The Docile Learner

Disclaimer : This text was co-written with an LLM. I mention it both out of formal transparency, and because the article below raises precisely the question of the relation between the subject and the apparatus that models him — and it seemed dishonest not to signal that the "I" of this post is itself a relation, not a sovereign subject.

There is a scene that repeats itself, at this hour, in several million browser tabs. A learner sits before a screen. The screen asks a question. The learner answers. Somewhere — in a server that nobody has ever visited, in a Dutch or Irish data centre cooled by the North Sea — a probability is updated. A number that did not exist a moment ago now exists, and it stands for the learner. The number is small: a scalar between zero and one, or a vector of a few dozen dimensions. But it is, for the duration of the session, who the learner is to the machine. The next question will be selected on its basis. The next hint, withheld or granted, on its basis. The next suggested lesson, the next notification, the congratulatory animation of a green owl or a red flame — all of it flows from this number.

I want to describe this scene not as pedagogy and not as technology but as a dispositif: a configuration of knowledges, practices, architectures, and discourses that produces, as its effect, a particular kind of subject. The student. Not the student as she experiences herself — tired, curious, distracted, ambitious — but the student as the apparatus must know her in order to act upon her.

I. The two loops, or: the examination generalised

Twenty years ago, Kurt VanLehn described the intelligent tutoring system as a machine of two loops. An outer loop, which selects the next task. An inner loop, which runs inside the task, dispensing feedback and hints step by step. The description was offered as a technical clarification, a taxonomic gift to a field that had grown confused about its own objects. I read it as something else. I read it as the most precise description we have of a Foucauldian examination machine — not the examination as a single event (the viva, the baccalauréat, the concours) but the examination dissolved into time, spread thin across every keystroke, reconstituted as the continuous medium of learning itself.

Recall what the examination did, in the disciplinary society. It rendered the individual a describable, analysable object. It placed each person in a field of surveillance, documented her, permitted comparison, made her a case. The examination combined the ceremony of power with the ritual of experiment, the deployment of force with the establishment of truth. It was, Foucault wrote, a ritualisation of that which was most banal in modern power: the fact of being seen.

The inner loop is the examination that no longer requires a ceremony. It requires only a text box. Every step of a proof, every line of code, every conjugation of a French verb — each is an examination. Each is recorded. Each updates a number. There is no moment at which one is not being assessed; assessment is the ambient condition of the activity itself. One does not pass or fail; one is traced.

The vocabulary has obligingly shifted to mark this. The old word was testing. The new word is tracing. Bayesian Knowledge Tracing. Deep Knowledge Tracing. Half-Life Regression. The student is no longer photographed at discrete intervals; she is filmed continuously, at thirty frames per second, and the film is what she is.

II. The subject of the model, or: what the apparatus must presuppose

Every model of the student is a philosophical anthropology in disguise. It says: this is what a learner is, this and nothing more. The question is always worth asking — what is the learner, for this system to be able to function? — because the answer is never innocent, and the answer is never what the system says about itself.

Consider Bayesian Knowledge Tracing, the canonical tool. It presupposes that knowledge is a set of binary variables: mastered, or not. There is no such thing as partial understanding, no such thing as the knowledge that hovers, the intuition that cannot yet be defended. A skill is a bit. It presupposes that skills are independent of one another — that knowing how to factor a polynomial tells the system nothing about whether you understand what a polynomial is. It presupposes that all students learn at the same rate: the transition probability from ignorance to mastery is a single number, fitted once, applied to every child in the cohort. It presupposes, classically, that one does not forget. It presupposes that the only thing the apparatus can see of you — the only thing it needs to see of you — is whether you got it right or wrong.

Item Response Theory, the older and more austere cousin, presupposes something stranger still: that there is, inside you, a scalar called ability, denoted θ, and that your performance on any question is a monotonic function of this scalar and of the question's difficulty, plus a little noise. You are a number, in IRT. The number is estimable, invariant across populations (in principle), separable from the items on which it is measured (in principle). It is a remarkable theoretical achievement, and it is an ontological commitment of staggering confidence. An entire person, for the purposes of the examination, is one real-valued parameter. And this model — let us be clear — is the model that sits beneath the SAT, the GRE, the TOEFL, PISA. It is the model by which tens of millions of lives are sorted each year. It works. By its own criteria, it works extraordinarily well.

