Abstract
Contemporary discourse about reading and writing treats each as a single measurable capacity whose erosion can be tracked through statistics on time allocation, comprehension scores, and book sales. This article argues that psychology's confident measurements of “reading” and “writing” in its experimental setups are not merely incomplete but structurally incapable of reaching their object, because the discipline's methodology requires converting an irreducibly multimediated, multirecursive coordination into a nonrecursive surrogate before it can be measured at all. The article treats reading and writing as symmetrical cases of the same overburdened category, each compressing five irreducible mediational domains, five recursivity levels, and three fundamentally different recursivity types into a single ordinary-language word. I develop a mirrored multimediational and multirecursive anatomy of reading and writing, arguing that generative AI functions as an epistemological gift that exposes the historical coupling between symbol production and recursive discernment rather than merely a threat to it. The question of generational decline is reframed as a question about which coordinative capacities have changed. The conclusion situates reading and writing among a wider class of ordinary-language concepts, including memory, thinking, attention, and creativity, that psychology has repeatedly mistaken for stable objects.
Introduction
We do not know what reading is. Nor, in quite the same way, do we know what writing is.
Not because either is mysterious, in the way that consciousness or time or meaning might be called mysterious. Reading and writing are among the most extensively studied, most heavily funded, most institutionally consequential objects of empirical research in the modern human sciences. Departments of education, cognitive psychology laboratories, national assessment programmes, publishing industries, and now generative AI companies all produce continuous streams of data about both. The mystery is not epistemic poverty. It is the opposite. We do not know what reading and writing are because the categories themselves have never been adequately described, and the sheer volume of confident measurement has made this fact almost impossible to notice.
This article makes a stronger claim than that observation might suggest, and the strength of the claim matters for everything that follows. It is not that psychology has, so far, failed to notice what reading and writing actually are, and might one day catch up with sharper definitions or better instruments. It is that psychology's own experimental methodology forecloses the discovery in advance. To become an object that a controlled experiment can measure, an activity must first be rendered nonrecursive: its outcome must not depend, in ways the instrument cannot control for, on who is doing it, for whom, why, and within what relationship of trust and purpose. Reading and writing are not, in their actual occurrence, nonrecursive. They are constitutively selfrecursive and interrecursive, in ways developed in detail below. An instrument capable of controlled measurement can only be built by stripping out precisely the dimensions that make reading and writing what they are. This is not a contingent limitation that a cleverer experiment could eventually fix. It is a structural consequence of what counts, within psychology's inherited methodology, as a measurable object at all. Psychology cannot know what reading and writing are, not because it has not yet looked closely enough, but because its methodology requires both to be transformed into nonrecursive objects before either can become measurable.
Consider a recent and unusually influential case. In the summer of 2026, The Atlantic published a lengthy cover story by staff writer Rose Horowitch titled "The End of Reading Is Here," arguing that the United States has crossed from a literate into what she calls a postliterate condition. The piece assembles an impressive body of evidence: declining rates of reading for pleasure, falling national assessment scores, shrinking sentence lengths in bestselling fiction, survey data on how many books English teachers now assign, brain-imaging comparisons between children who listen to stories and children who watch animated versions of the same stories, and reflections from literary scholars and philosophers on what is lost when reading recedes. The article moves confidently across all of this material as though it were evidence about a single, continuous phenomenon.
It is not. And the article is exemplary precisely because it does not invent this confusion. It inherits it, faithfully, from decades of psychological and educational research that has never paused to ask whether decoding a road sign, following the plot of a young-adult novel, holding the structure of a three-hundred-page argument in mind across a week of reading, and skimming a news article on a phone are the same activity measured at different intensities, or whether they are, in fact, different activities that happen to share an English verb.
Reading and writing are not unusual in this respect, and this is the second claim this article wants to establish alongside the first. Memory, thinking, attention, creativity, learning, problem solving: psychology's entire working vocabulary is inherited wholesale from ordinary English, and ordinary English was never in the business of individuating scientific kinds. It was in the business of coordinating everyday life, a business at which words like reading and writing succeed admirably and continue to succeed admirably, which is why this article is not a proposal to stop using them. The deeper argument, developed across the sections that follow and returned to in the conclusion, is that psychology has systematically mistaken this second, coordinative kind of success, a word's fitness for organising a classroom, a curriculum, a parent's bedtime routine, for evidence of a first, scientific kind of success, the individuation of a natural object with stable causal properties. It has built an entire experimental apparatus on top of that confusion, and the apparatus works, in the sense of producing publishable, policy-relevant numbers, precisely because it never tests the assumption on which it depends.
