

Does the sequence of retrieval and generative tasks matter?

If retrieval practice strengthens memory and generative activities like producing examples or explaining ideas support deeper understanding, then surely the order in which students do these things matters. A new laboratory experiment tested exactly this assumption, and found surprisingly little evidence for it.

University students studied short texts and then completed different sequences of learning tasks. Some generated examples before retrieving definitions; others retrieved definitions first. A third group restudied the material before completing generative tasks. Some students completed these activities immediately, others after a two-day delay. One week later, the sequence had made almost no difference. Students performed similarly regardless of whether they generated before retrieving or retrieved before generating.

What did matter was whether retrieval happened at all. Retrieval practice consistently produced better long-term retention than restudying, even when restudy was paired with generative activity. The implication is worth sitting with: teachers who agonise over the optimal choreography of learning activities may be solving the wrong problem. The critical variable is not the sequence but the presence of retrieval itself. Get students retrieving from memory, and the exact order of the surrounding activities appears to be a second-order concern.

Retrieval practice works but how much of this evidence actually comes from real classrooms, and does it work for all learners?

Tom Perry’s review of cognitive science in the classroom revealed an uncomfortable truth for many science of learning advocates: most research does not come from real classrooms. We know that retrieval practice is one of the most robust findings in cognitive psychology and over the past two decades, hundreds of studies have demonstrated what is often called the testing effect: attempting to retrieve information from memory strengthens long-term learning more than restudying the same material. A new perspective piece in npj Science of Learning analyses 23,850 publications on the testing effect and asks two important questions: how much of this evidence actually comes from real classrooms or from the laboratory? and does it generalise to all learners?

The answer is: mostly from the laboratory. Much of the testing effect literature comes from controlled experiments using short learning materials and university students. These studies provide strong evidence that retrieval practice works under tightly controlled conditions, but relatively few examine authentic classroom settings where instruction unfolds over weeks, involves multiple competing variables, and serves learners who are not psychology undergraduates. There is also limited research on diverse learner populations, including students with special educational needs or younger children, precisely the groups for whom the translation from lab to classroom is least straightforward.

None of this undermines the testing effect itself. The phenomenon is real and well-replicated. But classroom environments introduce what might be called a signal-to-noise problem: in a laboratory, the effect of retrieval practice is easy to detect because everything else is held constant. In a real classroom, that signal is diluted by differences in instruction, prior knowledge, motivation, and the sheer complexity of thirty minds working at different speeds on the same material. The review is a useful reminder that the distance between a robust laboratory finding and a reliable classroom practice is not a straight line, it is a translation, and translations always lose something.

Reading and Language Development

Teaching children background knowledge before they read related texts does improve their comprehension, but the benefit is not equally durable for all learners.

This study set out to test a deceptively simple but practically consequential question: if you deliberately teach children the knowledge they need to understand a text, does that instruction actually improve their comprehension of it? Seventy-three Year 2 students (aged around eight) from a single regional school in Victoria were taught four topic-based knowledge units covering artefacts, fossil fuels, Faith Bandler, and Nelson Mandela, using an explicit instruction model over two four-week blocks. Students were then assessed on how much of the knowledge they had retained, and whether that retained knowledge enabled them to make inferences when reading related texts. A key additional manipulation was timing: some students read the related text immediately after instruction, while others read it after a four-week delay during which the knowledge was neither reviewed nor revisited. The study found that explicitly teaching a knowledge base did enable both skilled and less skilled readers to make more accurate inferences. However, less skilled readers acquired the knowledge base less thoroughly within the same instructional time, and they also showed significantly greater comprehension decline when there was a delay between learning the knowledge and reading the text.

The study supports the case for knowledge-rich curriculum design: deliberately teaching domain knowledge does appear to translate into improved comprehension of related texts, including for weaker readers. But the study also reveals a structural problem embedded within that approach. Less skilled readers, the very students a knowledge-rich curriculum is most intended to help, are less likely to fully acquire the knowledge base within a standard unit of instruction, and more likely to lose access to it if it is not regularly revisited. The metaphor that comes to mind is scaffolding that is removed too soon: the knowledge built through instruction may simply not be stable enough, in less skilled readers, to bear the weight of inference-making weeks later. This implies that effective implementation of a knowledge-rich curriculum is not merely a matter of content selection, but requires deliberate, recurring review built into the instructional sequence itself, what the authors call schema maintenance.

How students read multiple texts may depend less on general strategy skill and more on whether they know what strategic reading actually looks like.

A new study in Contemporary Educational Psychology examined the relationship between what students know about reading strategies and how they actually behave when reading multiple texts. The researchers developed a measure of strategic knowledge, not whether students can execute strategies, but whether they can recognise what good strategic behaviour looks like in the context of multi-text reading. They then tested whether this knowledge predicted actual strategy use and, in turn, comprehension.

The findings suggest that strategic knowledge is a meaningful but underexplored predictor. Students who could identify effective approaches to integrating information across texts were more likely to use those approaches themselves, and this translated into better comprehension outcomes. The implication is worth pausing on: much of the strategy instruction in schools focuses on practising strategies, but relatively little attention is given to building students’ declarative knowledge about when and why those strategies work. You cannot deploy what you do not understand. This echoes a broader pattern in the science of learning: metacognitive competence rests on a knowledge base, not merely on procedural habit.

Fostering creativity in schools: a systematic review of what actually works for younger learners.

A systematic review in Educational Research Review examined interventions designed to cultivate creative thinking in primary and early secondary classrooms. The review distinguished between explicit interventions — those that directly teach creative thinking skills or techniques — and implicit interventions that embed creative opportunities within regular instruction without labelling them as such.

The question the review addresses is one that education has struggled with for decades: can creativity be taught, and if so, how? The findings suggest that both explicit and implicit approaches can produce measurable gains, but that the evidence base is stronger for explicit instruction in creative thinking strategies. This is a finding that aligns with broader patterns in instructional research: making the target of learning visible and explicit tends to produce more reliable outcomes than hoping it will emerge as a byproduct of open-ended activity. For those tempted to dismiss creativity instruction as inherently incompatible with structured teaching, the review offers a useful corrective. The most effective creativity interventions in the review were not unstructured. They were highly structured, but structured around the right things.

Teaching and Classroom Practice

Do we actually understand what makes feedback effective, or are we still describing what feedback looks like?

Two companion papers in the March 2026 issue of Contemporary Educational Psychology take stock of the feedback literature and arrive at a shared conclusion: the field knows more about the forms feedback takes than about the processes by which it works.

Daumiller and Meyer argue that feedback research has been disproportionately focused on surface-level features — timing, specificity, source — while paying insufficient attention to the cognitive and motivational processes that determine whether feedback is actually used. Their call is for a shift in focus: from what teachers do when they give feedback to what happens inside the learner when they receive it. Koenka, McManus, and Nicolai make a related but distinct point. They argue that feedback’s conceptual foundations remain underdeveloped, and that the field needs a more coherent theoretical framework to explain why the same feedback intervention can produce dramatically different effects depending on the learner, the task, and the context.

For practitioners, the message is both validating and cautionary. Teachers have long suspected that giving feedback is not the same as students learning from it. These papers confirm that suspicion and locate the gap precisely: the problem is not a shortage of feedback, but a shortage of understanding about what happens between the moment feedback is delivered and the moment it either changes thinking or is quietly ignored. If feedback research is going to be useful to teachers, it needs to move beyond describing inputs and start explaining mechanisms.

Teachers keep learning throughout their careers and 15 years in, their pedagogical knowledge is still growing.

A longitudinal study published in Contemporary Educational Psychology tracked teachers’ pedagogical and psychological knowledge over 15 years in the profession. The finding challenges a common assumption embedded in much of the professional development literature: that teacher knowledge plateaus early and that experienced teachers are merely refining habits rather than acquiring new understanding.

The study found measurable growth in pedagogical knowledge sustained well into mid-career. This is not the same as saying all teachers improve, or that experience alone is sufficient. But it does suggest that the profession’s knowledge base is not a fixed endowment determined by initial training. For school leaders, the implication is that experienced teachers may be significantly undervalued as sources of pedagogical insight — and that professional development models which treat all teachers beyond five years as essentially equivalent may be missing real differences in accumulated expertise.

Which teacher characteristics actually predict instructional quality? A meta-analysis finds the answer is more nuanced than “years of experience.”

A meta-analysis in Learning and Individual Differences synthesised the evidence on which teacher attributes most strongly predict effective instruction and student learning. The study examined a range of characteristics — content knowledge, pedagogical knowledge, beliefs, self-efficacy, personality — and assessed their relationships to independently measured instructional quality.

The findings resist easy summarisation, which is itself informative. No single characteristic dominates. Content knowledge matters, but its relationship to instructional quality is weaker than many assume. Pedagogical knowledge shows stronger and more consistent associations. Teacher beliefs and self-efficacy show modest but reliable effects. The pattern that emerges is of a complex, multi-dimensional skill that cannot be reduced to any single variable. For those designing teacher selection or development programmes, this is both a warning against silver-bullet thinking and an argument for broad-based approaches that develop multiple dimensions of teacher expertise simultaneously. What makes a teacher effective is not one thing. It is many things working in concert — and isolating any single factor risks mistaking a thread for the tapestry.

Technology, AI, and Learning

An AI tutoring system that personalises the order and difficulty of practice problems produced meaningfully better exam outcomes than a standard fixed sequence across 770 high school students over five months.

Most AI tutoring platforms are reactive: they wait for students to ask questions and then respond. The research team argues this paradigm misses something fundamental about how learning works. Drawing on mastery learning, they designed a system that actively guides what problems students attempt next. The core innovation is a reinforcement learning algorithm that estimates each student’s knowledge state using not just answer correctness but the quality of student-chatbot conversations and the nature of code editing behaviour. This richer signal feeds into an algorithm that dynamically selects the difficulty level of the next practice problem. Across ten Taipei high schools over five months, 770 students were randomly assigned to either this adaptive sequencing condition or a fixed easy-to-hard sequence. The adaptive group outperformed the control group by 0.15 standard deviations on an in-person, unassisted final examination, a pre-registered finding that held across multiple regression specifications.

For educators, the implications are encouraging. The study does not simply show that AI tutoring works; it shows that a specific mechanism, personalised problem sequencing, adds meaningful value over and above a generic AI chatbot that all students had equal access to. The control group was not deprived of AI support; they had the same chatbot and the same content, just in a fixed sequence. This is an important comparison condition, and it matters. Crucially, the gains were not achieved by making students do more work or spend longer on the platform in absolute terms. The mediation analysis suggests the benefit was carried through increased engagements, students in the adaptive condition spent more time on task and made more attempts per problem, and they interacted with the chatbot in qualitatively more productive ways. For school leaders considering AI tutoring deployments, the practical implication is that what you sequence matters as much as what you provide, and that keeping students productively challenged, not overwhelmed, not bored, appears to be the active ingredient.

Does AI-powered note-taking help students learn from video lectures, or does it quietly do the cognitive work that would have built understanding?

A study in Contemporary Educational Psychology examined what happens when students use AI-powered note-taking tools while watching STEM video lectures. The researchers assigned students to different conditions — some took notes with AI assistance, others without — and measured both learning outcomes and students’ perceptions of how helpful the tool was.

The study adds to a growing body of evidence that suggests the relationship between AI assistance and learning is not straightforward. The concern is one that will be familiar to anyone who has followed the cognitive offloading literature: when a tool reduces the effort required to process information, it may simultaneously reduce the depth of encoding that makes learning durable. The note-taking context is particularly telling because note-taking has always been understood as a generative activity — its value lies not in the notes produced but in the cognitive processing required to produce them. If AI handles the processing, the student receives a better product but may undergo a weaker learning experience. The question for instructional designers is whether AI note-taking can be configured to augment rather than replace the generative work — or whether the efficiency gains and the learning losses are inseparable features of the same mechanism.

When people know content was produced by AI, they trust it less — even when it is competent. What does this mean for AI-assisted learning?

A study published in Computers in Human Behavior: Artificial Humans examines what the authors call the “AI Penalty”; the systematic reduction in trust, perceived authenticity, and knowledge uptake that occurs when people are told content was generated by artificial intelligence. The paradox is that transparency about AI involvement, which we might expect to build trust, actually undermines it. Identical content is rated as less credible, less authentic, and less persuasive when attributed to AI rather than a human author.

For education, this finding cuts in two directions. On one hand, it suggests that student scepticism toward AI-generated feedback or explanations may be a real barrier to learning, even when the content is accurate. On the other, it raises an uncomfortable question about what happens when AI assistance is not disclosed. If students learn equally well from AI-generated content when they believe it was written by a human, but less well when they know its provenance, the disclosure paradox creates a genuine dilemma for educators committed to both transparency and effectiveness. The study does not resolve this tension, but it names it precisely — and that is a useful contribution.

From “AI Psychosis” to “Brain Rot” — how pseudo-diagnoses may be crowding out genuine understanding of technology’s effects on the mind.

A review in Computers in Human Behavior argues that the proliferation of informal diagnostic labels for technology-related mental states; “AI Psychosis,” “Brain Rot,” “doomscrolling disorder”, is actively impeding legitimate psychological research. The authors contend that these labels, amplified by social media and news coverage, create an illusion of scientific understanding where none exists. A catchy name is not a diagnosis. A trending term is not a construct.

The relevance for education is direct. The history of educational psychology is littered with pseudo-scientific concepts that were adopted because they were intuitively appealing rather than empirically grounded — learning styles, digital natives, the 10,000-hour rule. Each acquired diagnostic authority long before the evidence warranted it, and each proved remarkably difficult to dislodge once embedded in professional practice. The authors’ concern is that the same pattern is repeating with AI-related psychological claims: we are naming phenomena before we understand them, and the names themselves are shaping how we think about the underlying reality. For educators navigating the AI conversation, the paper is a useful reminder that the impulse to pathologise is not the same as the work of understanding.

Can large language models actually teach creativity? New research suggests cognitive gains but affective gaps and the distinction matters.

A study in Computers in Human Behavior: Artificial Humans investigated whether LLMs can enhance creative thinking. The researchers measured both cognitive outcomes, divergent thinking, idea fluency, originality, and affective responses, while also tracking neural patterns during the creative process. The headline finding is that LLM interaction did produce measurable improvements on standard creativity metrics. Students generated more ideas, and those ideas were rated as more original by independent evaluators.