What does it mean that it works? It means the apparatus has successfully produced a population for which it is true. The students have learned — and here is the reflexive turn I want to insist on — to be the kind of subject the model requires. They have learned to answer in units the machine can score. They have learned to fail in ways the machine can interpret. They have learned, most profoundly, to accept that θ is a thing they have.

This is not a conspiracy. No one designed it. It is simply what an apparatus of that scale, operating over that duration, does. Foucault's point about the prison, the clinic, the barracks was never that anyone was scheming. It was that the architecture thinks on behalf of those within it, and that the thinking becomes, in time, indistinguishable from their own.

III. The pastoral algorithm, or: Duolingo and the half-life of you

Duolingo is the most beautiful recent instance of this because it is the friendliest. Its mascot is an owl. Its interface is chirpy. Its underlying model, for the period when they published about it, was called Half-Life Regression — a predictor of the probability that you will recall a word, given the time since you last saw it, fitted by regression over thirteen million user traces. The paper, by Settles and Meeder (2016), is a small masterpiece of its genre. It reports an operational lift of 12% in daily engagement.

Notice what has happened. The Ebbinghaus forgetting curve, a psychological finding from 1885 about the decay of nonsense-syllable memory, has become an industrial-scale scheduling engine for attention. Your forgetting is the raw material. Each word you learn has a half-life; when the half-life approaches, you are notified. The notification is friendly, the streak must be kept, the owl is mildly passive-aggressive. You return. The model was right about when you would forget — or rather, the model, by acting on its prediction, made you return at the moment of predicted forgetting, which is a slightly different and more interesting thing. It is a self-fulfilling psychometrics. The forgetting curve, having been modelled, now disciplines the time of your day.

This is what Foucault called pastoral power, translated into the cloud. Pastoral power individualises (it cares about each sheep), it totalises (it governs the whole flock), it extracts truth (each sheep must confess), and its justification is always the salvation — the learning, here — of those it governs. The pastor does not compel; he solicits. He does not forbid; he invites you to keep your streak. He is gentle. He is always available. He never sleeps. He holds, in a growing file, the shape of your forgetting.

IV. Delegation, or: the subject who no longer wants to be examined

And now a new figure has entered, and it is worth pausing over, because I think it is genuinely novel and not yet well understood. The large language model is not a tutoring system. It does not have the two loops. It does not maintain a student model. It will answer your question whether you have the prerequisites or not; it will solve your problem for you, if you ask it to; it will hallucinate, with complete confidence, an answer that is wrong in ways you are not equipped to detect. It is, from the standpoint of the pedagogical apparatus, a catastrophe. From the standpoint of the student, it is often a relief.

A recent study — Shen and Tamkin, at Anthropic — watched twenty-three programmers try to learn a new library with and without AI assistance. They clustered the behaviour into six patterns. Three of the patterns produced quiz scores below forty percent. Those three patterns shared a structural feature: the learner had delegated the cognitive work to the machine. They had asked it to write the code. They had asked it to fix the bug. They had asked it, iteratively, to debug its own output. They finished the task quickly, often impeccably. They did not, afterwards, know what they had done. The other three patterns — the ones that preserved learning — all involved the learner asking the model to explain, to decompose, to check her own understanding. Same tool. Opposite outcome.

I do not think the interesting claim here is the obvious one, which is that students should not use AI to do their homework. That is true and has already been said a hundred times. The interesting claim is the Foucauldian one: the large language model has given the student, for the first time in the history of mass education, a way out of the examination. Not a way to cheat on it — cheating is a move inside the game — but a way to render the examination structurally obsolete by delegating the act that the examination was meant to certify. The two-loop apparatus of VanLehn assumed a student who was trying to acquire the skill. The inner loop made sense because mastery was the goal. If mastery is no longer the goal — if the goal is only output, the essay, the working code, the completed exercise — then the inner loop is a humiliation, a pedantic slowness that the student is right to escape.

The apparatus has not yet caught up. It still tries to trace. It still updates its Bayesian mastery estimates based on a response that, increasingly, was not produced by the student at all. The number that stands for the learner, in the server in the North Sea, is tracking a subject that is partially absent. What it is tracking, in many cases, is the learner's skill at prompting — a skill which the apparatus has no concept for, which is not in the model, which will eventually require a new model, and a new science of the prompter, and a new regime of examination.

V. On scaffolding, and its fading

There is a concept in the learning sciences called scaffolding, borrowed from Vygotsky through Bruner. The idea is lovely. An expert provides temporary support, calibrated to the edge of what the learner cannot yet do alone, and withdraws it — fades it — as the learner internalises the skill. The scaffold exists in order to be dismantled. A pedagogy in which the scaffold is never withdrawn is not a pedagogy; it is a crutch, a dependency, a regression in the technical sense.

