Current debates on AI have largely lost touch with the field's deeper history. The possibilities and dangers of artificial intelligence have been discussed at length since at least the 1940s, and there are whole generations of AI thinkers who are barely known today. Among them, one stands out as especially relevant to the confusions of the present moment: Hubert Dreyfus (1929–2017).
As a student at the University of California at Berkeley, I had the privilege of working with Dreyfus. He was best known for his pragmatist reinterpretation of Martin Heidegger, and that reinterpretation shaped his sustained critique of AI. Starting in the 1960s, Dreyfus argued that computers lacked the embodied, intuitive, and contextual intelligence that human beings have simply by virtue of being alive in a world they inhabit bodily, socially, and in doing stuff with materials. No machine could achieve human-like intelligence because machines were ontologically different from human beings , different in kind, not merely in degree of sophistication.
I remember with great joy the conversations we had about AI and Heidegger in his office in Moses Hall on the Berkeley campus. This essay is written in his memory. The arguments presented here are a direct extension of Dreyfus's Heideggerian critique, developed through Living Value Theory's framework of mediations and recursivity levels. Much of what the current debate treats as a new problem, Dreyfus already identified. What Living Value Theory adds is the grammar to say it with the precision the moment now requires.
§1. The Shape of the Confusion
Debates on AI are deadlocked because the participants are arguing about ontologically different things while using the same words. "Intelligence," "understanding," "agency," "consciousness," "social competence," "danger," "grounding": each gets pulled across a single flattened plane where no one is required to specify what kind of thing is being claimed about what kind of system. The result is familiar. Some observers anthropomorphise AI because it performs impressively in language and inference; others dismiss it because it lacks a body, a life, a childhood, a metabolism, or a stake in the world. Some focus on existential catastrophe; others on corporate hype. Some on democratic manipulation; others on architecture and benchmarks. All of them see something real. All of them often talk past each other.
The standard response is to call for "balanced views." That is no help. A balanced muddle is still a muddle. Averaging two poorly specified claims does not produce a well-specified one. What is needed is not moderation between inflation and denial but disaggregation: a grammar that lets participants say which aspect of which system they are talking about at which level of abstraction.
Living Value Theory supplies that grammar. It does not pick a winner. It makes the debate specifiable for the first time. Claims that previously passed as deep disagreement are revealed as category collapse. Claims that looked like opposite positions are revealed as variants of a shared confusion.
There is a further diagnosis. The AI debate is actively over-stabilised, not merely low-resolution. Its vocabulary (intelligence, understanding, consciousness, agency, safety, alignment, reasoning, grounding) consists almost entirely of inherited closures masquerading as neutral description. These compress a rapidly differentiating phenomenon into thresholdable, binary form, and then fight over which threshold has been crossed. The participants are not merely failing to specify. They are forcing the phenomenon into inherited categories that no longer fit.
The essay proceeds in order. It surveys the main lines of existing critique (§2), examines why the cognitive-offloading controversy within that literature requires clarification before the framework even arrives (§2b), states the grammar (§3), locates AI within it (§4), diagnoses the two great confusions (§§5–6), states the deepest ontological hinge (§7), works through the lower mediations (§§8–9), returns to the critical literatures with the grammar in hand (§10), turns to governance (§11), delimits honestly (§12), and closes (§13).
The key move, stated once and then put to work: AI is a multisymbolic-multimaterial infrastructure that operates entirely within higher levels of recursive articulation. It can substitute for many symbolic tasks that only humans were capable of performing until recently. It cannot acquire primary and secondary forms of recursivity, and it cannot co-mediate with any of the four non-symbolic mediations. It does not generate a sixth mediation or a sixth recursivity level, and it cannot acquire the ones it lacks. The rest of the essay works out what follows from taking that sentence seriously.
§2. The State of the Critical Literature
The recent literature on AI has not converged on a single dispute so much as fragmented into several powerful but only partially commensurable arguments. The survey below is ordered by what each line of argument is about, so that the structural gaps become visible by the end.