But the affective data tells a different story. The subjective experience of creating with AI assistance felt qualitatively different from unassisted creation. The emotional engagement, the sense of ownership, the felt difficulty of the work, all were diminished. This is a finding that resonates well beyond the creativity literature. It echoes the pattern visible in the cognitive offloading research: AI can improve measurable outputs while simultaneously degrading the quality of the experience that produces them. Whether this matters depends on what you think education is for. If the goal is to maximise the quality of the product, AI assistance is clearly beneficial. If the goal is to develop the capacity of the person, the answer is considerably less clear — because the struggle that AI removes may be precisely the struggle through which creative capacity is built.

Most adaptive learning systems adapt to generic signals like accuracy and speed. A new editorial argues this misses the point: what matters is the structure of the domain being taught.

An editorial introducing a special issue of Learning and Instruction makes a deceptively simple but consequential argument: adaptive learning systems cannot be truly adaptive if they are domain-blind. Strohmaier, Depaepe, Nickl, and Obersteiner argue that most current approaches to adaptivity — whether in AI tutoring platforms or human-led differentiation — rely on generic learner metrics: accuracy, response time, time on task. These signals tell you that a student is struggling. They do not tell you why, and the why is almost always domain-specific.

The authors identify three dimensions where this matters. First, what to measure: the relevant learner variables differ by subject. A meaningful error in fraction arithmetic — confusing the denominator with a whole number, say — is categorically different from a meaningful error in source evaluation during a history lesson. A system that treats both as “incorrect” has already discarded the most useful diagnostic information. Second, how to measure it: the same observable behaviour carries different meanings in different domains. A long pause before responding might signal productive mathematical reasoning in one context and a complete vocabulary gap in another. Third, how to respond: the instructional move that follows from a diagnosis should be shaped by the conceptual structure of the domain, not merely by a generic adjustment to difficulty level.

The special issue includes nine empirical studies across different subjects that collectively illustrate both the promise and the difficulty of building domain-specific adaptivity. For educators evaluating AI tutoring tools, the editorial offers a useful diagnostic question: is this system adapting to what my students are thinking, or merely to whether they are getting things right? The difference is the difference between a thermostat and a diagnostician. One adjusts to a signal. The other understands a system.

Motivation, Self-Regulation, and Individual Differences

Not all motivation fluctuates equally: achievement goals shape not just how motivated students are, but how much their motivation swings from day to day.

A study in Learning and Individual Differences moves beyond the question of average motivation levels to examine something more granular: daily variability. Using intensive longitudinal methods, the researchers tracked students’ intrinsic motivation across multiple days and examined how different achievement goal orientations — mastery goals, performance goals, and their various subtypes — predicted both the level and the stability of that motivation.

The finding that matters most for practitioners is that some goal orientations produce motivation that is not only higher on average but more stable over time, while others produce motivation that is volatile — high one day, absent the next. For teachers, this is a useful reframe. It suggests that the question is not simply “how do I motivate my students?” but “how do I produce motivation that persists?” A student who is intensely motivated on Monday and entirely disengaged on Wednesday may average out to “moderately motivated” on any standard measure, but the instructional challenge they present is fundamentally different from a student who is steadily, modestly engaged all week. The study suggests that the type of goal a student pursues may matter as much for motivational stability as for motivational intensity.

Adolescent self-regulated learning is not a single dimension and the profiles that emerge predict achievement in ways that simple averages miss.

A study in Learning and Individual Differences used latent profile analysis to identify distinct patterns of self-regulated learning among adolescents. Rather than treating self-regulation as a continuum from low to high, the researchers looked for qualitatively different profiles: students who might be strong on planning but weak on monitoring, or effective at metacognition but poor at time management.

The profiles that emerged were meaningfully different from one another, and they predicted academic achievement more accurately than any single self-regulation measure alone. They also showed associations with sociodemographic factors, suggesting that self-regulation profiles are not purely individual traits but are shaped by the contexts in which students learn. For teachers, the practical implication is that “teach students to self-regulate” is too blunt an instruction. Different students may need support with different components of the self-regulatory process, and a student who appears to lack self-regulation may in fact have a specific deficit in one area while being competent in others. The profile approach suggests that diagnosis should precede intervention — and that the diagnosis needs to be more fine-grained than a single score on a self-regulation questionnaire.

Does metacognition actually drive improvement in maths over time, or does it just track who is already doing well?

This longitudinal study followed 334 junior high school students in China over two and a half years, measuring both their metacognitive development and their mathematics achievement every six months. The results show a surprisingly uneven pattern. Metacognition rose in the first year, then declined across the following two years. More importantly, students clustered into three distinct developmental groups: those with consistently high and slowly rising metacognition, those with moderate but declining levels, and those with low and sharply declining levels. These trajectories were mirrored, in complex ways, in their maths attainment.

The key finding is subtle but important. Students with higher metacognition tended to have higher maths achievement at the outset, but changes in metacognition did not strongly predict changes in achievement over time. In other words, metacognition seems to be more of a marker of who is already doing well rather than a clear driver of improvement. For teachers, this challenges the common assumption that simply “teaching metacognition” will automatically raise attainment. It suggests that metacognitive skill sits alongside knowledge and achievement rather than acting as a straightforward lever to accelerate it.

Subscribe now

One of the things that cognitive science keeps discovering about learning is that much of the most important work happens invisibly, beneath conscious awareness, in the silent labour of long-term memory consolidating and connecting representations. We tend to think of reading as a visual act: the eye meets the word on the page, the brain decodes it, and meaning follows. But a growing body of research suggests that by the time a child encounters a word in print for the first time, their mind has already been preparing for that encounter. The bones of the word’s spelling have been quietly assembling themselves in memory, guided by nothing more than the sound of the word spoken aloud.

A new paper I’ve just read has introduced me to a fascinating term which describes this process in reading and spelling. The study by Ataman, Beyersmann, Castles, and Wegener, explores a simple question; when we hear a new word, do we start forming a guess about how it is spelled before we ever see it written down? The authors call these guesses “orthographic skeletons”. Using a novel word learning paradigm, adult readers were trained on spoken versions of made up words that included either inflected forms (e.g. past tense) or derived forms (e.g. agent nouns). They were then tested on how quickly and efficiently they processed the written stems of those words for the first time.

The Ghost Before the Word

The concept, first proposed by Wegener and colleagues in 2018, is based on a deceptively simple insight: when a child learns a new word orally, hearing it spoken, understanding its meaning, using it in conversation, their knowledge of how sounds map onto letters (phoneme-grapheme correspondences) allows them to generate an expectation about how that word might be spelled. Not a complete, fully formed spelling, but a partial sketch; a skeleton.

To test this, researchers use invented nonsense words like "vish" or "jayf," words that no participant has ever encountered before, so the experimenters can be certain that any spelling expectations were formed purely from oral training rather than prior reading experience. So take a reader who has been taught the spoken word "vish," its meaning and its use in sentences, but has never seen it written down. Their knowledge of English spelling patterns tells them that the /v/ sound is typically written as "v," the /ɪ/ sound as "i," and the /ʃ/ sound as "sh." Without ever seeing the word in print, they have already begun to assemble its orthographic form.

When that reader later encounters “vish” written on a page, the word is not entirely novel. It arrives into a cognitive space that has been prepared for it, a space where expectation meets confirmation. The result, demonstrated across multiple studies using both lexical recognition tasks and eye-tracking, is faster processing, shorter fixation times, and more efficient reading. The skeleton has done its invisible work.

What makes this more than a laboratory curiosity is the growing evidence for its robustness. Wegener et al.’s original 2018 study demonstrated the effect in Grade 4 children. Beyersmann et al. (2021) replicated it in adults using eye-tracking during sentence reading and extended it to morphologically complex words; those with inflectional endings like “vished” or “vishing.” Beyersmann et al. (2022) then found the same skeleton effect in children in Grades 3 to 5.

The previously mentioned new study pushes the finding further still. Using the same paradigm with over 160 adult readers across two experiments, one using a lexical recognition task measuring response latencies and the other using eye-tracking during sentence reading, they show that orthographic skeletons form not only when participants are trained on inflected forms but also when trained on derived forms (e.g., "visher," "vishist"). Crucially, the size of the skeleton effect did not differ between inflected and derived conditions. The absence of that difference is itself a key finding: it suggests the mechanism operates with equal facility regardless of morpheme type.

This last point is particularly interesting I think. Inflectional affixes like "-ed" and "-ing" are frequent, productive, and grammatically transparent. Derivational affixes like "-er" and "-ist" are less common, change the syntactic category of the word, and alter its meaning more substantially. One might reasonably expect that the more familiar inflectional forms would produce stronger skeletons. But they do not: the effect is equivalent.1

This suggests that the mechanism is not dependent on the familiarity or regularity of the affix. The mind is not merely memorising high-frequency patterns; it is generating predictions based on its deep knowledge of the sound-spelling system. As Stuart and Coltheart proposed as early as 1988, readers draw on their phoneme-grapheme knowledge to form expectations about the spellings of words they have heard but never seen. The skeleton is not a snapshot; it is an inference.

The Hidden Return on Vocabulary Instruction

Now dear reader, a caveat is important here. There is a huge limitation with this paper which needs to be addressed. The Ataman et al. study, like Beyersmann et al. (2021), uses adult skilled readers in a tightly controlled novel-word paradigm. It does not directly demonstrate that young children form orthographic skeletons for derived words. However, the child evidence from Wegener et al. (2018) and Beyersmann et al. (2022) shows the skeleton effect for simple and inflected words in primary-age readers, and the broader programme of research on the causal relationship between oral vocabulary and reading acquisition converges on the same, if not very similar conclusion. The adult paradigm just isolates the mechanism with more precision; the child studies confirm it operates in developing readers. Together I’d argue, they present a compelling picture, even if the full story is still being assembled.

The important implications for teaching converge on a point that the science of learning has been making for decades: knowledge in long-term memory is the engine of further learning.

Consider a primary school classroom where children are being taught about the water cycle. They hear the word "condensation" used repeatedly in context: the teacher explains it, models it, returns to it across lessons. They hear "evaporation" and "precipitation." These are words most seven-year-olds will not encounter in print for some time. But if the orthographic skeleton hypothesis is correct, those oral encounters are not merely building scientific vocabulary; they are laying down orthographic scaffolding. The child's phoneme-grapheme knowledge is quietly sketching the spellings of words they have never read, preparing the ground for a future encounter with text that will, when it arrives, be measurably easier to process.

This reframes to a certain extent, the relationship between oral vocabulary and reading. We have known for some time that the correlation is strong. Duff and Hulme (2012), Lee (2011), Nation and Snowling (2004), and many others have shown that early oral vocabulary is a significant predictor of later reading ability. Novel word training studies have demonstrated the causal direction: teaching a child the spoken form of a word facilitates the reading of that word when it is first encountered in print (Castles and Nation, 2006; Ricketts et al., 2016). But the orthographic skeleton hypothesis gives us the mechanism underneath the correlation. It tells us how oral vocabulary feeds reading: not through some vague process of “language enrichment,” but through the specific, measurable generation of spelling expectations in long-term memory.

Phonics Without Vocabulary is Half a Bridge

This finding also illuminates something that the reading wars, in all their acrimony, have often obscured: systematic phonics instruction and rich oral vocabulary are not competing priorities. They are two halves of the same cognitive mechanism, operating at different timescales and converging at the point where the orthographic skeleton is formed.

Phonics gives children the grapheme-phoneme correspondence rules: the systematic knowledge that /sh/ maps to “sh,” that /tion/ at the end of a word is typically spelled “-tion,” that a long vowel before a final consonant often signals a silent “e.” This knowledge is the skeleton’s building material. Without it, no amount of oral exposure will generate spelling expectations, because the child has no principled way to translate sound into letter.

But vocabulary gives them the words those rules will be applied to. A child who has been taught the phonics rules but possesses a thin oral vocabulary has the decoder but nothing to decode towards. The skeleton cannot form because there is no oral representation to generate expectations from. The child knows that /k/ maps to “c” or “k,” but has never heard the word “catastrophe,” so no skeleton of that word’s spelling exists in memory. When they later encounter it in a text, they must build the word from scratch; every letter a surprise, every syllable an act of effortful assembly.

Conversely, a child with a rich oral vocabulary but no systematic phonics instruction has the raw material for skeletons but no reliable way to construct them. They have heard “magnificent” a hundred times, but without systematic knowledge of how sounds map to letters, the mind cannot sketch even a partial spelling. The oral representation sits in memory, rich and available, but orthographically inert.

The skeleton forms at the intersection of these two bodies of knowledge. It is the product of phonics and vocabulary working in concert. This is why, as the evidence from Ataman et al. and the broader programme of research demonstrates, the effect is strongest for words with predictable spellings in the trained condition: the match between what the phoneme-grapheme rules predict and what the eye subsequently encounters is what produces the facilitation. When the spelling is unpredictable (when “jayf” is spelled “jayf” rather than the expected “jaif”), the skeleton actually creates a mismatch, a momentary surprise that the eye registers as longer fixation times and slower recognition. The skeleton, in other words, is not a passive residue of oral learning. It is an active prediction, and like all predictions, it can be confirmed or violated.

The Gap That No Reading Test Detects

There is a darker implication here, one that bears directly on questions of equity which I’ve touched on before. If orthographic skeletons are real, (and the converging evidence from multiple research groups using different paradigms, tasks, and populations suggests they are), then vocabulary-rich children arrive at school with a hidden cognitive advantage that no reading test will detect. They have heard “ridiculous” and “extraordinary” and “investigation” at the dinner table, in bedtime stories, in the overheard conversations of articulate adults. Their minds have been silently sketching the spellings of hundreds of words they have never read. When they open a book in Year 1 or Year 2, they are not starting from zero; they are reading into a prepared landscape.

Children from language-poor environments arrive without those skeletons. Their oral vocabularies are thinner, not because they are less capable, but because they have had fewer opportunities to hear and use complex language. When both groups encounter the same text, one group is reading words their minds have been quietly preparing for. The other is meeting each word cold, with no prior expectation, no skeleton to ease the recognition process.

This is the Matthew Effect in reading, but observed at a level of granularity that the usual discussions of “word gaps” and “literacy-rich environments” rarely achieve. The gap is not just one of exposure in the vague sense; it is a gap in the number of pre-formed orthographic representations stored in long-term memory. It is a gap in prediction. And it compounds: the child who reads more easily reads more, hears more words in the context of text, forms more skeletons, and reads still more easily. The child who struggles reads less, encounters fewer new words, forms fewer skeletons, and falls further behind.

This makes the case for early, intensive, explicit vocabulary instruction in schools with a precision that hand-waving about “word-rich environments” cannot match. It is not enough to surround children with language and hope for the best. The orthographic skeleton is not formed by ambient exposure; it is formed by the interaction of specific oral knowledge with specific phoneme-grapheme knowledge. Both must be taught, and both must be taught with the deliberate, sequenced, cumulative logic that characterises effective instruction.