This article extends the diagnosis developed in Unearned Confidence to this domain. That article argued that psychology's most influential popular findings, the marshmallow test, self-efficacy theory, the illusion of explanatory depth, growth mindset, achieve unearned confidence because they apply the methodology of a container ontology to phenomena that are constitutively multimediated and interrecursive. The present article sharpens that diagnosis into a stronger, more general claim about method: the container ontology is not an option psychology happens to choose and could in principle abandon while keeping its experimental apparatus intact. The experimental apparatus is the container ontology, operationalised. Wherever that apparatus is pointed, at reading, at writing, at anything else on the list above, it will manufacture a nonrecursive surrogate, measure the surrogate with impressive precision, and report the result as a finding about the interrecursive achievement the surrogate replaced.
The argument proceeds in seven movements, each treating reading and writing as strict structural parallels rather than as a primary case and a secondary extension. The first situates both alongside class as overburdened L4 categories, and widens the diagnosis to the broader class of ordinary-language psychological terms. The second traces the unnoticed conceptual drift within the Atlantic article, treating it as a diagnostic instrument for the reading side of the pair. The third and fourth show that reading and writing are each irreducibly multimediated and multirecursive, mirroring each other mediation by mediation and level by level. The fifth develops the article's methodological centrepiece: the demonstration that both are simultaneously nonrecursive, selfrecursive, and interrecursive, and the argument, sketched above, for why psychology's methodology cannot in principle reach the latter two. The sixth turns to generative AI, not primarily as a threat but as an epistemological instrument that has, for the first time, made the coupling between symbol production and recursive discernment visible by breaking it. The seventh returns to the question of generational decline. The conclusion situates reading and writing within the wider programme this series has been developing: the reconstruction of the human sciences' ontology by identifying where ordinary-language achievements have been mistaken for natural scientific objects.
Part I. The Overburdened Categories "Reading" and "Writing"
The companion article on this site, The Overburdened Category, showed that class began as a modest, low-bandwidth stabilisation in Ricardo's macroeconomic taxonomy and was progressively loaded with mediational and recursive freight until it became a category asked to explain almost everything about social life while fitting almost nothing precisely. Reading and writing have followed a strikingly similar career, though their inflation has occurred largely within a single century and across a different set of disciplines: experimental psychology, education research, library science, composition studies, and now, urgently, public commentary on artificial intelligence.
At its narrowest, reading names a nonrecursive perceptual-motor achievement: the decoding of graphemes into phonemes, or of printed symbols into their referents. Writing, at its narrowest, names the mirror-image nonrecursive achievement: the graphomotor or typographic production of legible symbols, a child's capacity to form the letter a or to strike the correct key. Both are genuine, tractable capacities, close in kind to Ricardo's class, narrow, structurally bounded, and reasonably well captured by the instruments built to measure them. Early-childhood literacy assessments test exactly these narrow senses, and they are right to.
But almost no discussion of reading or writing, in the scholarly or the popular register, stays at this level for more than a sentence. Reading immediately expands to cover comprehension, the capacity to extract propositional content from a passage. It expands further to cover retention, interpretation, synthesis across a whole book, a preference and a habit, an institutional achievement, and, in its most inflated form, a proxy for attention span, empathy, civic capacity, intelligence, even the health of democracy itself. Writing undergoes an exactly parallel inflation. It expands from graphomotor production to cover typing and dictation, then composition, the capacity to construct an argument or a narrative. It expands to cover thinking itself, memory, learning, expression, communication, and creativity. In its most inflated form, writing becomes a proxy for a person's intelligence, their moral seriousness, their fitness for a university place or a job, exactly as reading becomes a proxy for the reader's.
None of these are the same phenomenon, in the precise sense that matters for empirical research, on either side of the pair. Decoding a road sign under time pressure and spending three unhurried weeks with a difficult novel do not share a mediational profile, a recursivity level, or a recursivity type. Neither do typing a text message and drafting the fourth revision of a difficult argument. Yet the same word is used to name all of these in each case, and the same research literatures routinely move between the narrow and the inflated senses without marking the transition, exactly as class theory moved between Ricardo's factor-share taxonomy and Gramsci's claim that hegemony saturates the entirety of cultural and symbolic life.
This is what makes reading and writing overburdened categories in the precise LVT sense developed for class. A concept can be evaluated by its mediational bandwidth, the range of lived mediations it attempts to stabilise under one name. Reading has been asked to stabilise embodied habit, relational transmission between parent and child or teacher and student, the spatial and institutional conditions of libraries and classrooms, the material history of books and screens, and the symbolic apparatus of language and genre, all under one word. Writing has been asked to stabilise exactly the same range: the embodied history of the hand or the keyboard, the relational configuration of author and reader or student and examiner, the institutional spaces in which composition happens, the material history of the pen, the typewriter, the word processor, and now the language model, and the symbolic apparatus of argument and narrative construction.