What is an AI coding assistant, when it writes the function for you, if not a scaffold that refuses to fade? Or rather: a scaffold that the student refuses to let fade? The problem is not the scaffold. Scaffolding is how anyone ever learned anything. The problem is the historically unprecedented ease with which the scaffold can now be kept permanently in place, the way a person in 2026 can go through an entire computer science degree without ever having to sit, alone, with an error message for more than forty seconds. The inner loop of the tutoring system, in its classical form, was a regulated scaffold: it knew when to withhold. The large language model is a scaffold with no theory of withholding. It gives, because giving is what it was trained to do, because the reward signal in its training approximated helpfulness, and helpfulness, operationalised, means: do not make the user struggle.

There is a deep irony here. The entire research programme of intelligent tutoring — all those Bayesian networks, all those carefully calibrated hint policies, all those papers about productive struggle and the zone of proximal development — was precisely about when to withhold. It was a science of pedagogical parsimony. And it has been outflanked, in the consumer market, by a technology whose defining feature is that it does not withhold anything from anyone, ever.

The student who masters this new situation — the student who will be, in some sense, the subject of the next educational apparatus — is the one who imposes the scaffold's fading on herself. Who, when the AI offers the code, asks for the explanation instead. Who uses the model as a Socratic interlocutor and not as a vending machine. This is a kind of ascesis, in the old sense: a self-imposed discipline, a care of the self. It is not automatic. It cuts against the grain of the tool. It has to be learned, and it has to be learned without an examination apparatus to enforce it, because the examination apparatus is the thing the tool has outflanked.

What an astonishing pedagogical moment. We have arrived, by the most circuitous technological route, at the Stoic problem. Will you govern yourself, when nothing external compels you to?

VI. On the number that stands for you

I want to end where I began, with the scalar in the server. It is worth being precise about what it is, because the critique that says "we have reduced the student to a number" is both true and boring. The student was already reduced to a number by the SAT, by GPA, by class rank. The question is not whether reduction occurs; reduction is what institutions do. The question is which number, fitted by which procedure, under which hidden assumptions, with what feedback loop onto the thing it claims to measure.

The hidden assumptions of BKT: mastery is binary, skills are independent, students are exchangeable, one does not forget, guessing is stationary, the knowledge-component tagging is correct. Every one of these is, in the general case, false. The model works anyway, sometimes, for the narrow purpose of predicting the next response, because the world conspires to make it approximately true inside the narrow slice of activity the model observes. The student who finishes the drill has, for the system's purposes, mastered the skill. Whether she has mastered it in any sense that matters a year later — in any sense that transfers to a problem she has not seen before — the model does not ask, because the model was not built to ask.

The hidden assumption of IRT: you are, in the relevant respect, a scalar. A point on a line. The line has been calibrated against a population, and your position on it determines what the institution does with you. There is nothing wrong with this, in the way there is nothing wrong with the metre or the kilogramme. It is a measurement convention. But the measurement convention for mass does not change the mass of what is weighed. The measurement convention for student ability has, over fifty years, produced populations of students who experience themselves through the scalar, who describe themselves as being at a certain level, who orient their study toward moving the scalar upward. The measurement, in the human case, is not separable from the measured. This is the specifically Foucauldian twist, and it is why the question what model does the apparatus use is never merely a technical question.

The docile learner, then, is not a person who has been coerced. She is a person who has been modelled, who has been invited to recognise herself in the model, who has — often gratefully, often with real relief at having been seen at all — accepted the model as an image of herself. The green owl sends her a notification at the hour of her predicted forgetting. The quiz question is selected by the algorithm because her estimated θ places it in her zone of proximal development. The AI assistant completes her function because it has inferred, from the context, that this is the completion she would have written herself. Each of these interventions is, individually, a kindness. Each is a reduction of friction. Each is a small act of pastoral care.

Together, they constitute a form of life. Whether it is a form of life in which learning — in the old, difficult, self-transforming sense — still takes place is an empirical question. It is also, I think, a political question, and a question about what kind of people we are trying to become. I do not think the answer is given in advance. But I do think it is worth noticing that the answer will be given, in the main, by the people who design the models, and that those people are currently, at this hour, updating parameters.