What the Mind Does Before the Eyes Arrive

The thing that I think is so interesting about this idea of the orthographic skeleton is that it tells us that reading does not begin when the eye meets the page. It begins earlier, in the ear, in the mouth, in the rich associative networks of a mind that has been listening and speaking and building representations of language long before it encounters its written form. The skeleton is a testament to the generative power of well-organised knowledge: give the mind enough of the right building blocks, and it will begin constructing things you never explicitly asked it to build.

This is, in miniature, the story of all learning. The student who knows a great deal about the French Revolution does not merely recall facts; they generate inferences, make predictions, notice patterns that a novice cannot see. Their knowledge is not a static warehouse but a living architecture, constantly projecting forward, anticipating what comes next. The orthographic skeleton is simply this same principle, operating in the domain of written language: prior knowledge does not wait to be activated. It reaches out to meet the world halfway.

For instructional design, the practical takeaway is both simple and far-reaching. Every science lesson in which children hear and use the word “photosynthesis,” every history lesson that introduces “parliament” and “revolution,” every story in which a character is described as “reluctant” or “courageous” is doing double duty. It is building domain knowledge, yes. But it is also, invisibly, preparing the child to read those words when they next appear on a page. The skeleton is being formed. The bones are being laid down.

A key question for schools is whether we recognise this invisible work and design our instruction accordingly, or whether we continue to treat vocabulary and reading as separate line items on a curriculum plan, unaware that the mind has long since woven them together.

Three decades later, B.F. Skinner refined this intuition into something more sophisticated. He called it a teaching machine which advanced one small frame at a time, each requiring a correct response before the next would appear. But the premise was the same: arrange the material so that the student is almost always right, and learning will take care of itself. Again, correctness, in Skinner’s framework, was the learning. The reinforcement of right answers would shape behaviour, and shaped behaviour was all that mattered.

This idea still runs beneath a lot of adaptive learning software, many gamified apps, and AI tutoring systems that treat accuracy as both the measure and the engine of understanding. But a recent study suggests that this idea may have had the relationship exactly backwards. Correctness shows up after learning has already occurred. When learning is happening at the edge of understanding, correctness is often invisible.

The Study: When Failure Became a Signal

In the study, the researchers took a language model and gave it a set of mathematical problems it could not solve. Not inconsistently, it failed completely. They then tried to improve its performance by training it directly on those problems. Nothing happened. The model was stuck: the problems were too hard to provide any useful signal, and without signal, there was no learning.

What came next was the interesting part. Rather than simplifying the problems or introducing easier ones, the researchers designed a framework in which the model generated its own intermediate tasks. These were not easier versions of the target problems, but structurally related ones that shifted in content and style as the learner improved: from word problems and basic formulas to concise, equation-heavy exercises in algebra and calculus.

Most of the answers to these problems were wrong. Accuracy remained low. And yet, learning restarted, and did so rapidly. Performance on the original hard problems began to improve, even though the model could not solve them at the outset and direct training on them alone had produced no progress.

If correct answers were the mechanism of learning, this result would make little sense. What mattered was not success, but rather exposure to structure. The stepping-stone problems revealed something about the organisation of the domain: the kinds of relations that recurred, the representations that mattered, the transformations that preserved meaning. Even incorrect attempts reshaped the model’s internal representations in a productive direction. Correctness arrived later, as an outcome rather than a cause.

Perhaps the most interesting detail in the study is this: the “teacher” models that generated the intermediate problems were themselves largely wrong. Only 32.8% of their solutions were fully correct. And yet the “student” models still learned. As long as the problems were well-posed and conceptually relevant, they provided what the researchers call a “learnable gradient”: enough structure for the learner to gain traction. The accuracy of the answer mattered far less than the architecture of the question.

There is also a telling comparison buried in the results. A separate group of teachers, trained to optimise for a different signal, actually produced more correct answers: 55% accuracy. Yet their students learned less. Their questions collapsed into a narrow conceptual range, and in one case caused complete student failure. More correct answers, worse learning. The relationship between correctness and learning is not merely loose; under certain conditions, it inverts entirely.

This is a finding that I think is really interesting for instructional design because it suggests that learning happens not by mimicking a perfect answer, but by engaging with the right conceptual structure. A well-constructed problem, even one accompanied by a flawed solution, can open up a region of understanding that a correct but poorly structured exercise never touches. Structure beats accuracy. In other words, the map matters more than the destination marker.

The Productive Failure Problem

This also helps clarify a persistent confusion in education around the idea of productive failure. Repeated failure on the original hard problems produced no learning signal at all. It was only once the learner encountered problems that were, in some meaningful sense, learnable that failure became informative. Learning did not arise from struggle per se, but from struggle within a structure that could be partially apprehended. Struggle without structure is not a desirable difficulty. It is simply noise.

This distinction matters because the debate around productive failure has often been poorly drawn. As has argued, the research base for productive failure is weaker than its advocates suggest, and the ethical implications of intentionally causing students to struggle are rarely examined. He is right on both counts. Intentionally designing lessons around failure, without specifying the conditions under which that failure becomes informative, is not a pedagogy; it is an abdication. The students most likely to flounder are precisely those who are already least advantaged: the ones who bring the fewest resources to a problem and stand to lose the most from another experience of not knowing. As Ashman puts it, intentionally causing struggle and then mitigating the negative effects is “like intentionally making a hole in the bottom of a row boat and then grabbing a bucket to bail out the water.”

At the Edge of Learnability

The study reinforces this concern: unstructured failure on problems beyond the learner’s reach produced nothing at all. But it also reveals something the productive failure debate has largely missed. What restarted learning was not success, nor explicit instruction, nor even easier problems. It was problems with the right internal structure: problems that made the organisation of the domain partially visible, even when the learner could not yet solve them. The question is not whether students should struggle. It is whether the task they are given has enough conceptual architecture to make that struggle legible.

Seen this way, the concept of scaffolding also sharpens. Scaffolding is often misunderstood as making tasks easier, or as providing hints, encouragement, or additional support. But none of these capture its essential function. A scaffold is not a ramp; it is a lens. It changes what the learner can see and work with. It makes the structure of the domain visible and usable, altering what the learner can do later, not how well they perform now.

The paper also exposes a limitation in many AI tutoring systems, and in more than a few classrooms. These systems adapt to performance indicators: correctness, speed, confidence etc. Yet at the edge of learnability, these signals are largely meaningless. Learning is either beginning or failing to begin long before performance becomes stable or interpretable. A system that optimises for immediate accuracy at the frontier of a student’s understanding is like a gardener who judges growth by pulling up the seedling to check the roots.

A genuinely intelligent tutoring system would not optimise for immediate success. It would optimise for learning trajectories: the slow, often invisible reshaping of representation that precedes competence. It would tolerate incorrect answers when they indicate progress in understanding, and it would judge instruction by delayed transfer rather than local accuracy. The 32.8% figure from this study is a provocation: it implies that the quality of the question a system poses may matter more than whether it can supply the correct answer.

What Correctness Cannot Tell Us

For teachers, the lesson is familiar, but still worth stating plainly. Good instruction does not aim to maximise success in the moment. It aims to make success possible later. As notes, the purpose of feedback is to improve the student and not the work.

Rosenshine was right that effective teachers obtain a high success rate, but the art lies in understanding what makes that success rate possible: not the avoidance of difficulty, but the careful construction of tasks whose structure a learner can get hold of. The right problem, at the right time, with enough architecture to make the struggle productive.

I have written before about what happens when correct answers arrive without learning: watching my daughter completing a phonics puzzle by matching colours, producing perfect scores while the cognitive architecture beneath remained untouched. This study reveals the inverse: what happens when learning arrives without correct answers. Together, they suggest that correctness and learning are far more loosely coupled than our systems assume. Correct answers tell us what learning has already achieved. They tell us very little about how to bring it about.

Skinner's teaching machine was an intricate piece of engineering in service of a half-truth. Correctness is not irrelevant; it is the destination. But it is not the mechanism. In 1929, the Knapp Electric Questioner offered a closed circuit for the right answer and silence for the wrong one. Nearly a century later, most of our educational technology still operates on the same principle. The SOAR study suggests we have been listening for the wrong signal. Learning does not announce itself with a light. It begins in the silence, in the space where the circuit has not yet closed but the structure of the problem is slowly becoming visible.

Subscribe now

Bainbridge’s central argument was this: the more sophisticated an automated system becomes, the more demanding, not less, the human role within it. She called this the “ironies of automation”, and the irony she had in mind was structural rather than incidental. Designers of automated systems, she observed, tend to view the human operator as the weak link: unreliable, inconsistent, prone to fatigue and error. The aim of automation is therefore to eliminate the human wherever possible.

But the designer who tries to eliminate the operator still leaves the operator to do the tasks which the designer cannot think how to automate. What remains after automation is not a simplified role but an arbitrary residue of the most demanding, most ambiguous, and least supported work in the entire system. The human is not replaced. In other words, the human is paradoxically left with the hardest parts, and given almost no preparation for them.

Forty years on, Bainbridge’s paper reads less like a historical document than a prophecy. It describes, with an accuracy that might trouble us, exactly what is now happening to knowledge workers across every sector in which AI has taken hold. And it raises, with particular accuracy, a question that education has barely begun to confront.

Work Doesn’t Shrink. It Expands.

Two recent studies, published within weeks of each other in the Harvard Business Review, have now supplied the empirical substance that Bainbridge anticipated over forty years ago1. The first, by Aruna Ranganathan and Xingqi Maggie Ye of UC Berkeley, followed around two hundred employees at a technology company over eight months, tracking in granular detail how the introduction of generative AI tools changed the texture of their working lives. The results were not what the productivity optimists had promised. AI did not reduce work. It intensified it, consistently and across every role the researchers examined.

Workers expanded their remit because AI made previously inaccessible tasks feel suddenly tractable. Parkinson’s Law holds that work expands to fill the time available; what this study suggests is a more unsettling corollary: work also expands to fill the capacity available. They blurred the boundaries between work and rest because prompting an AI tool felt less like working than like chatting, and so slipped imperceptibly into lunch breaks, evenings, and the margins of meetings. They took on more tasks simultaneously because AI gave them a sense of momentum, of always having a capable partner at hand. And as a result, they found themselves doing more at once, feeling more pressure, and experiencing less genuine recovery than before they had adopted the tools that were supposed to liberate them. As one engineer in the study put it with disarming honesty: the expectation had been that greater productivity would mean working less. What actually happened was working the same amount, or more.

The Cognitive Governors

Part of what drives this dynamic is the disappearance of friction. Before AI, the effort required to begin a task served as a natural governor on the number of tasks undertaken. The blank page, the research required before drafting, the time needed to think through a problem from scratch: all of these imposed a rhythm on cognitive work, structuring attention and creating the stopping points that allowed genuine recovery. When AI lowers the cost of beginning almost anything, those governors are removed. The scope of work expands not because anyone demands it but because the tools make expansion feel possible, even effortless. What disappears is not effort but the friction that once gave effort its shape.

This is Bainbridge’s first irony, translated from the control room to the open-plan office. The system absorbs the tractable parts of the job and returns to the human a workload that is larger in volume, broader in scope, and denser in the kind of high-stakes, high-attention judgement that cannot be delegated. The work does not diminish. It concentrates.

The Workslop Problem: Why Oversight Is Harder Than It Looks

There is a particular kind of modern frustration that many of us will recognise since 2023 which had no name until recently. It looks like this; you receive a report. It is long, it is formatted, it has an executive summary and a set of recommendations and a conclusion that circles back to the introduction with apparent tidiness. You read it. You realise, slowly, that it has told you nothing. Someone has produced it with AI without thinking about it, and now the thinking has been transferred to you.

This phenomenon underpinning this is explored in the second study, conducted by a research team at Boston Consulting Group and published in early March 2026. Surveying nearly fifteen hundred full-time workers across industries, roles, and seniority levels, the researchers found that intensive oversight of AI tools was the single most mentally taxing form of engagement their participants described. Workers required to monitor AI agents closely reported fourteen percent more mental effort, twelve percent more mental fatigue, and nineteen percent greater information overload than those whose AI engagement was less demanding.

A distinct phenomenon emerged from the data, which the researchers called “AI brain fry”: a state of acute cognitive exhaustion characterised by difficulty focusing, slower decision-making, and what participants variously described as mental fog, a buzzing sensation, or the feeling of having a dozen browser tabs open in one’s head simultaneously. Fourteen percent of those using AI at work reported experiencing it; among those in marketing and operations roles, the figure rose to over a quarter.

The consequences were not merely personal. Workers experiencing AI brain fry made major errors thirty-nine percent more frequently than those who did not. They were thirty-nine percent more likely to be actively seeking to leave their jobs. The cognitive tax was, in the most precise sense of the phrase, a business cost.

What makes this finding particularly interesting I think is the nature of the work that produces the fatigue. AI output does not arrive as raw material awaiting construction: it arrives looking finished. This is what Niederhoffer, Robichaux and Hancock at Stanford have termed workslop: AI-generated content that masquerades as good work but lacks the substance to meaningfully advance a given task. The particular cruelty of workslop is that it doesn’t announce its own inadequacy; it arrives fluent, confident, and formatted, offloading the cognitive labour of detecting its failures entirely onto the recipient. In other words, the person who did the least thinking ends up doing the least work, and the person who receives it ends up doing the most.

I think what makes this burden so difficult to manage is what has called the jagged frontier2 of AI capability: the fact that AI is immensely powerful on some tasks and fails completely or subtly on others, and that these failures bear no reliable relationship to how difficult or straightforward the task appears to be. The frontier is invisible, which means the human overseer cannot calibrate their vigilance to the actual risk; they must attend to everything with equal suspicion. For example, AI can produce a technically accurate summary of a research paper it has never read, and a technically plausible summary of a research paper that does not exist, and the two outputs will be indistinguishable to anyone who does not already know the answer.

A generated text is fluent and confident; generated code compiles and may even pass tests; a synthesised argument has the surface coherence of careful reasoning. This is precisely the condition under which what Parasuraman and Riley identified as automation bias takes hold: the tendency to over-trust automated output, following it when it is wrong and failing to notice when it has failed.

Their subsequent synthesis with Manzey, published in 2010, showed that this bias manifests in two distinct forms: omission errors, where the human fails to notice the system has gone wrong, and commission errors, where the human actively follows the system’s incorrect recommendation. Critically, neither form can be overcome by training or experience alone; automation complacency and bias are found in expert and novice users alike, because the underlying mechanism is attentional rather than epistemic. It is not that people lack the knowledge to catch errors; it is that the conditions of oversight suppress the vigilance required to deploy that knowledge.