The overburdening did not happen through any single theoretical intervention. It happened through accretion, on both sides of the pair, at roughly the same pace. Cognitive psychology added comprehension and working-memory load to reading, and planning and revision processes to writing. Developmental psychology added the link between early reading and later life outcomes, and between early writing fluency and later academic achievement, in moves that should by now sound familiar from the marshmallow test's confusion of interrecursive trust with individual self-control. Media theory and neuroscience added claims about attention and imagination to reading, and claims about cognitive offloading and the extended mind to writing. Education policy added achievement and equity to both. Cultural commentary, from Socrates in the Phaedrus worrying that writing would destroy memory, through worries about the typewriter and the word processor, to the present debate about generative AI, added claims about memory, civilisation, and the fate of thought itself, to both categories simultaneously, since the anxiety about writing's effect on the writer's mind and the anxiety about reading's effect on the reader's mind have always been two faces of the same worry. Each addition was locally reasonable. None was accompanied by a corresponding differentiation of either concept into the distinct coordinative achievements it was now being asked to name.
The consequence is a pair of concepts that retain enormous rhetorical and political force while losing analytical precision. To say that reading is declining, or that students can no longer write, carries the emotional weight of a claim about civilisational health, while resting empirically on a handful of narrow, mutually incommensurable measurements on each side. This gap between rhetorical scope and empirical grounding is the structural signature of an overburdened L4 category, and it invites a question this article now wants to press further than the companion piece on class did: is this overburdening peculiar to reading, writing, and class, or is it the default condition of psychology's working vocabulary.
It is the default condition. Memory, thinking, attention, creativity, learning, and problem solving are not exceptions to the pattern traced above. They are further instances of it. Each is an ordinary-language word, inherited by psychology rather than constructed by it, and each was doing genuinely useful coordinative work in ordinary language long before any psychologist attempted to operationalise it. Ordinary English did not individuate memory into procedural, episodic, semantic, and working components before psychology arrived, any more than it individuated reading into decoding, comprehension, and sustained multi-day engagement with an argument. It simply had a word, memory, that worked well enough for telling someone to remember to lock the door, and a word, reading, that worked well enough for asking a child what they did after school. Psychology's founding methodological move, across its entire history, has been to take words of exactly this kind, words optimised for everyday coordination rather than for causal individuation, and treat them as though they already picked out natural kinds waiting to be measured more precisely. This article's analysis of reading and writing is offered as one detailed case of a mistake that runs through the discipline's entire vocabulary, and the conclusion returns to what a corrected vocabulary would need to look like.
Part II. The Atlantic as a Case Study in Conceptual Drift
This section focuses on reading, whose public case study is the more developed of the two; the exactly parallel drift within contemporary discourse about student writing, and about what generative AI now does to both practices, is picked up directly in Parts V and VI below. Treating "The End of Reading Is Here" as a case study rather than a target of refutation requires a specific interpretive discipline. The claim of this article is not that Horowitch's reporting is inaccurate, or that her sources are unreliable, or that the statistics she assembles are wrong. Every individual data point she cites is plausible and well sourced: the decline in the share of American adults reading for pleasure, the fall in National Assessment of Educational Progress scores among fourth and eighth graders, the rise in the proportion of thirteen-year-olds who say they rarely read for fun, the shortening of sentences in bestselling fiction, the shrinking number of full books assigned in secondary-school English classes, the brain-imaging research comparing story listening to animated viewing. The claim is that the article's rhetorical power depends on treating all of these as evidence about one continuous, declining phenomenon, when they are in fact evidence about several distinguishable phenomena that happen to share a name.
Trace the drift across the piece. It opens with the deepest and most concrete sense of the word: whether Americans, in raw behavioural terms, sit down with a printed or digital book at all. It moves quickly to attention span, citing the difficulty of sustained engagement with long texts against the pull of short-form video. It moves from attention span to a claim about thinking itself, framed through Kwame Anthony Appiah's warning that handing the labour of reading and writing to a machine would make us a different kind of creature. It moves to education, through survey data on how few books secondary-school teachers now assign. It moves to memory, through the Socratic worry, voiced in Plato's Phaedrus, that writing itself was once feared as a threat to memory, and the observation that sustained oral memorisation of epic poetry is no longer a living practice. It gestures toward empathy and social cognition, through the wider discourse the piece situates itself within about literary fiction's capacity to simulate other minds. It touches democracy, through the broader claim that a common culture of long-form reading underwrites shared civic reasoning. It returns to cognition, through the brain-imaging comparison suggesting that children draw less on imagination-related neural systems when watching an animated version of a story than when hearing it read aloud. It moves to book ownership and library use, through reporting on which titles now circulate most at a major public library. It moves to time allocation, through the striking comparison of reading against gambling as competing leisure activities. It moves to literary participation, through the observation that today's bestselling fiction uses substantially shorter sentences than the bestselling fiction of the mid-twentieth century. And it moves, finally, to academic success, through the testimony of a Harvard humanities administrator describing students who regard assigned reading as an unnecessarily inefficient way of acquiring information that a professor could simply state directly.