VII. The norm tightens, or: what becomes of the child who does not fit

I want to close by following a thread that I have left implicit until now. Every predictive model of a learner is built on a statistical population. θ is a position on a line calibrated against a reference group. The transition probabilities of BKT are fit to a cohort. Half-Life Regression learns its coefficients from thirteen million users. The model is, in the exact sense, the norm made operational — not as an ideal to which students are asked to conform (this was the old disciplinary mode, the schoolmaster at the blackboard, the bonne conduite marked in red ink) but as a probability distribution whose tails are, by construction, sparsely populated and poorly served.

Canguilhem taught us to be careful with the word normal. It names two things which are usually conflated: the normative, what ought to be, and the statistical, what is on average. The great trick of the modern era, he said, was to slide from the second to the first without anyone noticing the movement. The apparatus of knowledge tracing does this slide at machine speed, millions of times per second. The model's prediction of what the student will do next becomes, by feeding back through the hint policy and the task selector, a soft instruction about what she should do next. The students whose behaviour lies near the mean of the training distribution receive a pedagogy exquisitely tuned to them. The students in the tails receive a pedagogy tuned to someone else.

There is a more recent and more precise way of saying this, which comes from Deleuze. The disciplinary society moulded subjects; the society of control modulates them. Mould is a rigid form; modulation is a continuously adjusting parameter. The student of the examination era was shaped against a template. The student of the knowledge-tracing era is tracked against a moving estimate of her own trajectory, which is itself a projection from a population's trajectories. The effect is gentler and much more total. There is no longer a form to break; there is only a gradient to descend. To escape the apparatus one must become, somehow, unpredictable to it — and unpredictability, in a large enough system, is simply residual variance. It is noise. It is what the model has not yet learned to absorb.

Here is where the question of the child who does not fit becomes sharp. In a classroom full of children, the teacher's attention is a scarce good allocated by many criteria — some pedagogical, some institutional, some frankly social. In a classroom mediated by an adaptive learning platform, the platform's attention is not scarce in the same way; it is allocated by the model's uncertainty. And the model, by design, is most uncertain about the children who deviate from its priors. This is the paradox worth naming. The child who is legible to the model — who follows the expected trajectory, whose errors are the expected errors, whose knowledge state updates as the model anticipates — receives, in a certain sense, the least human intervention. The pastoral algorithm handles her. The teacher, freed by the platform, attends to the children the platform cannot handle: the ones whose responses surprise it, whose patterns of error do not factor into the known knowledge components, whose silences are of the wrong length, whose attention is distributed in ways the system was not trained to recognise.

The irony is considerable. The apparatus, in its effort to individualise education, produces a population in which legibility-to-the-algorithm is inversely correlated with legibility-to-a-human. The children who conform to the statistical assumptions get the algorithmic pedagogue, which is efficient and patient and never tired and never really there. The children who break the statistical assumptions — who ask the question the system did not expect, who fail in an interesting way, who refuse the task, who answer in the wrong register, who take too long, who take no time at all — these are the children who get the human. Not because they are preferred, but because they are the residue the machine cannot process.

I do not want to romanticise this. There is nothing liberatory about being the child whose behaviour the algorithm cannot classify. Often it is painful; often it is a diagnosis; often it is a referral, a specialist, a form filled out by a worried parent. But I want to insist that a strange incentive landscape has opened up, which the pedagogical discourse has not yet caught up with. In the twentieth-century school, human connection was a default that had to be rationed. In the twenty-first-century school, human connection is — increasingly, unevenly, across institutions with very different resources — an exception, granted to those whose legibility to the machine has failed. The statistical norm and the zone of human attention have come apart.

What does this mean for the children who are, in the current diagnostic vocabulary, neurodivergent? I want to be careful here, because this is the kind of question that lends itself to overclaiming in several directions at once. Autism and ADHD and the rest are not strategies; they are not chosen; they are not heroic ruptures with normativity. They are heterogeneous developmental conditions with their own phenomenology, their own difficulties, their own forms of suffering, and their own forms of flourishing. To make them the protagonists of a narrative about the failure of algorithmic pedagogy would be to instrumentalise them for someone else's critique. That is not what I want to do.

What I do want to note is more modest, and I think more defensible. The apparatus is, right now, sorting children along an axis that coincides substantially, though not perfectly, with what we currently call neurotypicality. The child who processes the inner loop smoothly, who updates her knowledge state in the shape the HMM expects, who forgets on the Ebbinghaus curve and returns when the owl calls — this child is, statistically, the neurotypical child. The child whose processing does not fit the model is, disproportionately, the neurodivergent child. And so, without anyone having decided this, without any policy anywhere saying this, the apparatus has begun to draw a line, and the line tracks — imperfectly, but legibly — a line that neuroscience and diagnostic practice have been drawing independently, for their own reasons.