The evaluative burden this places on the human is therefore heavier than it might initially appear. The user is not correcting an obvious gap but assessing something that presents itself with the authority of a completed product, searching for the subtle error, the misplaced assumption, the quietly wrong inference that the system’s fluency has carefully concealed. This is cognitively exhausting in a way that is quite distinct from the effort of original production; and it is, in the most uncomfortable sense, a task that demands precisely the expertise that the system was supposed to make less necessary. AI, in other words, assumes the very competence it risks undermining.

This is Bainbridge’s monitoring irony in its contemporary form. She had identified, with uncomfortable clarity, that asking humans to oversee automated systems is a near-impossible task: vigilance decays rapidly, the system’s internal logic is opaque to the overseer, and the human is held responsible for catching failures that the machine cannot flag and that the human has no reliable means of detecting. The BCG study confirms that this impossibility is not confined to nuclear operators watching instrument panels. It applies equally to the marketing manager reviewing AI-generated copy, the engineer checking code produced by a large language model, the teacher evaluating the output of an AI tutoring platform. The work looks like oversight. What it actually requires is a form of sustained, expert, hypervigilant attention that human cognition was simply not designed to sustain.

The Exhaustion of Evaluation

But there is a lived texture to all of this that the research papers, for all their rigour, cannot quite capture. Siddhant Khare, a software engineer who builds the infrastructure that powers AI agents in production systems, published an account of his own experience in February 2026 that functions as an inadvertent case study in Bainbridge’s thesis. He had shipped more code in the previous quarter than in any quarter of his career. He had also felt more drained than at any previous point in his working life. These two facts, he noted at the outset, were not unrelated.

What Khare described was not the exhaustion of creation but the exhaustion of evaluation: the relentless cognitive labour of reviewing outputs he had not produced, from a system whose reasoning he could not trace, whose errors were often subtle rather than obvious, and whose behaviour was fundamentally probabilistic rather than deterministic. Before AI, his job had been to think about a problem and build a solution. After AI, his job had become to prompt, wait, read, evaluate, decide, correct, and repeat; a quality inspector on an assembly line that never stopped. He found himself unable, by midweek, to make simple decisions. His working memory was not full of code: it was full of judgements about code.

The most unsettling passage in Khare’s account concerns what he calls “thinking atrophy”. During a design review meeting, asked to reason through a concurrency problem at a whiteboard without his laptop, without AI, he struggled. Not because he lacked the relevant knowledge but because he had not exercised that particular cognitive muscle in months. The capacity for unassisted first-draft thinking had quietly degraded, in direct proportion to how consistently he had outsourced it. This is precisely what Parasuraman and Riley described as the out-of-the-loop performance problem: operators removed from active engagement with a system lose not just practice but situation awareness, the accumulated, dynamic sense of what the system is doing and why, which is the very thing required to intervene effectively when it fails. He reached, without naming her, precisely Bainbridge’s deskilling insight: the skills that atrophy during smooth automated operation are exactly the skills required when the automation fails. The former expert, in the moment of crisis, has become the novice.

The Final Irony

Bainbridge concluded her 1983 paper with a characteristically dry observation that has become I think, even sharper since it was written. The final irony of automation, she wrote, is that the most successful automated systems, those with the rarest need for human intervention, are precisely the systems that require the greatest investment in human skill. The longer the machine runs without incident, the more degraded the human backup; and yet it is in those rare, high-stakes moments of failure that the human is most needed and least prepared.

The BCG study offers an unexpected confirmation of this from a completely different direction. When AI was used to eliminate genuinely repetitive, low-value tasks, workers in their sample reported lower burnout scores, more engagement, and more time for the kind of human connection and creative work that the technology was supposed to enable. The problem was not AI itself. The problem was the way AI had been deployed: not to remove toil but to expand scope, not to free attention but to multiply the demands upon it, not to simplify the human role but to intensify it while stripping away the natural governors, the pace constraints, the breathing room, that had previously made it sustainable.

Ranganathan and Ye put this plainly: without intention, AI makes it easier to do more but harder to stop. The natural tendency of AI-assisted work is not contraction but intensification, and the costs of that intensification accumulate quietly, in the degradation of judgement, the erosion of skill, the slow compression of recovery time, and the rising frequency of errors made by people who are producing more than ever and thinking less carefully than they know.

The Positive Case for AI in Education

However, the case against careless AI adoption is not the same as the case against AI itself, and the distinction matters enormously. The BCG data already contains, embedded within its warnings, the outline of a genuine positive case. When AI was used to eliminate genuinely repetitive, low-value tasks rather than to expand the scope of work indefinitely, burnout fell, engagement rose, and workers reported more time for the kind of human connection and creative thinking that gives professional life its meaning. The technology is not the problem. The absence of intentional design is.

Consider first the signal to noise problem in education. What do we actually know about student learning? In my experience, assessment in most schools operates like a kind of faulty compass: it produces readings, generates confidence, and often points consistently in the wrong direction. The granular, timely, diagnostic information that would actually improve instruction is precisely what our current systems cannot supply. Instead we have built vast administrative infrastructures around measurements that tell us remarkably little about what students actually know or can do.

The spreadsheet of predicted grades is one of education’s great epistemic embarrassments: a document that looks like knowledge but tells the teacher almost nothing. Grades are entered, targets are set, progress is RAG-rated, and the resulting data is presented with a confidence entirely disproportionate to its explanatory power. AI’s most under-appreciated contribution to education may not be instruction at all but measurement: the possibility of replacing these crude, lagging, heavily interpreted signals with something finer-grained, more timely, and more honest about what it does and does not know about the learner in front of it.

The second argument concerns scale. Teaching quality is the single most important in-school factor in student outcomes, and it varies enormously; a student in one classroom may receive instruction that is transformatively better or worse than a student in the classroom next door, for reasons that have nothing to do with the student and everything to do with the professional lottery of which teacher they happened to be assigned.

The children who most need the best teachers are, by every measure we have, often the least likely to encounter them: disadvantaged students are disproportionately taught by less experienced practitioners, in schools with higher turnover, in systems that have never found a credible answer to the distribution problem. That problem is not a resource problem or a political problem in the first instance; it is a human problem. Expertise of the kind that makes a great teacher cannot be manufactured on demand, franchised, or guaranteed to persist under the conditions of a real school. AI does not solve this problem. But it may be the first tool that meaningfully addresses it: not by replicating the best teachers, which it cannot do, but by raising the floor of curriculum design, instructional quality and assessment in the classrooms where that floor is lowest.

I first heard this scaling argument from Joe Liemandt, when I met him in Austin in December last year; I found it uncomfortable in the way that only true things are uncomfortable which challenge long-held beliefs. If the system is dependent on those 1 in 20 superhero teachers who burn out after five years, we should at least be honest that what we are running is not a scalable model of education but a moral subsidy paid for by the most dedicated people in the profession.

How AI Can Further The Science of Learning

There is a third argument, less discussed but in some ways the most exciting of all (for me at least), in terms of the science of learning. AI does not merely offer a new delivery mechanism for existing instructional content; it offers, for the first time, the computational infrastructure to take the science of learning seriously at the level of the individual learner. Consider spaced repetition, one of the most robust findings in memory research: the evidence that distributing practice across time, rather than massing it, produces dramatically stronger long-term retention. The principle has been known since the nineteenth century; it has been replicated hundreds of times; and it has been almost entirely ignored by faculties of education and academics, in favour of a revisionist formulation of people like Vygotsky or Piaget.

But as strong as the evidence is, the reality is that we have decades of research from laboratories and post grads but as Tom Perry showed in 2021, almost nothing from real teachers, with real students in real classrooms. One reason why we struggle to implement retrieval and spacing in schools is for the simple reason that calculating the optimal review interval for each piece of knowledge, for each learner, at each moment in time, is computationally impossible for a human teacher managing a class of thirty.

However, projects like FSRS4Anki, an open-source scheduler that uses machine learning to model the precise dynamics of an individual's memory and predict the exact moment a piece of knowledge is about to be forgotten, represent something genuinely new: not the replacement of learning science by technology, but its realisation through it. For decades, the gap between what cognitive science knew about memory and what classrooms could actually implement has been an almost impossible to implement at scale. AI is the first tool capable of possibly addressing that.

That positive case becomes considerably more compelling when set against the historical baseline. Despite the fact that the most cited tutoring study from the Edtech world, Bloom’s 2 Sigma is actually based on really poor evidence (see here for more on this), more recent evidence provides a better foundation. a landmark 2011 meta-analysis, Kurt VanLehn reviewed decades of research comparing human tutoring, intelligent tutoring systems, and conventional classroom instruction. His findings were interesting: human one-to-one tutoring produced an effect size of around 0.76 over whole-class teaching, a substantial and consistent advantage that held across domains, age groups, and study designs. Crucially, intelligent tutoring systems achieved an effect size of around 0.40: roughly half the benefit of human tutoring, but still a meaningful and reliable improvement over conventional instruction. Even if there is a measurement error that reduces the true effect considerably, the convergence across studies and contexts suggests something real is being captured, not manufactured.

For most of human history that kind of individualised instructional support has been available only to the privileged few, a luxury of wealth and social capital that no amount of pedagogical goodwill could redistribute at scale. AI tutoring systems powered by the science of learning, even with their real limitations and the well-documented voltage drop between controlled trials and classroom deployment, represent a genuine attempt to offer something approaching that experience to every learner regardless of background. The equity argument for well-designed AI in education, cautiously and precisely stated, is genuine; and it is, for those who care about educational justice, a powerful one. To dismiss the technology entirely is to dismiss the possibility it carries for the children who have historically been most failed by the systems we already have.

The other thing which is, again, counterintuitive is that a model like this does not mean less human interaction but more. has spent decades arguing that the most powerful intervention in education is not any particular curriculum or technology but the quality of the feedback loop between teacher and learner: the moment a teacher notices where understanding has broken down and responds in real time with the precisely calibrated question or explanation that moves the student forward. That moment requires something AI cannot replicate: genuine knowledge of the individual, accumulated over time, in a relationship of trust. What AI can do is remove the hundred smaller tasks that currently prevent teachers from arriving at that moment with sufficient attention and energy to make it count.

The positive case for teachers is equally compelling, provided the crucial condition is met. If AI absorbs the most mechanical elements of formative assessment: flagging routine errors, generating retrieval practice at appropriate intervals, tracking the accumulation of misconceptions across a class over time, it could in principle return to teachers the cognitive bandwidth required for the work that only they can do. The relational, diagnostic, and contextual dimensions of teaching; the capacity to read a room, to detect the student whose quiet compliance masks genuine confusion, to repair a fractured relationship with a subject or with learning itself; these are not peripheral features of good teaching. They are its irreducible core. A teacher freed from the drudgery of marking thirty near-identical practice responses is, in principle, a teacher with more of herself available for the interactions that matter most.

The condition in all of these cases is the same word: deliberately. The same technology that deskills when deployed carelessly is the technology that augments when deployed with intention. The distinction between AI that replaces human cognition and AI that strengthens it is not a feature of the technology itself; it is a design choice, made or unmade by the people who commission, build, and implement these systems. Spell-checkers did not inevitably stop people learning to write. Calculators did not inevitably stop people learning to think mathematically. The question in each case was always the same: is this tool being used to bypass the cognitive work through which understanding is built, or to support the learner in doing that work more effectively? AI poses the same question, at greater scale and with higher stakes than any tool that has come before it.

What This Means for Education

The educational implications of all this are not difficult to see, even if they remain largely unexamined. Students now have access to systems capable of producing fluent explanations, essays, and problem solutions in seconds. These systems remove many of the visible barriers to completing academic work. But to evaluate whether an AI-generated answer is correct requires domain expertise; to detect a subtle error requires conceptual understanding; to judge whether an argument is coherent requires genuine familiarity with the underlying ideas. As I’ve said before, you cannot connect the dots if you don’t have any.

Without that knowledge, the student becomes a supervisor of information they don’t have the understanding to properly assess. The technology, in short, assumes the very expertise it risks undermining: and nowhere is that irony more consequential than in the domain whose entire purpose is the development of that expertise in the first place.

What all of this points toward is not the rejection of AI but the deliberate design of the human role within it. Automation never removes the human from the system; it recasts the system as an inescapable human-machine team, which must then be designed explicitly rather than inherited by default. That design question, of how much cognitive work to delegate, how to preserve the skills that delegation erodes, and how to structure oversight so that it does not simply exhaust the people performing it, is one that every organisation deploying AI at scale now faces. Including, with particular urgency, schools.

Bainbridge said all of this in four pages, in 1983, about nuclear power plants. It has taken us forty years and several peer-reviewed studies to confirm that she maybe have been presciently right, and that the domain she was describing was not industry but the future of work itself. The question now is not whether AI intensifies the cognitive burden on the humans who use it. The evidence on that point is, at this stage, fairly clear. The question is whether the organisations deploying these tools at scale are willing to take that finding seriously: to design human roles that are genuinely sustainable, to invest in the skills that AI cannot replicate, and to resist the seductive arithmetic that mistakes higher output for better thinking.

The ironies of automation do not resolve themselves. They compound.

Memory, Learning, and Knowledge

Does spaced retrieval practice work in all subjects? New study suggests the honest answer is sometimes, and we do not yet fully know when

This study examined whether the well established laboratory finding that spacing retrieval improves long term retention actually generalises across real university STEM courses. The researchers embedded spaced versus massed retrieval opportunities into bi weekly quizzes across eight introductory STEM courses, and then examined performance on a final criterial test. A ninth course, calculus, had been published previously but was included in a combined analysis. The headline finding is mixed: significant spacing benefits emerged clearly in calculus and in Chemistry for Health Professionals, small positive effects appeared in a few other courses, but most individual courses did not show statistically significant gains. When all nine courses were meta analysed together, the average benefit of spacing was small, around two to three percentage points, and highly variable across subjects.

Spacing may help in some STEM contexts, especially where procedures and integration across topics matter, but it is not guaranteed to produce gains everywhere. The authors raise important practical moderators: the nature of the questions, the feedback provided, the extent of initial learning, and how students engage outside quizzes. In short, spaced retrieval works beautifully in the lab. In real STEM classrooms, the picture is messier.