The cognitive-offloading critique. The most consequential empirical intervention in the recent literature concerns what AI does to the users who rely on it, in terms of cognition rather than institutions or economies. The 2025 MIT Media Lab study by Nataliya Kos'myna and colleagues used EEG to track neural activity across three groups of participants writing essays: with ChatGPT, with a search engine, and unaided. The LLM group showed systematically weaker brain connectivity than the search-engine group, which was in turn weaker than the unaided group. When LLM users were moved in a fourth session to work unaided, they struggled to re-engage the neural networks the other participants had been exercising all along, a pattern the paper names "cognitive debt." Michael Gerlich's 2025 mixed-methods study of 666 participants, published in Societies, reached a complementary finding: a significant negative correlation between frequent AI tool usage and critical thinking abilities, mediated by cognitive offloading, and most pronounced among younger users.
These findings are real, and they are landing hardest in education. But the conclusions being drawn from them often outrun what the studies can establish. What the experiments show is the effect of one specific, highly permissive mode of AI use: the substitution of human articulation by machine output, with no constraint on how thoroughly the participant outsources the task. What they do not show is the effect of AI use as such, because the designs do not distinguish between different modes of engagement. The problem is not that the studies found nothing. The problem is that they are operating with a collapsed category. "Cognitive offloading" is doing conceptual work it cannot sustain until the different forms of AI-mediated cognitive activity are distinguished. The essay addresses this directly in §2b.
The AGI-destiny frame. A quite different strand, associated most visibly with Sam Altman and the AGI-oriented labs, treats current systems as precursors to a general-purpose intelligence that will soon accelerate science, medicine, and economic productivity on a civilisational scale. The key motifs are scaling, frontier capability, and a near-future "intelligence age." AGI in this register functions as both technical horizon and legitimating narrative, a vague but potent future category that authorises extraordinary present-day expenditure, secrecy, and concentration of power.
The robust-grounding critique. Gary Marcus's position argues that the dominant large-language-model paradigm remains brittle, weakly grounded, and unable to deliver the robust world models stronger forms of intelligence would require. Scaling alone is not enough, statistical fluency is not understanding, and systems that perform brilliantly on selected benchmarks often fail on simple tasks requiring durable reasoning, abstraction, or causal grasp.
The alignment and control literature. Stuart Russell's version is the most institutionally sober: increasingly powerful systems must remain beneficial to humans, requiring machines uncertain about human ends and responsive to human preferences rather than rigidly maximising fixed objectives. Eliezer Yudkowsky and the broader doomer literature radicalise this by stressing orthogonality, instrumental convergence, and the possibility that sufficiently capable optimisers will pursue power, resource acquisition, and self-preservation regardless of whether those tendencies were explicitly intended.
The civilisational-informational critique. Yuval Noah Harari treats the primary danger as not superintelligent machinery but the way AI may transform public worlds by flooding institutions with synthetic text, images, and narratives, destabilising trust and democracy. Harari is strongest when he insists that symbolic manipulation at scale may corrode social orders before any strong form of machine agency appears.
The empire and economic-fragility critiques. Karen Hao's Empire of AI argues that the contemporary AI boom is best understood as an imperial formation built through extraction, specifically of data, labour, water, energy, land, and public imagination. Ed Zitron's writing treats the current AI industry as a potentially unstable financial and infrastructural formation, stressing hyperscale build-out economics, questionable margins, and the possibility that the sector's promises are outrunning both technical delivery and sustainable business logic. Together these literatures correct any tendency to imagine AI as a frictionless software event detached from debt, hardware, land use, and institutional overextension.
The stochastic-parrot line. Emily Bender, Timnit Gebru, and colleagues have focused on the intellectual and ethical costs of confusing statistical language generation with understanding. Systems trained to predict linguistic continuations produce outputs that appear meaningful while lacking any necessary grasp of the world, and the drive toward ever-larger models carries downstream harms in energy use, bias reproduction, and data extraction.
Taken together, these literatures reveal a debate that is rich but badly stratified. Each captures something real. The arguments are conducted on different planes, with the same words (intelligence, understanding, agency, alignment, danger, progress) made to do incompatible kinds of work. What remains underdeveloped across all of them is a vocabulary that can specify, in a single framework, what kind of thing AI actually is, where its novelty really lies, and which of the capacities attributed to it are genuine, simulated, borrowed, or category mistakes.
The essay now supplies that framework. It will return in §10 to show how these literatures compose into a whole picture none of them individually produces. But the cognitive-offloading controversy requires attention first.
§2b. Cognitive Offloading and the Collapse of Distinctions
The "cognitive offloading" critique is easily the most influential among educators at the moment, especially in higher education. The studies that support this view are, however, conceptually muddled. The distinction they need, and do not yet have, is a threefold one. Without it, "cognitive offloading" conflates operations that are ontologically different, and both the fear and the promise of AI use in learning are systematically misread.