Each of these individual moves is defensible in isolation. The problem is that the piece treats them as steps in a single argument about a single declining capacity, rather than as evidence about a dozen loosely related capacities that have each moved, to different degrees and for different reasons, in a broadly similar direction. Nobody reading the article pauses at any individual transition to ask whether the sentence-length data from bestselling fiction and the brain-imaging data from children watching cartoons are actually measuring the same thing. The prose carries the reader smoothly across the seam, and the seam is where the entire argument's persuasive force is manufactured.
This is not a criticism unique to Horowitch. It is a structural feature of the genre she is writing in, and, more importantly, of the underlying scientific literature she is faithfully summarising. The psychological and educational research on reading has itself never adequately distinguished these senses, which is precisely why a skilled journalist synthesising that literature produces an article that moves through all of them without friction. The article is not creating the confusion. It is transmitting it, clearly and at scale, which is exactly what makes it valuable as a diagnostic instrument. It shows, more clearly than any single academic paper could, how thoroughly the underlying category has come apart without anyone noticing. The same faithfulness to an unindividuated underlying category will reappear, in mirrored form, when the discussion turns to writing, and to what happens when a machine can produce the nonrecursive residue of either practice on demand.
Part III. Reading and Writing Are Not One Mediation
The first constructive move is to show what the Atlantic article's drift was tracking without naming it, and to show the same thing for writing at the same time: neither is a single cognitive operation housed in one mediational domain. Each is a coordination that draws on all five simultaneously, in every instance, whether the reader is a child sounding out a picture book, the scholar working through a monograph, the writer drafting a difficult paragraph, or the student typing an essay against a deadline.
Embodiment is present in both practices in ways psychometric research routinely brackets out. Reading is something eyes do, saccading, fixating, regressing across a line. Writing is something hands do, forming letters, striking keys, the specific fatigue of a long drafting session. Both are something the whole body does. The restlessness that interrupts a long reading session has its mirror in the restlessness that interrupts a long writing session, and the breath that slows during absorbed reading has its mirror in the breath that changes during a difficult sentence that will not resolve. A child's embodied history of reading, comfort or punishment, a parent's lap or a remedial classroom, shapes what reading is as a bodily practice. A child's embodied history of writing, praise for neat handwriting or shame at a red-inked page, shapes what writing is as a bodily practice, long before any comprehension question is asked or any essay is graded.
Being-with saturates both practices even when each looks like the most solitary of activities. Every reader reads with an author, an imagined presence whose intentions and reliability are continuously modelled. Every writer writes for a reader, an imagined presence whose expectations and patience are continuously modelled. Every reader who learned to read, and every writer who learned to write, did so within a specific relational configuration: a parent who read aloud or did not, a teacher who made reading or writing a site of shame or of pleasure, a peer culture in which either was admired or was social death. When Margaret Rennix describes Harvard students who experience assigned reading as an arbitrary obstacle between them and information a professor could simply state, she is describing a being-with failure. The mirror failure on the writing side is the student who experiences an assigned essay as an arbitrary obstacle between them and a grade a rubric could simply assign, and who therefore sees no relational reason to invest the selfrecursive labour composition actually requires.
Dwelling, in the strict LVT sense of non-human-made environmental conditions, plays a smaller but real role for both: the seasonal and daily rhythms that make sustained reading or sustained writing possible, the quiet of a particular season historically associated with retreat and reading, or with the concentrated solitude a difficult piece of writing requires.
Multimateriality is where the most visible recent transformation has occurred on both sides. The physical book, the printing press, the public library, the e-reader, and now the large language model condition what reading is possible. The pen, the typewriter, the word processor, the predictive keyboard, and now the same language model condition what writing is possible. The finding that today's bestselling fiction uses shorter sentences than the bestselling fiction of 1958 is a finding about the multimaterial and economic conditions of publishing as much as about readers. The parallel finding, if a research programme bothered to establish it, that professionally drafted prose today passes through more revision cycles per hour than prose drafted on a typewriter, would be a finding about the multimaterial conditions of composition as much as about writers.
Multisymbolisation is, unsurprisingly, central to both, and it is the mediation existing research on each already attends to, though usually without distinguishing its internal complexity. Reading involves the symbolic system of language, genre convention, and argumentative structure. Writing involves the same symbolic system deployed generatively rather than receptively, the writer's task being to construct the very structures the reader's task is to follow. A comprehension-test passage engineered to have one extractable correct answer is a different symbolic object from a three-hundred-page argument. A five-paragraph school essay engineered to satisfy a fixed rubric is a different symbolic object from a genuinely difficult piece of original argument. Treating performance on the constrained version as evidence about capacity for the open version, on either side, conflates two different symbolic architectures under one measurement.