Not all revision is equal: testing beats restudy for retention. Students may dislike practice tests, but they work better than re-reading.
In authentic high-stakes exam preparation, testing only improves later performance when it includes feedback, and even then it is no better than focused restudy in the short term. This paper reports three experiments with Chinese teacher-training students preparing for a real, high-stakes educational psychology exam. Students revised key course concepts using one of four strategies: testing with feedback, testing without feedback, restudy, or an unrelated control task. Two experiments took place shortly before a high-stakes course exam, while a third occurred after the exam in a low-stakes context. Crucially, the practice materials and the final assessments used different questions, meaning any gains depended on transfer rather than simple item repetition.

Main findings: In high-stakes contexts, both testing with feedback and restudy produced similar immediate gains. However, only testing with feedback reliably transferred to better performance on the actual course exam days later. Testing without feedback was largely ineffective, and none of the strategies worked at all in low-stakes conditions. For teachers, the message is uncomfortable but clear: without feedback and without genuine stakes, its benefits largely disappear. If the goal is short-term exam performance on varied questions, structured restudy can work just as well as quizzing.

Instruction and Practice

Can we really measure student engagement?

This study set out to validate the Secondary School Student Engagement in Learning Activities Short Scale, a 12 item self report tool designed specifically to measure engagement in learning activities rather than broader school belonging. Drawing on self determination theory, the scale measures four dimensions: cognitive engagement, affective engagement, behavioural engagement, and agentic engagement. Using a sample of 566 Portuguese secondary students, the authors conducted confirmatory factor analysis and found good model fit for the four factor structure. The scale showed acceptable reliability overall, strong concurrent validity with an existing engagement measure, and some predictive validity for competence, relatedness, self reported achievement, and Portuguese grades.

Main takeaway is that engagement is not one thing. Students can be thinking deeply, enjoying learning, putting in effort, and actively shaping instruction in different combinations. The inclusion of agentic engagement is particularly important. Students who ask questions, express preferences, and influence their learning environment appear to gain psychological benefits beyond effort and enjoyment alone. The scale is short enough to be used for monitoring trends or identifying at risk students, but it is not a diagnostic instrument. It gives a snapshot, not a full psychological profile.

Does this paper address Rob Coe’s claim that engagement is a poor proxy indicator of learning? Well this paper is a measurement validation study, not a causal test of whether engagement leads to learning. It shows that four forms of self reported engagement are associated with self reported achievement and Portuguese grades. This paper neither proves nor disproves it. It shows engagement correlates with achievement, but it does not test whether engagement tracks durable knowledge acquisition.

For me the critique still stands: A student can look engaged and not be learning, and a student can look disengaged and still encode effectively. Observable engagement is not a guaranteed indicator of knowledge change.

Do decorative extras in lessons really help, or do they quietly get in the way of learning?
This meta analysis suggests that seductive details have a small but reliable negative effect on learning, mainly because they increase extraneous cognitive load rather than helping learners understand or transfer knowledge.

This study tackles a long running question in instructional design: what happens when we add interesting but irrelevant material to learning resources. These seductive details might be decorative pictures, entertaining anecdotes, background music, or eye catching but non essential content. Drawing on 177 effect sizes from 50 studies, the authors found a small overall negative effect on learning outcomes. That pattern held across recall, comprehension, and transfer, with the strongest negative effect for comprehension. In simple terms, the extras may make materials seem more appealing, but they tend to slightly reduce what people actually learn.

Subscribe now

Reading and Language Development

Do popular reading comprehension strategies actually improve students’ performance on real reading tests?

This meta analysis set out to answer a deceptively simple question: which instructional approaches genuinely improve reading comprehension, particularly when measured using standardized assessments rather than researcher designed tests. That distinction matters because many interventions look effective when measured with tests closely aligned to the taught material, but the effect often shrinks when students are assessed using broader comprehension measures.

To answer this, the authors screened 1,557 studies drawn from ERIC, Education Source, and earlier meta analyses, eventually analysing 70 studies and 116 effect sizes that met strict inclusion criteria such as experimental or quasi experimental designs and sufficient statistical reporting.

The results show generally small average effects on standardized reading comprehension outcomes. Weighted effect sizes were modest for most approaches: vocabulary instruction around d = .11, content instruction d = .17, cognitive strategy instruction d = .21, and metacognitive strategies d = .06. Reciprocal teaching stood out with a larger effect of d = .45, though the evidence base for this approach was relatively limited.

One of the most interesting findings is that combinations of instructional approaches produced larger effects than individual pedagogies. For example, when vocabulary instruction was combined with content knowledge and cognitive strategy instruction, the effect size rose to around d = .51, suggesting a synergistic effect between these components.

Another key point is that instruction often improves comprehension of texts related to the taught material but does not reliably transfer to generalized reading comprehension across new texts. This transfer problem helps explain why effect sizes frequently shrink when standardized tests are used.

This study has proved to be somewhat controversial. Olivia Mullins and Natalie Wexler have written about this paper and I have also responded here:

Reading and writing appear to depend on the same underlying language system.

The study followed 261 children from kindergarten to Grade 2 in the United States to investigate the relationship between reading comprehension and written composition and the skills that explain that relationship.

Using confirmatory factor analysis and structural equation modelling, the author examined whether shared language and literacy skills account for the connection between reading and writing. The analysis showed that Grade 2 reading comprehension and writing quality were very strongly correlated (r = .82) but were still statistically distinct constructs.

Crucially, once oral discourse skills (listening comprehension and retell) and lexical literacy skills (word reading and spelling) were included in the model, the correlation between reading comprehension and writing almost disappeared. Handwriting fluency contributed to writing quality but not to reading comprehension.

For teachers, these findings imply that the common practice of teaching reading and writing in isolated silos may be inefficient. Since oral discourse and word-level skills (spelling/decoding) are the "pillars" supporting both domains, strengthening these foundations in the early years provides a double benefit. Teachers might focus more on integrated lessons where oral storytelling, vocabulary development, and spelling are used to bridge the gap between understanding a text and producing one, ensuring that children have the cognitive tools necessary to succeed in both.

Classrooms, Behaviour, and Attention

Do teachers really need charisma to be effective, or is the idea of the “charismatic teacher” more myth than method?

The paper reviews literature from the scholarship of teaching and learning (SoTL) and identifies three recurring ways researchers describe charismatic teachers. First, charisma is often portrayed as an uncultivable natural gift possessed by only a few individuals. Second, charismatic teachers are frequently associated with humour, which is assumed to increase engagement and approachability. Third, charisma is linked to high energy performance through expressive delivery, enthusiasm, and emotionally charged classroom interactions.

The author argues that each of these ideas is problematic. Treating charisma as a rare gift risks discouraging teachers who do not see themselves as naturally entertaining. Linking charisma to humour overlooks the risks of inappropriate or culturally misinterpreted jokes. And equating effective teaching with emotional intensity may place unrealistic expectations on teachers while privileging a performative style that is not necessary for learning. The paper ultimately suggests that education may benefit from shifting focus away from theatrical charisma.

Does strict, controlling teaching actually help unmotivated students learn better? A one year study of nearly 5,000 students suggests controlling teaching reduces engagement and achievement.

This paper examines a persistent belief in education: that strict or controlling teaching may be necessary for students who lack motivation. The study approaches this question through the lens of Self-Determination Theory (SDT), which argues that motivation and engagement depend on the satisfaction of three psychological needs: autonomy, competence, and relatedness. Controlling teaching styles undermine these needs by pressuring students to behave, think, or learn in ways dictated by the teacher.

The researchers followed nearly 5,000 Chinese middle school students across a full academic year. Their goal was to understand how perceived controlling teaching influences learning outcomes and whether the impact differs depending on students’ motivation profiles.

The results suggest that students who perceived their teachers as more controlling showed lower behavioural engagement (participation and effort) and lower cognitive engagement (deep thinking, strategy use, and persistence). Those drops in engagement were in turn associated with lower academic achievement one year later.

A key contribution of the study is its focus on motivation profiles rather than average motivation levels. Using latent profile analysis, the authors identified four types of students: controlled motivation, moderately motivated, strongly motivated, and autonomous motivation. The researchers then tested whether the effects of controlling teaching differed between these groups.

The main finding is that the negative effect appeared across all motivation profiles. Even students with lower motivation did not benefit from controlling instruction. While the strength of some pathways varied slightly between groups, the direction of the effects was consistently negative. This directly challenges a common classroom assumption: that strict control might be necessary or effective for disengaged students.

SEN, Individual Differences and Inclusion

For many children with ADHD, improving sleep quality may do more to support focus, behaviour, and learning than medication alone.

A large brain-imaging study suggests that common stimulant medications for ADHD (like methylphenidate/Ritalin and amphetamines/Adderall) do not primarily boost attention networks in the brain as previously assumed. Instead, these drugs seem to act on brain circuits that control alertness (wakefulness) and reward/motivation, helping people feel more awake and increasing the perceived value of tasks that they would normally find uninteresting.

This matters because it reframes what ADHD medication is actually doing. If stimulants mainly increase alertness and motivation rather than fixing attention itself, then poor sleep is not a side issue but a central factor. A tired brain will struggle to focus no matter what, and medication may simply be masking sleep deprivation rather than solving the underlying problem. For some children, getting consistent, high-quality sleep may therefore deliver many of the same benefits as medication, without the trade-offs, and should be treated as a first-order intervention rather than an afterthought.

By age six, children are already on different fraction trajectories whether we teach fractions or not.

This study finds that first graders cluster into three clear profiles of early fraction understanding, and those profiles meaningfully predict end of year mathematics achievement. The researchers examined 204 first graders at the start of the school year, focusing on early, mostly informal fraction knowledge across nonsymbolic and symbolic tasks. Using latent profile analysis, they identified three distinct groups. Profile 1 showed strong nonsymbolic fraction knowledge. Profile 2 was similarly strong but weaker specifically on nonsymbolic equivalence tasks. Profile 3 showed broadly limited fraction understanding across areas.

Crucially, these early profiles predicted later outcomes. By the end of first grade, children in Profiles 1 and 2 outperformed those in Profile 3 on mathematics achievement. Interestingly, the initial weakness in nonsymbolic equivalence shown by Profile 2 largely disappeared by year end. In other words, children who were strong on part whole relations and equal sharing but weaker on equivalence appeared to “catch up” in that specific area over time. For teachers, the message is nuanced: early strengths in equal sharing and part whole reasoning matter, and a specific early weakness in equivalence may not signal long term difficulty. However, broadly weak fraction knowledge at school entry is a more serious concern.

AI, Technology, and Cognitive Tools

AI amplifies the Matthew Effect: without strong foundational knowledge, inequality deepens.

A new Brookings report argues that artificial intelligence will not automatically transform education for the better or the worse. Its impact depends on how it is integrated into teaching and learning. AI has the potential to expand access, personalise instruction, support neurodivergent learners, and free teacher time. But under current conditions, the risks outweigh the benefits. The report warns that AI can weaken cognitive development through overreliance, erode trust between teachers and students, undermine assessment integrity, disrupt social and emotional growth, and deepen equity divides.

A central claim is that AI amplifies existing differences. Students with strong foundational knowledge can use AI to extend and refine their learning, while those without such knowledge are more likely to misinterpret outputs, outsource thinking, and fall further behind. The authors therefore call for a shift: embed AI within sound pedagogy, strengthen knowledge rich instruction, build holistic AI literacy across the curriculum, establish national guardrails, and prioritise child safety, privacy, and developmental needs.

How much damage does a single phone notification really do to thinking?

The researchers asked university students to complete a demanding Stroop task while realistic smartphone notifications appeared on screen. Crucially, the notifications mimicked real iPhone pop ups and varied in whether participants believed they were their own, generic, or visually blurred. Across conditions, notifications caused a clear and measurable slowdown in response speed lasting roughly seven seconds, even though accuracy mostly recovered more quickly. The disruption was strongest when notifications were believed to be personally relevant, weaker when they were generic, and weakest when they were visually degraded.

The study matters because it moves beyond artificial lab alerts and shows that ordinary notification cues hijack attention automatically. The effect does not require opening the phone or reading the message. Simply seeing the alert is enough to disrupt thinking, and this disruption stacks up across a day filled with interruptions.

New research suggests machines learn better when they “mumble” to themselves. Inner dialogue may be as important for machines as it is for humans.

This study, published in Neural Computation, examined whether adding self directed inner speech to an AI system with working memory would improve multitask generalisation. The researchers designed models with multiple working memory slots and trained them to “mumble” to themselves a specified number of times during task processing. They tested these systems on tasks of increasing difficulty, such as reversing sequences, reconstructing patterns, and switching between tasks. Models with multiple memory slots outperformed simpler ones, and performance improved further when structured inner speech was added. The gains were most pronounced in multitasking and multi step problems, and importantly the system reportedly required less training data to generalise.

Why does this matter? The core claim is that learning is shaped not only by architecture but by internal self interaction during training. In human terms, it suggests that structured self talk plus working memory scaffolding supports flexible problem solving. For educators, this resonates strongly with decades of research on metacognition, self explanation, and overt verbalisation. If AI systems benefit from explicit internal rehearsal and structured working memory supports, that reinforces the idea that prompting students to articulate reasoning and manage cognitive load is not superficial but central to flexible thinking.

Can humans still tell the difference between a child’s painting and an AI imitation of it?

The study tested a version of the Artistic Turing Test. Seventeen oil paintings by a single child served as the human baseline. For each one, the researchers crafted a detailed prompt and used DALL E 2 to generate a corresponding AI image. Eighty seven adults were shown the images in pairs and asked to decide which was human made.

Human accuracy was 49.76 percent, effectively no better than chance. Precision, recall, and F1 scores were all in the low fifties. By contrast, a ResNet 18 classifier using leave one out cross validation achieved 97.06 percent accuracy, with an AUC of 99.65 percent. Put simply, people could not reliably tell the difference, but a machine could.

The implication is not that AI possesses human creativity. It is that surface stylistic cues are no longer dependable signals of authorship. Framing the works as child art may also have led participants to interpret quirks and imperfections as authentic. Generative models have reached a level of perceptual plausibility that challenges visual judgement, even if deeper questions about meaning and intention remain open.

Does artificial intelligence actually improve learning in universities or are we mostly describing possibilities rather than evidence?

This systematic review of 43 studies concludes that AI is widely associated with innovation and pedagogical change in higher education, but the evidence base mainly describes potential benefits rather than demonstrating clear causal improvements in learning outcomes.

This study set out to synthesise research on the use of artificial intelligence in higher education between 2021 and 2025. The authors reviewed 43 articles retrieved from Scopus, ScienceDirect, and Google Scholar and grouped them into three main themes: AI driven learning innovation, pedagogical transformation, and implementation challenges. The review argues that AI systems can support personalised learning, automated feedback, predictive analytics for identifying struggling students, and generative tools such as ChatGPT that assist with writing and content generation. Across the literature, AI is portrayed as a driver of digital transformation in universities that may reshape teaching practices and institutional decision making.