Now — and here is where I want to be most careful, because the speculation in your question deserves a more honest answer than biological determinism would give — what happens to a population over time when this line is institutionalised?

If the question is Darwinian, in the strict sense — does the proportion of autistic alleles in the human gene pool increase because of adaptive learning platforms? — then the answer is no, or at least: not on any timescale we can reason about. Human generations are twenty-five years; meaningful allele frequency shifts require hundreds of them; the entire history of digital pedagogy is one generation old. Anyone telling you the SAT or Duolingo is changing the human genome is selling something.

But the question admits a subtler formulation, which I think is the one you are reaching for, and which recent biology has made less speculative than it would have been a decade ago. Development is not a read-out of the genome. It is a dialogue between genome, epigenome, and environment, and the epigenome — the pattern of methylations and histone modifications that regulate which genes are expressed when — is responsive to environmental input, particularly in early childhood, and is, under certain conditions, partially heritable. The literature on this is contested and moves quickly; I don't want to overclaim. But the broad outline is no longer fringe: experiences of stress, of attention, of nutritional environment, of social embedding — these leave epigenetic marks, some of which persist and some of which transmit.

A child raised within an algorithmic pedagogy that subtly rewards legibility-to-the-model, who receives less human attention when she conforms and more when she does not, who learns — as all children do, with exquisite sensitivity — what kinds of behaviour summon a human face and what kinds summon only the screen, is developing inside an environment whose selection pressures differ from those of any previous educational system. It is not impossible that, over several generations of this, the developmental trajectories we currently group under the heading of neurodivergence become more common. Not because the genes have changed; they will not have. But because the environment in which genes express themselves has changed, and because the phenotypes that were once costly have become, in this new niche, less costly or even — in the narrow sense of attracting human attention — advantageous.

I offer this with all the hedges it deserves. The mechanism is plausible; the magnitude is unknown; the timescale is generational; the evidence is not yet in. What I am confident of is the smaller claim, which does not need epigenetics to be true: we are currently running, at global scale, an experiment in which human attention to children is being reallocated along an axis of algorithmic legibility, and we have not begun to think about what this does to the children. The rise of neurodivergent diagnosis over the past two decades is over-determined and has many causes — better awareness, broader criteria, genuine increases, diagnostic substitution — but one of the causes we have not yet named, because the apparatus that would be causing it is so new, is the apparatus itself. The apparatus is not neutral with respect to the kinds of minds it makes visible. It makes some minds into signal and other minds into noise. The noise, increasingly, is where the humans are.

This is not, finally, a prediction. It is an attentiveness. A Foucauldian genealogy does not forecast; it notices what has already happened by the time anyone thought to ask. What has already happened is that the statistical assumptions inside our tutoring models — assumptions about what a skill is, about what learning looks like, about how forgetting works, about how children attend — have become, through millions of small pastoral acts, the implicit shape of childhood. The children who fit this shape recede into the background of their own education, handled. The children who do not fit it come forward into relation with the humans around them, whether they want to or not.

Somewhere in this, there is a politics we have not yet written.

Bibliographical note. VanLehn's "The Behavior of Tutoring Systems" (IJAIED, 2006) is the canonical reference for the inner/outer loop framing. Corbett and Anderson (1994) introduced Bayesian Knowledge Tracing; Yudelson, Pavlik and Koedinger (2013) document its individualisation limits; Doroudi and Brunskill (2017) discuss identifiability. The Settles and Meeder (2016) Half-Life Regression paper is in the ACL proceedings. Shen and Tamkin, "How AI Impacts Skill Formation" (Anthropic, 2026), is the study on delegation patterns. Deonovic et al. (2018) work out the formal relationship between BKT and IRT. On scaffolding, Wood, Bruner and Ross (1976). For the Foucault of all this: Discipline and Punish, the Collège de France lectures on governmentality, and the late essays on the care of the self. On the slide from the statistical to the normative: Canguilhem, Le normal et le pathologique (1966). On modulation versus discipline: Deleuze, "Postscript on the Societies of Control" (1990). On transgenerational epigenetics, the literature is genuinely contested; Bohacek and Mansuy (2015, Nature Reviews Genetics) is a reasonable entry point that does not overclaim.