No actual instance of reading, and no actual instance of writing, occurs in only one of these five mediations. A single reader or a single writer, on a single afternoon, moves through embodied fatigue, relational trust or distrust, the dwelling conditions of the room and season, the multimaterial affordances of the object or screen at hand, and the multisymbolic architecture of the specific text, all at once, continuously modulating each other. The instruments that dominate the public discourse about both, comprehension tests and book-completion counts on one side, rubric scores and word counts on the other, are calibrated almost entirely to register outputs of the multisymbolic mediation under controlled multimaterial conditions. They are structurally blind to the other four mediations even while those four mediations are doing much of the causal work in determining whether the multisymbolic outputs occur at all.
Part IV. Reading and Writing Occupy Every Recursivity Level
The mediational analysis already implies a recursivity-level analysis, and here too reading and writing move together, level for level, though the underlying research communities studying each have rarely noticed the parallel.
At L1, both occur as seamless, unreflective coordination. The fluent adult reader does not experience decoding as effortful. The fluent adult writer, drafting familiar material in a familiar genre, does not experience sentence construction as effortful either. Word recognition and syntactic parsing proceed automatically on the reading side. Lexical retrieval and basic grammatical assembly proceed automatically on the writing side. This is the level at which the most successful reading and the most successful writing happen, and, by the same logic developed in Unearned Confidence for embodied expertise generally, it is the level that generates the least measurable signal on either side.
At L2, something disrupts that seamlessness. On the reading side: confusion at an unfamiliar word, the felt friction of a passage that does not parse easily. On the writing side: the specific resistance of a sentence that will not come, the sense that a paragraph is not saying what it should. Much of what current discourse calls attention span, on the reading side, and writer's block, on the writing side, is best understood as a claim about tolerance for extended time at L2, for sitting with difficulty without reaching for an easier alternative or abandoning the attempt, rather than a claim about L1 fluency, which need not have declined in the same way.
At L3, both become explicitly articulable. The reader can name what confused them, ask a question, annotate a passage. The writer can name what is not working, articulate why a draft fails, explain a revision to a peer reviewer. This is the level at which classroom discussion, writing workshops, and most assessment instruments actually operate, since it is the first level at which either activity produces a verbal report an instrument can register.
At L4, both are stabilised into institutional and disciplinary forms. Reading is stabilised into curricula, canons, and the category of assigned reading. Writing is stabilised into the five-paragraph essay, the standardised rubric, the genre conventions of the academic paper or the cover letter. Much of the Atlantic article's evidence about reading, and much of parallel commentary about student writing, operates at this level. It is evidence about institutional stabilisation, not directly about individual cognitive capacity, though the two are of course related.
At L5, both become objects of theoretical reflection: literary theory and media theory on the reading side, composition studies and rhetorical theory on the writing side, and, in both cases, the long tradition of anxiety about decline that runs from Socrates' worry in the Phaedrus about what writing would do to memory, through worries about the typewriter and the word processor, to the present debate about generative AI. This article, like the Atlantic piece it analyses, operates at L5 with respect to both.
The consistent error across the contemporary discourse is to isolate one level on either side, most often L3 or L4, and generalise findings from it to the whole. A survey asking how many books a teenager completed, or how many essays a student drafted without assistance, is an L4 institutional and behavioural measure. A study of sentence-formation fluency under time pressure is closer to an L1 or L2 measure. Treating these as convergent evidence for a single underlying decline, on either side of the pair, silently assumes that L1 seamless engagement, L2 tolerance for difficulty, L3 articulable response, and L4 institutional participation all move together, in lockstep, as expressions of one variable. There is no reason, given the multimediational analysis above, to expect this, and considerable reason to expect the levels to move somewhat independently on both sides, since each is differently exposed to embodied, relational, material, and symbolic pressures that need not coincide.
Part V. nonrecursive, selfrecursive, Interrecursive: Why Psychology Cannot Know What Reading and Writing Are
This is the article's central methodological claim, and it is here that the sharpened thesis stated in the introduction does its main work: psychology cannot know what reading and writing are because its experimental methodology requires converting each into a nonrecursive object before either becomes measurable, and the conversion discards exactly the dimensions in which their value actually resides.
The nonrecursive dimension of reading is real and is what current instruments are best built to detect: words per minute, books completed, comprehension-test scores on standardised passages, national assessment results. The nonrecursive dimension of writing is equally real and equally well instrumented: words per minute typed, essay length, rubric scores on standardised prompts, grammatical error counts. Both dimensions are legitimate, well-defined, nonrecursive measures in the specific sense that the measurement does not depend on a relational loop between reader and researcher, writer and grader, or either party and the institution administering the test. A word count, on either side, is in principle indifferent to what the person being measured believes about the test. This dimension is exactly what the Atlantic article's statistical spine, and the parallel statistical spine of concern about student writing, are built from. All of it is genuine and worth taking seriously. None of it is reading or writing in the fuller sense either discourse's rhetoric claims to be diagnosing.