For educators and institutions, the key message is that AI appears to change the structure of teaching rather than simply adding new tools. Teachers increasingly act as designers, facilitators, and interpreters of learning analytics while AI systems automate routine tasks such as grading or content generation. However, the literature also emphasises concerns about privacy, algorithmic bias, digital inequality, and teacher readiness. The practical implication is that adopting AI in universities requires not only technological infrastructure but also ethical governance, staff training, and careful integration into existing pedagogical practices.

Housekeeping

Listen to my conversation with the brilliant here

The Dutch Version of Instructional Illusions is out now. Just got my copy and what a lovely cover is is too.

I have rescheduled my Australia dates that were in March until May. For information on all the upcoming dates in May, check my website.

I’ll be at three upcoming researchED events; Tokyo on April 18th, Warrington on 25th April and Crieff on the 9th May.

Finally, I have written a review of Jared Cooney Horvath’s new book the Digital Delusion.

One of the more counterintuitive things about large language models is that the prompt contains almost none of the knowledge required to generate the response. A few words typed into ChatGPT can produce paragraphs of detailed, coherent text, not because the prompt contained that information but because the model has already absorbed vast amounts of language and knowledge during training. The prompt activates patterns that already exist inside the system. Remove that prior knowledge and the same prompt produces nothing at all.

Reading works in much the same way. A text does not carry its meaning fully formed inside it. It relies on knowledge that already exists in the reader’s mind: vocabulary, background information, conceptual frameworks built over years of experience. When those structures are present, meaning emerges quickly and almost effortlessly. When they are absent, the words remain visible but the understanding never arrives.

The parallel is not merely metaphorical. In cognitive psychology, Walter Kintsch’s construction–integration model describes comprehension as a process in which a text activates a network of possible meanings drawn from the reader's prior knowledge; the mind then settles on the interpretations that best fit. Large language models operate in a very similar way: a prompt activates learned associations and the system converges on the most coherent continuation. In both cases, the crucial ingredient is not the input but the knowledge already present in the system.

“Teach big words to little kids, and do it interestingly.”

In education, the idea that reading comprehension is a teachable, generic skill that can be applied to any given text is an error, but a natural one, which is partly what made it so durable. We observed what comprehension failure looked like: students who could not identify the main idea, who could not draw inferences, who could not synthesise information across paragraphs. And then we did something that felt entirely logical but was, in retrospect, a category error: we turned the description of the failure into the curriculum. We taught “main idea” identification, inferencing, synthesis etc. We built an entire instructional architecture from the symptoms of the problem rather than from its cause. The description of what students could not do was mistaken for an explanation of why they could not do it.

For me, the deeper reason why strategy instruction fails to transfer is that what we call “strategies” are really meta-labels for underlying linguistic principles. You cannot summarise a passage if you do not understand the semantic relationships between its sentences. You cannot infer an author’s purpose if you do not know what half the words mean. You cannot “find the main idea” if the main idea is expressed in vocabulary that might as well be a foreign language. The strategy is downstream of the knowledge, not upstream of it. It is like teaching someone the rules of chess on a board where half the pieces are invisible. The rules are not the problem. The missing pieces are.

One of the most powerful vehicles for helping students understand a text and also one of the most misunderstood, is vocabulary instruction. This point is brilliantly made by Isabel Beck who says that what we should do about this is “Teach big words to little kids, and do it interestingly.” The phrase is deceptively simple. What it recognises is that comprehension depends on the language students bring to a text. The size of a child's vocabulary is not a measure of how many words they have memorised but of how much of the world they have been taught.

What a New Study Reveals

A new meta-analysis on reading comprehension that reexamined and reanalysed studies from three previous meta-analyses, breaking the results down by a distinction that turns out to be decisive: standardised assessments versus researcher-created measures. The headline finding is stark. Nearly all effect sizes shrank substantially when measured on standardised tests. Cognitive strategies, the bread and butter of comprehension instruction for the past thirty years, produced a weighted effect size of just 0.09. Metacognition strategies showed no significant effects at all. For anyone who has sat in a professional development session being told that teaching students to “learn about learning” or “monitor their own thinking” is the key to comprehension, this should be a sobering corrective.

However, a surprising finding is that content instruction appeared to show no effect either. But this is where the picture requires a second, harder look. has published a detailed response arguing that the collection of papers on content knowledge and comprehension included in Hansford et al. are simply not an appropriate set from which to draw conclusions about knowledge-building. Critical papers were excluded. Some included papers had dubious methodology. The data, she argues, needs to be taken in context, because the broader research base tells a rather different story.

The meta-analysis told us something important about strategies. But its treatment of content and knowledge instruction is where we need Mullins’ corrective. And that corrective, as it turns out, reframes something much larger than a single study.

But What About Fluency?

There is an obvious objection here, and it deserves to be taken seriously. Reading comprehension does not rest on vocabulary alone. Fluency matters: the speed and accuracy with which a reader can process text. If decoding is slow and effortful, working memory is consumed by the mechanics of turning print into language, and there is nothing left over for meaning. Hirsch put the problem precisely in 2003: if decoding does not happen quickly, the decoded material will be forgotten before it is understood. Anyone who has tried to follow a film in a language they half know will recognise this immediately; by the time you have translated one sentence, the next has already passed, and the connections between them dissolve.

But here is what makes fluency an ally of the vocabulary argument rather than a rival to it. Fluency is not purely a decoding phenomenon. It is also a knowledge phenomenon. A reader who knows a domain reads that domain faster, not because their eyes move differently but because their mind can chunk familiar information into larger units. This is the lesson of de Groot’s famous chess experiment: grand masters can reconstruct a mid-game board from a five-second glance, not because they possess superior memory but because they recognise patterns that novices cannot see. Arrange the pieces randomly and the advantage vanishes. The “skill” was knowledge all along.

Hirsch draws the analogy to reading directly: word knowledge speeds up word recognition, and world knowledge speeds up comprehension of textual meaning by offering a foundation for making inferences. The three factors, fluency, vocabulary, and domain knowledge, are not independent columns holding up the same roof. They are the same column, seen from different angles.

Spencer, Quinn, and Wagner confirmed this at enormous scale. They tested 425,000 children from first through third grade on decoding, vocabulary, and comprehension. Among those who had adequate decoding and adequate vocabulary, fewer than one per cent scored poorly on comprehension. Fewer than one in a hundred. Once the word-level problems are solved, the decoding and the vocabulary, comprehension takes care of itself in virtually every case. The finding does not diminish the importance of fluency. It clarifies its place: fluency is necessary but not sufficient, and in the presence of adequate vocabulary, it is almost always enough.

However the bottleneck, for the vast majority of struggling readers, is not the speed at which they process words. It is whether they know what those words mean. And even within Hansford’s own data, Willingham’s observation holds: students appear to gain all the benefits of strategy instruction after approximately ten hours, and increasing instructional time by as much as 400 per cent produces no further gains. Rosenshine found that spending six classes on teaching comprehension skills had the same effect on reading comprehension as spending twenty-five.

Strategies are not a deep competence that rewards sustained practice; they are a thin procedural layer that maxes out almost immediately. When you strip away the researcher-created measures, the flattering proximal tests, the assessments designed to detect exactly what was taught, strategy instruction produces almost nothing on the measures that matter.

Knowledge-Building Is Not an Intervention

But the meta-analysis did something else, something I think it perhaps did not intend. In its treatment of content instruction, it revealed a methodological bias so deep that it calls into question how we evaluate knowledge-building interventions altogether. This is where Mullins’ critique becomes essential. Her central argument deserves to be quoted at length:

“Knowledge building is not a short intervention. In fact, it’s not an intervention at all. It’s a gradual, cumulative, lifelong process. We already understand that short bursts of content instruction will not affect general reading comprehension.”

I think this is a hugely important point and a big weakness in the science of learning as a field, which sees everything as an intervention but in a domain where the signal to noise ratio is often imperceptible. The entire architecture of the randomised controlled trial assumes that “content instruction” is a sort of treatment: something that can be administered in a controlled dose, measured against a control group, and evaluated after a fixed period. But knowledge does not work like a treatment. It compounds. It accretes. It builds the very architecture through which future comprehension becomes possible. To measure it as though it were a pill is to fundamentally misunderstand the mechanism.

Barbara Oakley has recently argued for what she calls "cognitive realism": the principle that there are facts about how brains encode, consolidate, and retrieve information, and that any instructional framework which cannot be falsified by those facts is not a framework but an ideology. The deeper problem is categorical. Knowledge-building is not a strategy; it is what I have elsewhere called an instructional invariant: a condition, derived from how human cognition actually works, that must hold true continuously for learning to function. Invariants are not treatments to be tested; they are constraints to be maintained. Asking whether knowledge-building 'works' in a six-week trial is like testing whether foundations 'work' by building a house without them and checking if it falls down in the first month. The house might stand for six weeks. But it won’t stand for six years.

Mullins identifies ten reasons to look beyond the meta-analysis, and several of them are damning. The content studies ranged from three days to one school year; averaging across these is like averaging the temperature of ice and boiling water and calling it warm. Important studies were excluded, including the Romance and Vitale work, perhaps the cleanest quasi-experimental studies we have, which produced an effect size of 0.56 after a full year of integrated science and literacy instruction. Many included studies used student-led approaches unlikely to build knowledge effectively. But here is the critical detail: even within Hansford’s own data, there was a significant positive correlation (r = 0.43) between instruction duration and effect size for content instruction. Content instruction was the only treatment where this relationship was found. For every other approach, including strategies, longer duration was associated with smaller effects. Knowledge-building was the only intervention that got stronger the longer it ran. This single finding should have been the headline of the paper.

And what happens when we let that compound interest run? As pointed out, a rigorous study of more than 2,000 students using Core Knowledge Language Arts, a curriculum that systematically builds knowledge from kindergarten, found that after four years of implementation, students from low-income families matched their higher-income peers on state reading comprehension tests. Four years. Not six weeks. The knowledge-building did not merely help; it closed the gap entirely. But you would never see that result in a study that measured outcomes after a single term.

Drops in an Ocean

But there is a complication, and it is one I am not entitled to avoid, because I believed it myself for years. If the problem is vocabulary, then surely the solution is to teach more vocabulary: pre-teach the key words before each text, drill Tier 2 lists, front-load definitions. This is the naive version of the argument, and it is also wrong.

Direct vocabulary instruction produces meaningful effects on comprehension of texts containing the taught words. This much is well established. But Cervetti and colleagues found in 2023 that direct word teaching does not reliably generalise to untaught vocabulary breadth. You cannot simply pre-teach twenty words per week and expect comprehension to blossom across all texts. The words you teach help students read the passages that contain those specific words. They do not, on their own, build the kind of broad vocabulary knowledge that underpins general reading comprehension.

The threshold research explains why this matters so acutely. Laufer found in 1989 that students need 95 to 98 per cent known-word coverage for comfortable comprehension. Hu and Nation raised that estimate in 2000 to 98 to 99 per cent for unassisted reading. When a student encounters a passage where even five per cent of the words are unknown, the signal-to-noise ratio is too low for any strategy to rescue meaning. The text becomes opaque. Not difficult, not challenging, but genuinely impenetrable.

But here is the arithmetic that should trouble us. If a student needs to know 98 per cent of the words in a passage, and they are encountering academic texts with thousands of distinct word forms, and direct instruction of individual words does not generalise beyond the words taught, then the gap cannot be closed one word at a time. Schools teaching 300 to 400 Tier 2 words per year through explicit instruction are adding drops to an ocean. The scale of the problem dwarfs the scale of the solution.

So the question becomes: is there an approach to vocabulary instruction that gives us a better bang for our buck? which is multiplicative rather than additive? One that does not just teach individual words but teaches the generative logic from which families of words become comprehensible?

Words Are the Visible Surface of Knowledge

This is where vocabulary instruction and knowledge-building converge, and where a much more cohesive theory of how reading actually develops begins to emerge.

Vocabulary does not live in word lists. It lives in schemas, in morphological families, in the accumulated residue of having learned about the world. To teach words well is to teach the content from which they derive their meaning. Beck and McKeown’s framework calls for six to twelve robust encounters before a word enters long-term usable vocabulary. Not flashcard reviews; encounters in context where the word does real semantic work. This only happens within content-rich curricula, across time, through accumulation.

But it can also happen through morphology. Nagy and Anderson estimated in 1984 that roughly 60 per cent of the new words students encounter in print have meanings that can be predicted from their morphological parts. Sixty per cent. That is not a marginal gain; it is the difference between a closed system and an open one. Lyn Stone, a linguist and literacy specialist, puts the principle sharply in Reading for Life: it would be impossible to teach the definition of every single word in English, but that does not mean we cannot teach children to be generative. Knowing how words work, having an understanding of a core of prefixes, roots, and suffixes, helps children generate the meaning of new words.

So if the arithmetic of direct vocabulary instruction is so dispiriting, the question becomes whether there is an approach that teaches not just individual words but the generative logic from which families of words become comprehensible. Not 400 words a year, but the structural principles that unlock thousands. This is where the work of and the Word Mapping Project offers something that the research literature demands but rarely provides: a concrete, classroom-tested illustration of what generative vocabulary instruction actually looks like.

Morrisey’s approach centres on morphology matrices. Take the base sect, from Latin, meaning “cut.” Arrange its combinatorial possibilities: prefixes on the left (bi-, dis-, in-, inter-, mid-, sub-), suffixes on the right (-ion, -or, -ed, -ing, -cide, -vore, -ous, -al). Now give students the matrix and ask them to build words.

What emerges is not a vocabulary list but a generative system. From a single base, properly understood, students can construct bisect, dissect, insect, intersection, midsection, subsection, insecticide, insectivorous, and dozens more. They are not memorising words. They are learning the code from which words are made.

The pedagogical architecture is as important as the morphological content. Students begin by investigating the matrix: working alone or with a partner, building as many words as possible, writing word sums on the left (bi + sect + ion) and the whole word on the right (bisection). This is retrieval-generative practice, not passive reception. The teacher lesson includes full morphological breakdowns with definitions rooted in the base meaning: bi + sect = to cut into two parts; dis + sect = to cut apart; in + sect = literally “cut into,” a segmented creature. The etymology is not decoration. It is the explanatory mechanism. Students learn why words mean what they mean, not just that they do.

Extended vocabulary passages then group semantically related words into clusters. Not isolated definitions but constellations of near-synonyms that force discrimination. One passage, for instance, presents arduous, strenuous, laborious, gruelling, and demanding alongside effortless, smooth, unchallenging, painless, and seamless. Students partner up, reread the passage two to three times, then work on questions that require distinguishing nuance between words that share a broad category but differ in shade and weight.