The selfrecursive dimension is where a person's relationship to their own reading or writing state shapes the state itself. On the reading side: am I actually following this, should I reread that paragraph, why does this feel harder than it should. On the writing side: is this sentence actually saying what I mean, should I cut this paragraph, why does this argument feel weaker than it did an hour ago. In both cases the monitoring changes what is being monitored, which is precisely why it is constitutively different from a nonrecursive word count. A reader who notices their attention drifting and consciously reorients produces a different reading experience than one who does not notice at all. A writer who notices an argument going slack and revises accordingly produces a different piece of writing than one who does not, even where both would register similarly on a comprehension test or a rubric administered to the finished product. Current instruments capture almost none of this on either side, because the monitoring leaves no trace in the nonrecursive output an instrument can see. A finished essay that passed through fifteen rounds of selfrecursive revision and a finished essay that emerged in one uninterrupted draft can be, on the page, indistinguishable, exactly as a comprehension-test score cannot distinguish the reader who reread three difficult paragraphs from the reader who skimmed past them and guessed correctly.
The interrecursive dimension is the one the current discourse almost entirely misses on both sides, and it is where the diagnosis becomes most consequential, and where the methodological argument can now be stated precisely. To design a controlled experiment on reading comprehension, or on writing quality, a researcher must hold the interrecursive variables constant: the same passage, or the same prompt, is given to every participant, with the same stated purpose, the same audience fiction, the same institutional stakes, regardless of what any individual participant's actual relationship to a real author, a real reader, or a real institutional purpose would have been. This is not a design flaw the researcher could correct with more careful instructions. It is what a controlled experiment is. Controlling a variable and eliminating its interrecursive content are, for exactly the dimensions that make reading and writing interrecursive, the same operation. The experiment therefore does not measure interrecursive reading or interrecursive writing more crudely than it might otherwise. It cannot measure them at all, because the object it has built, a text with a fixed, institutionally imposed relevance structure applied uniformly to every participant, is precisely a nonrecursive surrogate for the interrecursive achievement it is presented as measuring.
Margaret Rennix's Harvard students, who experience assigned reading as an inefficient way of being handed information a professor could simply state, are not, on this account, primarily failing at decoding, comprehension, or even sustained attention in the abstract. They are making an interrecursive judgment about the value exchange the institution is offering them, whether the extra cognitive labour reading requires is repaid by something a summary or a lecture could not provide. The same students, asked to write an essay whose interrecursive purpose is legible to them, for a reader they trust, in service of an argument they actually want to make, write differently than they do for a rubric whose only real reader is a grader working through a stack of forty near-identical submissions. No controlled writing experiment can register this difference, because registering it would require abandoning exactly the standardisation that makes the experiment controlled.
This reframing does real explanatory work that the nonrecursive statistics alone cannot do. It explains why interventions aimed purely at the nonrecursive level, more assigned reading, stricter word counts, more explicit rubrics, tend to produce disappointing results on both sides, because they do not touch the interrecursive judgment about whether the effort is worth making, and because a more explicit rubric, in particular, can actively confirm to a student that the institution itself now treats writing as a checklist rather than an act of communication addressed to a real reader. It explains why the same student can be a voracious reader of fan fiction and a prolific writer of fan commentary, nonrecursively enormous quantities of text on both sides, while reporting an inability to get through an assigned novel or produce a passable assigned essay. The interrecursive relationship with the fan community is trusted, reciprocal, and chosen, while the interrecursive relationship with the assigned syllabus may not be. And it exposes precisely the pattern identified in the marshmallow test: a phenomenon that is substantially interrecursive is measured with instruments built for a nonrecursive phenomenon, and the resulting gap is then narrated as an individual or generational failure of will, attention, or capacity, when it is at least as plausibly a change in the interrecursive terms on which the labour of reading and writing is being offered and evaluated. Psychology did not overlook this dimension by accident. It could not have registered it without ceasing to run the kind of experiment that makes it psychology in its current methodological form.
Part VI. Generative AI as Epistemological Gift
The Atlantic piece's engagement with generative AI, and the broader anxiety it reflects, is usually framed as a threat, to reading, to writing, to the cognitive capacities each was thought to cultivate. This article wants to reframe it. Generative AI is also, and perhaps primarily, an epistemological gift. It has done, in a few years, what several decades of psychological research failed to do: it has made the difference between symbol production and recursive discernment impossible to ignore, on both the reading and the writing side of the pair.
Consider writing first, since the thought experiment is sharpest there. Suppose a person dictates a rough set of points to a language model, receives a full drafted paper in return, rejects eighty percent of it, reorders what remains, and rewrites every substantive argument by hand. Has that person written the paper. Almost everyone, on reflection, says yes, and says it without hesitation. But notice what that answer immediately demonstrates. If the person has written the paper despite having typed almost none of the words that survive into the final draft, then typing was never writing. The thing that makes a piece of writing the writer's own was never the graphomotor or typographic act of producing the string of characters. It was the selfrecursive discernment involved in judging which eighty percent to reject, the interrecursive judgment about what this particular argument, for this particular reader, actually required, and the labour of rebuilding the argument once the discernment had done its work. The language model performed the nonrecursive task, string production, at high speed. It could not perform, and was never asked to perform, the selfrecursive and interrecursive task that made the result a piece of that person's writing.
The same thought experiment works, mirrored, on the reading side. Suppose a person asks a language model to produce three different summaries of a difficult book, compares them against each other, reads several passages of the original to check a claim that seems doubtful, and revises their own understanding, and their notes, in light of the discrepancies this comparison exposes. Has that person read the book. The ordinary category begins to strain in exactly the way it did for writing. The person has not read the book in the narrow, nonrecursive sense the word usually implies, cover to cover, every sentence. But they have done something that looks very much like the interrecursive and selfrecursive core of reading: forming a judgment about what the text says, checking that judgment against evidence, revising it under pressure from a discrepancy. What they have not done is the specific multimaterial and nonrecursive labour of moving their eyes across every page in sequence. AI exposes, on this side too, that the multimaterial labour and the recursive achievement were never the same thing, however tightly coupled they used to be.
This is not incidental to summarisation specifically, though summarisation makes it unusually vivid. A summary was never a smaller version of a text preserving its content in miniature, the way a thumbnail preserves a photograph. It is an interrecursive judgment, relevant to whom, for what purpose, under what constraints, and a good summary for an examiner checking comprehension can be a poor summary for a colleague deciding whether to read the source in full. Psychological experiments that ask participants to summarise the text, full stop, with no audience and no downstream purpose specified, were never testing summarisation. They were testing a manufactured nonrecursive surrogate, compress this string into a shorter string, and reporting performance on the surrogate as though it settled a question about the interrecursive achievement it replaced. A language model is very good at exactly this surrogate task, which is precisely why AI summarisation can feel like an adequate substitute for reading: it succeeds completely at the task the literature had already been measuring, because that task was never the full achievement in the first place.
One reason these confusions remained largely invisible throughout the twentieth century is that the multimaterial implementation and the recursive labour were tightly coupled. If one wanted to produce a sophisticated piece of writing, one generally had to perform the recursive work oneself because there was no alternative route to fluent symbolic production. Generative AI has broken that historical coupling. It has therefore become possible, for the first time, to see clearly that symbol production and recursive discernment were never identical activities. The technology has not created a philosophical problem. It has exposed one that was already there.
This reframing has a direct consequence for the assessment instruments discussed in Part V. A rubric-scored essay and a comprehension-test score were never actually measuring selfrecursive or interrecursive achievement, even before generative AI existed. They were measuring the nonrecursive residue that this achievement, in a tightly coupled world, reliably left behind: correct grammar and on-topic content on the writing side, a correctly bubbled answer on the reading side. Generative AI has not made these instruments newly gameable. It has revealed that they were always measuring the residue rather than the achievement, and that the residue can now be produced, cheaply and at scale, by a system with no selfrecursive or interrecursive relationship to the material whatsoever. What AI genuinely cannot do, or can only do on explicit instruction, is the interrecursive work itself: deciding, for this reader, in this relationship, for this future purpose, what matters and what can safely be omitted, or noticing that a claim on page forty qualifies a claim on page two hundred in a way that changes what the passage means for the specific argument a reader is trying to build. Being changed, in the sense Kwame Anthony Appiah gestures toward in the Atlantic piece, by the duration and difficulty of sustained engagement, is not a service a model can perform on a person's behalf, because it is not a service at all. It is the thing itself, and it happens, if it happens, only in the reader or the writer, not in the tool either one is using.
Part VII. Generational Decline or Category Failure?
The question that organises the contemporary public discourse, exemplified vividly by the Atlantic piece on the reading side and by a parallel, less consolidated discourse about employers who report graduates unable to write a coherent paragraph on the writing side, is whether the current generation of young people can still read, or still write, in some unitary sense each question presupposes exists. The multimediational and multirecursive analysis developed above suggests a different and more tractable question: which specific coordinative capacities, at which specific mediations and recursivity levels, have changed, in which direction, and why, on each side of the pair.
Some changes are real and are not in dispute. nonrecursive measures, minutes spent reading, books completed, standardised comprehension scores under timed conditions, essay word counts, grammatical accuracy rates, have genuinely moved by the metrics both discourses cite, and there is no reason internal to this article's argument to doubt those numbers. What is in dispute is what those numbers are measures of, and what has replaced the coordinative work they used to index.