This for me is Beck and McKeown’s “robust encounters” principle made concrete: words appear in context, doing real semantic work, and students must think about the differences between them, not just their shared meaning.

On subsequent days, retrieval takes unexpected forms. “Would you rather” questions: Would you rather diminish the amount of broccoli on your plate or diminish the number of commercials in a show? These are low-stakes, high-engagement, and they force students to use the word in a new context, exactly the kind of productive use that moves vocabulary from receptive recognition to active command. Reader’s theater scripts feature characters named after vocabulary words: Stricken, Protected, Bickering, Accord. The words become agents in a story. Students cannot perform the scene without internalising what the words mean.

But the finding that matters most, the one that answers the Cervetti problem, is this: Morrisey’s standardised spelling test data (Test of Written Spelling, 5th Edition) showed substantial growth over just four months, and crucially, on words that were not specifically taught.

This is the transfer that Cervetti’s research says does not happen with standard direct vocabulary instruction. Morphological instruction generates transfer because it teaches the underlying code, not individual items. A student who understands that sect means “cut” and that -ion makes a noun does not need to be explicitly taught dissection. They can construct its meaning from its parts. This is the difference between teaching a student a hundred fish and teaching them the structure of aquatic ecosystems.

But What About Critical Thinking?

There is one more thing that needs to be said, because the argument above invites a misreading I want to anticipate. Nothing here suggests that inference, critical analysis, or the capacity to weigh competing perspectives are unimportant. They are profoundly important. But my claim is that they are dependent: they require a substrate of organised knowledge before they can function. You cannot think critically about the causes of the First World War if you do not know what the causes of the First World War were. You cannot weigh both sides of an argument about gene editing if the word genome is a stranger to you. The higher-order thinking we rightly value is not an alternative to knowledge; it is what knowledge makes possible.

Nor does the argument reduce to “teach more facts.” Isolated facts are almost as instructionally inert as isolated strategies. What matters is connected knowledge: words understood in relation to other words, concepts organised into schemas, information structured so that new learning has something to attach to. This is why morphological instruction works where word lists do not, and why content-rich curricula outperform topic-of-the-week approaches. The goal is not to fill children’s heads with disconnected information. It is to build the architecture of understanding, so that when they are finally asked to infer, to analyse, to synthesise, they have something to infer from.

If higher-order thinking depends on knowledge, and knowledge depends on vocabulary, then the students with the smallest vocabularies are not merely behind in reading. They are being locked out of the very capacities we claim to value most: the ability to reason, to question, to see a problem from more than one angle. A child who cannot access the language of a debate cannot participate in it. A child who does not know what the words mean cannot think critically about what they say. The vocabulary gap is not just a literacy problem. It is, at bottom, an equity problem, and it is one that begins earlier and compounds faster than most people realise.

Running the Same Race From a Different Starting Line

Which brings me finally to the students for whom all of this matters most. Hart and Risley found in 1995 that vocabulary exposure gaps exceed thirty million words between socioeconomic groups by age three. The vocabulary gap is visible by the end of Grade 2, often exceeding 3,000 root words between the highest and lowest performing groups. After that point, low-vocabulary children may grow at the same rate as their peers, but they are running the same race from a different starting line.

Before age ten, more than 80 per cent of words are learned from language experiences in context (e.g. explanations from parents, being read to, conversation, etc) rather than from inference from reading. The school curriculum is their primary source of the background knowledge and word meanings that more advantaged children absorb at home.

The sociologist James S. Coleman, after spending a career examining the characteristics of effective schools, concluded that the most important feature of a good school programme is that it makes good academic use of school time. The reason is as simple as it is consequential: for disadvantaged children, school time is the only academic learning time, whereas advantaged students learn a great deal outside of school. A good programme is inherently compensatory, because it has a bigger effect on those who depend on it most. Coleman’s principle reframes the vocabulary gap not as a problem of ability but as a problem of exposure, and it places the responsibility squarely on what schools choose to do with the hours they have.

The point that should keep curriculum designers awake at night is that even if content instruction truly had no effect on comprehension, which the evidence does not show, knowledge-building curricula would still be the right choice. Integrated content instruction does not hurt comprehension. It does improve content knowledge, vocabulary, and writing. There is simply no competing coherent argument for scattered, content-poor approaches to literacy instruction.

The students most harmed by the strategy-heavy, knowledge-light approach are the ones who arrive at school with the smallest vocabularies. They are the ones for whom the school curriculum is not a supplement but a lifeline: the primary means by which they will acquire the words and the knowledge that make comprehension possible. To deny them systematic, cumulative, content-rich instruction, instruction that builds morphological knowledge, that teaches the generative logic of language, that treats vocabulary as knowledge rather than a checklist, is to treat a knowledge deficit as a skills deficit.

For Weil, attention was not merely a cognitive capacity but a moral and spiritual orientation toward the world. To attend fully to something, whether a geometry problem, a passage of Greek, or another human being in their suffering, was to practise a form of self emptying. It required the suspension of ego, impatience, and the urge to master and move on. Attention, in Weil’s formulation, was a kind of liberation, what she describes as “waiting, not seeking.”

I return to Weil now because her understanding of attention illuminates something about our present condition that many contemporary analyses, for all their sophistication, tend to miss. We are adept at diagnosing the attention economy, at mapping its mechanisms of capture and exploitation. What we struggle to articulate is what is lost when it is fragmented and commodified. What kind of self fails to form in its absence.

Weil understood attention not as something the self deploys, like a tool, but as something through which the self comes into being. To attend is to be formed. Iris Murdoch, who extended Weil’s thinking into moral philosophy, made this point explicit in The Sovereignty of Good: the quality of our attention determines the quality of our moral life. What Murdoch called “the fat relentless ego”, or the self that insists on its own centrality, that reduces every encounter to a mirror, is precisely the self that cannot attend. And attention, she argued, is not a preliminary to moral action. It is moral action. The just and loving gaze directed upon an individual reality is the characteristic and proper mark of the active moral agent.

If Murdoch is right, then the crisis of attention is not simply a matter of productivity or information management. It is a crisis of moral formation. A self that has never learned to attend, to hold still, to look carefully, to resist the impulse to categorise and move on, is a self that remains, in some important sense, unformed. The attention economy does not merely distract us. It arrests a process of becoming ourselves.

The Architecture of Capture

In 2006, Facebook introduced the News Feed. The internal metric they used to evaluate its success was called “engagement”, a word borrowed, with no apparent irony, from military terminology for making contact with an enemy. What it measured was not whether users had learned anything, or felt enriched, or understood the world more clearly. It measured whether they had stayed. The longer the better. The ideology encoded in that metric, that the purpose of a human being’s attention is to be captured and held, has since become so pervasive that we barely notice it. It is the water in which we swim.

But it is worth pausing to consider what happened: an entire civilisation’s relationship to its own attention was restructured around a metric that equated value with time-on-screen. No one voted for this. It was not a decision at all, in any sense that democratic societies would recognise. It was an optimisation. And by the time anyone thought to ask what, exactly, was being optimised away, the habits it had shaped were already too deep to feel like habits. Against this backdrop, the act of reading, real reading, the kind that demands sustained attention over hours and days, begins to look strange. It starts to resemble a refusal.

To sit with a difficult book is to decline the invitation to be elsewhere. It is to resist the pull of the feed, the ping of the notification, the low level itch to check. It is to say: I will be here, in this sentence, with this difficulty, for as long as it takes. It is to practise the kind of attention Weil describes: not grasping but receptive, not seeking but waiting.

That kind of reading is now a rare form of resistance, though not in the familiar sense of protest or spectacle. It is closer to what Václav Havel called “living in truth”: a refusal to participate in the lies a society tells itself. Havel’s celebrated example was not the dissident on the barricade but the greengrocer living under Communist rule who stops placing the party slogan in his window. The act is quiet, almost invisible. No one applauds. But in ceasing to perform a loyalty he does not feel, the greengrocer reclaims something essential about his own relation to reality.

When the Book Became the Feed

In 2022, the publishing industry celebrated a surge in book sales driven by BookTok, the corner of TikTok devoted to reading recommendations. The books that went viral shared certain characteristics: they were fast, they were emotionally immediate, they were describable in a single sentence. The platform had, in effect, selected for books that behaved like the platform itself. What was celebrated as a renaissance of reading was, on closer inspection, the final absorption of the book into the logic of the feed. The form survived. The practice it once required did not. It was as if a culture had preserved the ritual of going to church but replaced the liturgy with a three-minute highlight reel.

We have now built an entire culture of simulated reading: an elaborate apparatus for performing the appearance of engagement with ideas while evading the labour that real engagement actually requires. The book summarised in a thread. The argument reduced to a pull-quote. Ideas encountered only through someone else's paraphrase, passed along with the confidence of first-hand knowledge. What circulates in such a culture is not understanding but its simulacrum; a fluency untethered from the slow, often uncomfortable process by which real understanding forms. We have become, in a phrase Havel might have recognised, a society of intellectual greengrocers: placing slogans in our windows that we neither believe nor have earned the right to display. When actually reading a book counts as an unusual act, something has gone profoundly wrong with a culture’s relationship to its own intellectual life.

What would it mean to take reading seriously as resistance?

It is tempting to dismiss concerns like these as luxury beliefs, as the indulgence of those with the time and security to worry about the quality of their attention. And it is true that the conditions for deep reading have never been equally distributed. But the erosion of contemplative life is not merely a private loss. It is a civic one. A society of distracted, exhausted selves is a society incapable of deliberation, incapable of listening, incapable of the slow work of understanding that democratic life requires.

What would it mean to resist the feed? It would mean treating attention as something that must be practised, not assumed. The habits of distraction are deeply grooved, and the devices that fragment attention are rarely more than an arm’s length away. This is not accidental cultural drift. It is an environment deliberately engineered to make sustained attention feel unnatural. Resistance therefore requires discipline: repeated, intentional acts that slowly reshape our capacities. Reading, in this sense, is an askesis; a training of attention.

The word is worth dwelling on. Askesis is the root of asceticism, but its original meaning is broader: it refers to any disciplined practice undertaken for the formation of the self. The monastic tradition, which took the training of attention more seriously than any culture before or since, understood that sustained focus was not a natural state but a hard-won achievement. The Desert Fathers’ writings on acedia, that peculiar restlessness of the spirit, the noonday demon, are essentially a phenomenology of distraction avant la lettre.

John Cassian, writing in the fifth century, described the monk whose mind wanders ceaselessly: to the next meal, to idle gossip, to the conviction that he would be better off almost anywhere else, doing almost anything else. The resemblance to the experience of trying to read seriously with a phone nearby is uncanny, and it is not coincidental. The underlying structure of distraction has not changed; what has changed is that an entire economy now depends on inducing and sustaining it.

The monks understood this restlessness as a spiritual problem requiring a disciplined response. They developed practices such as fixed hours of reading, the slow repetition of texts, the physical rhythm of lectio divina, designed to train the wandering mind back toward sustained presence. They were, in effect, building a technology of attention from the inside out.

What social media platforms have done is the precise inverse. In 2009, Aza Raskin, the designer who invented infinite scroll, described it as a way to remove “stopping cues”; those natural moments of friction where a user might pause, reflect, and choose to do something else. The design philosophy behind it owes more to B.F. Skinner than to any tradition of human flourishing. Skinner’s operant conditioning chambers, colloquially known as “Skinner boxes”, were designed to keep rats pressing a lever by delivering rewards at variable, unpredictable intervals. The News Feed, the infinite scroll, the pull-to-refresh gesture are in essence Skinner boxes rebuilt at civilisational scale. Raskin later said he regretted the invention. But by then the logic had spread everywhere, and the gesture had become so instinctive that most people no longer recognised it as a behaviour being performed upon them.

We, by contrast, tend to treat our distraction as a design problem requiring a better app. We download focus timers and website blockers, as if the difficulty were merely technical. The ancient world trained attention through practice; the modern world engineers its dissolution for profit, then sells us tools to recover what was taken. There is something almost perfectly circular about this, and almost perfectly cynical. But perhaps the deepest consequence is one that nobody designed at all. When attention becomes a resource to be optimised, difficulty becomes waste. And a culture that treats difficulty as waste will, sooner or later, stop reading anything that refuses to be easy.

But if reading is to be seriously taken as an act of resistance, it would also mean recovering a reverence for difficulty. The contemporary assumption is that difficulty is a defect, a failure of communication, an unnecessary barrier. The idea of hard books as a kind of “torture narrative”, the framing of challenging reading as punishment rather than invitation, has left us suspicious of any text that does not yield its meaning immediately. But difficulty is often the point. The resistance of the text is what exercises attention.

The urgency of this becomes clearer when we consider how much of our reading life now takes place inside systems that do not want us to read at all. All social media platforms reward speed, reaction, and performative fluency rather than understanding. They train us to skim, to respond, to move on. Increasingly, algorithmic powered systems promise to read for us: to summarise, extract, simplify, and pre digest. What is presented as convenience is, in practice, a quiet substitution of attention itself.

Notice the language. When algorithmic reading tools are marketed, they are never described as replacements for thought. They are described as tools for “empowerment.” They “unlock” knowledge. They “democratise access.” The rhetoric is borrowed, almost verbatim, from the liberation movements of the twentieth century, but redirected toward the consumption of pre-digested information. What was once a political demand, that everyone should have the right to read, to learn, to encounter difficult ideas has been quietly transformed into a commercial promise that no one should ever have to. The struggle for access has become, by a sleight of hand so deft it is almost admirable, a struggle against effort itself.

What is happening when a machine summarises a book for us is not merely a loss of informational fidelity, as though the summary were simply a lower-resolution version of the original. It is the severing of the relationship between understanding and time.

Understanding, in the Weilian sense, requires duration. It requires the reader to undergo something, to be pulled through the resistance of a difficult argument, to feel the weight of an accumulating narrative, to sit with ambiguity long enough for it to resolve into insight. A summary, however accurate its propositional content, delivers conclusions without the journey that makes those conclusions meaningful.

A culture that stops reading in this deeper sense does not just lose books. It loses the capacity to confront and sit with difficulty, to tolerate uncertainty, and to allow meaning to emerge slowly. In such a culture, attention becomes brittle, thought becomes shallow, and the inner life gradually thins out, like soil from which too much has been extracted and to which nothing has been returned.

Weil wrote that “unmixed attention is prayer”. I am not a religious man, but I increasingly find that place of quiet contemplation is a fundamental part of human flourishing, and one that is slowly dwindling away.