There is reasonable evidence for at least the following changes, each belonging to a different mediation or recursivity level rather than to one collapsing capacity, on both sides. There has plausibly been a decline in sustained L1 to L2 coordination with long, single, linear print texts specifically, and a parallel decline in sustained L1 to L2 coordination with long, single, linear pieces of original composition, the capacity and inclination to remain with one extended argument or narrative, as reader or as writer, across many uninterrupted hours or days. There has plausibly been an increase in a different kind of coordinative skill on both sides: rapid interrecursive relevance extraction across many short, competing symbolic sources when reading, and rapid interrecursive audience calibration across many short, competing symbolic formats when writing, a skill barely exercised by earlier generations navigating a comparatively low-choice media environment. There has plausibly been a shift in the relational, being-with structure of both practices, away from a model in which an institutional authority's assignment was largely sufficient to justify the labour involved, toward a model in which that justification is actively and continuously evaluated against readily available alternatives, some of which do provide comparable value for a fraction of the cost. And there has plausibly been a change in symbolic switching capacity on both sides, the fluency with which attention moves between text, image, video, and audio, a genuine multisymbolic skill, differently distributed and differently valued than the older skills of sustained linear reading and sustained linear composition, and not obviously inferior to them as a general cognitive matter, whatever its costs for those specific practices.
None of this licenses complacency, and this article does not argue that nothing of value is being lost on either side. The specific coordinative achievement of sustained, single-threaded engagement with a long, demanding, linear argument or narrative, whether as reader or as writer, does appear to be declining, for reasons the multimediational analysis makes legible. It requires embodied tolerance for extended L2 friction, relational trust in the text's or the institution's claim on one's time, and multimaterial conditions of uninterrupted attention that a smartphone-saturated environment actively works against, several at once, and a decline in any one of these mediational supports is sufficient to produce the behavioural pattern the statistics register on either side. Kwame Anthony Appiah's worry that we would become a different kind of creature if this specific capacity atrophies broadly is worth taking seriously on its own terms, for readers and for writers alike.
What the multimediational reframing changes is not the conclusion that something is happening, but the explanatory shape of the claim, and therefore the kind of response it calls for. Current psychology, inheriting categories built for a print-centred age in which the various senses of reading, and the various senses of writing, moved together closely enough that the collapse into one word did relatively little damage on either side, cannot detect that they have now come apart, because its instruments were never built to distinguish them. A discipline that measures only the nonrecursive residue of a multimediated, multirecursive achievement will report a single number and call it a decline in reading, or a decline in writing, when what has actually occurred is a redistribution of coordinative effort across a landscape of technologies, mediations, and recursivity types that neither word was ever precise enough to track in the first place.
Conclusion. Ordinary-Language Achievements, Not Natural Kinds
Reading and writing are not failed concepts. They are ordinary-language achievements, and they work extraordinarily well for the purpose ordinary language was built for: coordinating the mesocosm of everyday life. A parent asking whether a child has done their reading, a teacher assigning an essay, a friend recommending a novel, all rely on reading and writing doing exactly the compressive work these words have always done, and nothing in this article is a proposal that they stop. The mistake begins only at the point where psychology, education research, and the public discourse that draws on both, borrow this same everyday compression and present it as though it had already identified a natural scientific object with a single measurable magnitude that can rise or fall. It has not. Like class, health, technology, intelligence, and attention, reading and writing are compressed L4 stabilisations whose apparent unity conceals radically different mediational and recursive architectures, five mediations, five recursivity levels, three recursivity types, each moving somewhat independently of the others, on each side of the pair.
This is not a claim that nothing has changed, or that concern is misplaced. Something has changed, in specific, nameable ways this article has tried to specify precisely: a decline in sustained L1 to L2 engagement with long linear texts, on both the reading and the writing side, a shift in the interrecursive terms on which institutions ask for the labour of either, a rearrangement of which mediations carry the coordinative weight that print and the pen once carried more or less alone. What has not changed, because it was never true in the first place, is the assumption underneath a century of psychological research: that reading and writing name single, natural, nonrecursive capacities that a controlled experiment can isolate and a policy intervention can directly repair. Psychology could not have discovered this assumption was false while continuing to build experiments that require the assumption to be true in order to run at all. That is the specific, methodological sense in which this article has argued that psychology cannot know what reading and writing are, not merely that it has not yet found out.
The future science of reading and writing, if such a thing is to be built, will therefore not begin by measuring either more precisely. It will begin by replacing the inherited categories with coordinative profiles that specify exactly which mediations, which recursivity levels, which recursivity types, and which forms of interrecursive relevance are actually under investigation in any given study, curriculum, or policy claim. And reading and writing, on the argument of this article, are not exceptional. Memory, thinking, attention, creativity, learning, problem solving: the same reconstruction is owed to each of them, and the case developed here for two of the human sciences' most consequential ordinary-language words is offered as a template, not a conclusion, for the wider project this series has been pursuing across each of its cases, identifying, one overburdened category at a time, where the human sciences mistook a word for a discovery.