An Unhurried Refusal

John Keats, whose portrait by Severn captures exactly this quality of absorbed attention, gave us another language for it. His “negative capability”, the capacity to remain “in uncertainties, mysteries, doubts, without any irritable reaching after fact and reason” is, in essence, Weil’s attention under a different name. Both describe a mode of being in which the self’s urgent demand for mastery is suspended, and something else. What we might call receptivity, openness or grace, is allowed to enter. We are transformed not by grasping but by waiting.

Reading, at its best, is an education in this kind of attention. It teaches us to wait, to dwell, to let meaning emerge in its own time. It forms in us the capacity to be present to something other than ourselves. And here the private discipline reconnects with the public necessity. The person who has learned to attend; to hold still in the presence of difficulty, to listen before responding, to let an unfamiliar idea exist without immediately assimilating or rejecting it, is the person upon whom civic life depends. Deliberation is not merely the exchange of opinions; it is the willingness to be changed by what one hears. That willingness is a form of attention. And attention, as Weil understood, must be trained.

The behavioural economists who designed the attention economy borrowed a phrase from the political philosopher Cass Sunstein: “choice architecture.” The idea was that you could design environments that nudged people toward better decisions. What happened instead was that the architecture was optimised not for better decisions but for more decisions; faster, more reactive, less considered. The feed is a choice architecture designed to prevent the one choice that matters most: the choice to stop, to think, to attend. In this light, the person who sits down with a book is not merely engaging in a private leisure activity. They are defecting from an architecture of compulsion. They are, in the smallest and most stubborn way, rebuilding the conditions under which deliberation becomes possible.

Severn's portrait of Keats shows a young man reading. That is all. He is not performing comprehension or curating insight for an audience. He is simply there, with the book, in a posture of absorbed and unhurried attention. It is the most ordinary image imaginable, and today it looks almost countercultural. Weil would have recognised what Keats is doing in that painting. She would have called it a form of waiting. And she would have understood, I think, that in a civilisation that has made waiting almost impossible, the capacity to sit with a book; unoptimised, unmeasured, present, is not a nostalgic indulgence. It is the precondition for every other form of seriousness we might still hope to recover.

His answer was the Law of Requisite Variety. Buried in Chapter 11, amid discussions of game theory and information channels, Ashby made an observation that is a core axiom of instructional design.

“If R’s move is unvarying, so that he produces the same move whatever D’s move, then the variety in the outcomes will be as large as the variety in D’s moves. D now is, as it were, exerting full control over the outcomes.”

Put simply, Ashby was describing a game between two players: R, (the regulator) trying to achieve a particular outcome, and D, (the source of disturbance). The regulator wants to keep the outcome within acceptable bounds. The disturbance wants to push it outside them. Who wins depends on one thing: whether R has enough different responses to counter D’s different moves.

This condition applies wherever one system attempts to shape the behaviour of another. Teaching and instructional design, at its core, is an attempt to stabilise particular cognitive outcomes in the face of learner variation. It is in essence, a regulatory enterprise.

Consider a child learning to read in a reception classroom. The teacher is R: the regulator, trying to ensure the child learns to decode words by mapping graphemes to phonemes and blending them fluently. The child’s repertoire of strategies is D: the source of variation. The child has a range of strategies; they can sound out the letters properly, or guess from the picture, or memorise the shape of the word, or rely on the first letter and fill in the rest from context, or simply repeat what a classmate has just said. Five strategies, perhaps more.

The teacher asks the child to read a word from a simple book. The word is horse. The child says “horse.” The teacher hears the correct response and moves on. One response. Correct or incorrect.

Ashby’s law tells us what follows. If the regulator’s move is unvarying, or to put it more simply, if the only thing being checked is whether the word sounded right, then the variety in the child’s strategies determines the variety in the outcomes. If the child guessed from the illustration, or from the first letter and context, that strategic variety has passed through unblocked. The outcome variable is satisfied. The internal process is unregulated.

In that moment, control has shifted. The teacher believes reading has occurred. The system records success. But the child may not have decoded anything at all. The appearance of learning is preserved, however the underlying cognitive operation is not.

The law can be stated in a single sentence: to control a system, you must have at least as many possible responses as the system has ways of going wrong. Or as Ashby put it even more tersely: “only variety can destroy variety.”

In other words, if learners have more ways to succeed than your system has ways to detect and block easy shortcuts, the shortcuts will win. Not because learners are lazy or devious, but because they are rational agents navigating a system that has left paths of least resistance wide open. Learning will flow toward whatever strategy produces the right answer with minimal effort. This is not a character flaw in students; it is a predictable consequence of inadequate regulatory variety. The fault lies not in the student but in the design.

Why Differentiation Fails

One strategy that attempted to address low-variety regulation, or the tendency to check only whether answers are correct rather than how they were reached, in conventional classrooms was differentiation. Faced with obvious variation between pupils, the instinct was to increase variation on the instructional side: different tasks for different groups, scaffolded worksheets, extension problems, adjusted levels of difficulty. In principle, this appears to expand the teacher’s repertoire of responses.

In practice, it often failed because it mistook variation in performance for variation in strategy. It multiplied tasks without increasing the system’s capacity to detect how those tasks were being solved.

Ashby’s law is not about surface diversity. It is about whether the instructional system’s variety matches the disturbance that actually matters. If the disturbance lies in strategy, in how the child arrives at the answer, then varying the difficulty of the text does little to address it. A child who guesses from pictures will continue to guess from pictures on an easier book and also on a harder one. A pupil who has learned to pattern-match in mathematics will pattern-match on the basic questions and on the extension ones.

The deeper problem is informational. An instructional system can only regulate what it can detect and many learning environments rely on a channel of extremely low capacity: correct or incorrect.

Correct or incorrect carries almost no information about process. It does not distinguish decoding from guessing, understanding from memorisation, reasoning from elimination. If the system cannot detect how success was achieved, it cannot constrain how success was achieved. The range of strategies available to the learner, what Ashby calls “the disturbance space”, remains uncontrolled.

This is why so much classroom performance is fragile. Students appear fluent in the moment, yet falter when the context shifts. They were “regulated” at the level of outcome, not at the level of process. The system ensured that answers were right often enough, but it never ensured that the right thinking had occurred. The system is, in Ashby’s terms, informationally impoverished; and no amount of pedagogical enthusiasm can compensate for what the channel cannot carry. In other words, you can’t out-teach a system that cannot see.

When Correct Answers Conceal Cognitive Failure

I’ve written before about the problem of students getting answers “correct” but learning very little. From a cybernetics perspective, a binary answer of correct or incorrect is a channel with low capacity. It tells us little about how an answer was produced. And if a system cannot detect how success was achieved, it cannot constrain how success was achieved. In Ashby’s terms, the disturbance retains control.

This does not mean binary assessment is inherently flawed. In some cases, it is entirely sufficient. If the task tightly aligns with the target cognitive operation, and there are few viable shortcuts, then correct or incorrect may be all the regulation required. A simple factual retrieval question asked in open response form leaves little room for alternative strategies. The disturbance space is small. The regulator has enough variety. Right or wrong might also be a very good way of building knowledge and diagnosing misconceptions early.

The problem arises when the disturbance space is large. By disturbance space, I mean the range of different ways a learner can approach a task while still arriving at an acceptable answer. It is the full repertoire of strategies, shortcuts, heuristics and workarounds available in that moment.

When multiple cognitive strategies can produce the same correct answer, binary scoring becomes blind. Recognition can masquerade as recall. Elimination can masquerade as understanding. Pattern familiarity can masquerade as reasoning. The system records success, but it has no way of knowing what produced it.

Consider a pupil revising the causes of the First World War. The retrieval question reads:

Which of the following was a long-term cause of the war?
A. The assassination of Archduke Franz Ferdinand
B. Militarism
C. The sinking of the Lusitania
D. The Treaty of Versailles

The pupil selects B. But what just happened cognitively? Perhaps the pupil genuinely retrieved a schema in which militarism, alliances, imperialism and nationalism form an interconnected explanation of escalating tensions across Europe. That is one possibility.

Or perhaps the pupil simply remembered that the assassination was a trigger event, not a long-term cause. Two options can be eliminated immediately. The Treaty of Versailles obviously came after the war. The sinking of the Lusitania happened later in the conflict. “Militarism” is the only one that sounds plausibly abstract and structural. So B must be right.

The correct answer is the same. The cognitive operations are not. In one case, knowledge has been retrieved and integrated. In the other, superficial cues have been exploited. Binary scoring cannot distinguish between them. The regulator sees success. The strategic variety that produced it remains invisible.

This in part explains a fundamental problem in formal schooling; students to perform well on class assessments over the year and then underperform in an exam. It is tempting to attribute this to nerves, poor revision habits, or exam technique, and sometimes that is true. But often something more structural is at work.

During the year, the disturbance space was constrained. Question types were familiar. Contexts were predictable. Teachers gave hints, prompts and scaffolds. Retrieval was supported by cues embedded in the classroom environment. In effect, the regulator tolerated a degree of strategic shortcutting because the surface behaviour looked acceptable.

Ashby’s law does not condemn binary scoring. It simply reminds us that regulation requires sufficient variety. When the disturbance space grows and the regulator remains simple, control shifts. The learner adapts. The system records success. Learning, meanwhile, may remain fragile.

The Promise and Peril of Adaptive Systems

This is where adaptive learning platforms and AI tutors become genuinely interesting, and where Ashby’s framework becomes most revealing about their limitations.

In principle, these technologies offer something that traditional instruction cannot: a dramatic expansion of regulatory variety. An AI tutor can observe not just whether the answer is correct, but how long the learner took, what errors they made along the way, whether they asked for hints, how they responded to follow-up probing, whether their pattern of success suggests genuine understanding or strategic elimination.

Where a paper test carries one bit of information per problem, correct or incorrect, an intelligent system can potentially harvest dozens of signals per interaction. Where a human teacher monitoring thirty students can respond to perhaps a handful of strategic variations per lesson, an adaptive system can in principle track each learner’s idiosyncratic patterns continuously.

That is regulatory variety of a genuinely different order. The channel widens. More of what actually matters becomes, at least in principle, visible.

But increased variety in principle is not the same as requisite variety in practice, and this is where the current generation of adaptive platforms reveals its limits. Most adapt along a single dimension: difficulty. If the learner performs well, the system serves harder questions; if performance drops, it serves easier ones. This is regulation of a kind, but of a peculiarly narrow kind. It addresses one dimension of variation, the learner’s current performance level, while leaving strategy entirely unregulated.

A learner who has discovered that process of elimination works reliably on multiple-choice questions will sail through such a system. The platform will interpret sustained correct answers as mastery and serve progressively harder material. The learner will apply the same shortcut to those as well. The system will report that learning has occurred. The learner will have acquired nothing except the knowledge that process of elimination works. The regulator has variety; it simply does not have the variety that matters.

This is the central disappointment of much educational technology: not that it lacks sophistication, but that its sophistication is deployed in the wrong direction. Engagement metrics, gamification, personalised difficulty curves, these are genuine technical achievements. They are also, from Ashby’s perspective, regulatory activity aimed at the wrong variable. A system that optimises for time-on-task while leaving cognitive strategy unmonitored is not solving the problem; it is measuring a proxy for it and mistaking the proxy for the thing itself.

Designing for Requisite Variety

Taking Ashby’s law seriously means treating instructional design as an engineering problem rather than an art, and that shift in framing has real consequences for how we build and evaluate educational systems.

It begins with a question that many educators do not ask systematically: how else could a learner succeed here without doing the intended thinking? Every task has alternative solution paths, and those paths will be found. The designer’s job is to map them and close them, not by making tasks harder, but by making shortcuts unreliable. A task well-designed against Ashby’s law is one in which the target cognitive operation is not the easiest path to the correct answer; it is the only path. This is an instructional invariant; a non-negotiable condition of effective design, as fixed and indifferent to our preferences as any law in engineering.

All this requires a somewhat different conception of assessment. The question is not “did they get it right?” but “what would they have to have done cognitively to get it right in this way?” Diagnostic power matters as much as correctness. Requiring worked solutions, varying problem formats so that pattern-matching fails, asking for explanations that cannot be generated without understanding, probing with follow-up questions that are only answerable if the first answer was genuinely reasoned rather than guessed: these are not merely good pedagogical habits. They are mechanisms for expanding the instructional system’s regulatory variety to match the strategic variety of learners.

For adaptive systems and AI tutors, the implication is architectural rather than cosmetic. Adjusting difficulty is insufficient if the learner’s strategy works across difficulty levels. The system must adapt to strategy, not merely to performance: detecting when a learner is pattern-matching and breaking the pattern, detecting when a learner is guessing and making guessing unreliable, detecting when a correct answer was produced by a process that will not transfer and probing until that becomes apparent. This is not a matter of prompting an AI more cleverly. It is a matter of building diagnostic and constraint mechanisms into the system’s design from the outset.

And then, finally, there is the upper bound that no amount of ingenuity can dissolve. No system can regulate what it cannot detect. Learner ingenuity will always exceed designer foresight; there will always be shortcuts that were not anticipated, strategies that were not mapped, paths that were left open by accident. Requisite variety is an asymptote, not a destination. The honest response to this is not despair but iteration: observe what learners are actually doing, diagnose which strategies are succeeding without learning, close those paths, and observe again.

The Upper Bound: What the System Cannot See

Ashby was not thinking about classrooms. He was thinking about thermostats, organisms, and the abstract mathematics of control. But I find this engineering mindset fascinating when applied to learning and I think his insight cuts across domains because it concerns the logic of regulation itself, and that logic applies wherever one system attempts to constrain another.

Instructional design is, at its core, an attempt at regulation. We want particular knowledge structures to form and consolidate; we want others not to. The learner arrives with immense strategic variety, countless ways of navigating any task we set, many of which lead to the appearance of success without the substance of it. The instructional system is the regulator: its job is to match that variety, to close the paths that lead to performance without learning, to ensure that the only reliable route through the task is the cognitive operation we actually intend.

If the system cannot do this, the learner’s variety determines the outcome. The platform reports mastery. The assessment records a pass. Learning, in any meaningful sense, may not have occurred.

This is not a counsel of despair about teachers, or technology, or learners. It is a precise description of what instructional design must accomplish and why it so often falls short. The problem is not effort or intention; it is regulatory variety. The system checks too little, detects too little, and consequently controls too little. Learners, being rational, take the paths that are left open.

“Only variety can destroy variety.”

A reminder to Australian readers that I will be in Melbourne, Sydney, Canberra and Auckland next month. More info here.

Subscribe now