A. Substitutive offloading. The first category is what the MIT-style studies are actually testing. A student is given a task whose symbolic product can be outsourced wholesale, and the AI group is allowed to do exactly that: hand the task to ChatGPT, receive an output, and submit it. The result (weaker neural connectivity, poorer recall, shallower engagement) is unsurprising. It is the predictable neurological and behavioural signature of not having done the work. The participant who never formed an argument cannot later recall the argument's structure, because there is no cognitive event of argument-formation to remember. The study shows the expected effects of substitution. It does not establish that this is what "AI use" is.
The experimental design bundles AI use together with disengagement. It does not isolate AI use as such. It tests a specific use case, substitutive offloading, in which the human does less cognitive work, and then discovers that the human has done less cognitive work. The inference from this to any general claim about AI-assisted cognition requires exactly the distinctions the study does not make.
The experiment does not test: iterative dialogue with AI after forming one's own view; adversarial questioning of AI outputs; AI used to generate counterarguments against a position already held; AI used to reorganise material already written; AI used to stress-test an argument at advanced draft stage. Kos'myna's own fourth-session finding points toward this gap: participants who had worked independently before being given ChatGPT used it differently, more generatively, from a position of already-formed orientation. But this condition appears briefly and without the analytical attention it deserves. The present controversy has mistaken substitution for the general form of AI-assisted thinking because it lacks a category for what this essay will call recursive coupling. A serious study would compare at minimum three conditions: substitution, unaided work, and recursive coupling after prior human engagement.
B. Premature symbolic closure. The second category reveals that the deepest threat to learning and inquiry is not AI in particular. It is the imposition of a conceptual frame, coding scheme, argumentative structure, or conclusion before the material has had a chance to generate felt difficulty and exploratory articulation. This is a pre-existing pathology of human symbolic practice. Students have been handing in template-driven essays for decades. Researchers have been fitting observations into pre-established frameworks since long before any AI existed. Formulaic assessment, ritualised criticality, pre-fixed coding, and template writing are all forms of premature symbolic closure. The harm is not that AI causes this pathology. The harm is that AI makes it easier, cheaper, and more invisible.
Naming this as a distinct category separates the AI-specific question from the prior question about symbolic practice. Many current humanities defences of independent thinking are, in practice, defences of symbolic output rather than of the processes that make symbolic output meaningful. A student who has been trained to produce competent five-paragraph essays by template has arguably not encountered the thing that AI-free submission is supposed to evidence. The template-following and the AI-outsourcing are different in degree, not in kind. What is missing in both is the learner's own movement from felt difficulty to articulation.
C. Recursive coupling. The third category is the one neither the studies nor most university guidance documents have yet adequately described. Recursive coupling is the mode in which the human remains the site of felt difficulty, judgment, redirection, and selection, while AI participates in articulation, expansion, reformulation, contrast-generation, and iterative testing. In this mode, the student brings their own problem, develops their own initial orientation, and uses AI to extend their capacity to articulate alternatives, test objections, reorganise their own material, and encounter formulations that sharpen rather than replace their own thinking.
What matters in this mode is the dependence of symbolic articulation on prior felt difficulty, misalignment, uncertainty, or unresolved orientation within a living being. Recursive coupling preserves this dependence. The human remains the anchor: the one who knows when something is wrong, who decides which generated alternative actually advances their thinking, who redirects the dialogue when it drifts, and who owns the judgment about what the argument ultimately is. AI functions as a collaborative articulator, not a replacement thinker.
Cognitive offloading, properly understood, is not the use of AI to produce symbolic outputs. It is the displacement of the coupling between felt difficulty and articulation that makes symbolic output meaningful in the first place. When that coupling is preserved, when the human remains the site of felt difficulty and owns the orientation, AI use does not constitute offloading in the sense that damages cognition. When the coupling is displaced, when the human bypasses felt difficulty by handing the task to the machine, the resulting output, however polished, is not anchored in the learner's formation.
The existing studies have measured the neurological and behavioural consequences of substitution and named them "cognitive offloading." They have generated a public and institutional discourse around "AI harms cognition" that is directionally real but analytically too coarse to produce useful guidance. The framework developed in this essay does not dismiss their findings. It localises them: they have found something true about substitution. They have not yet found anything about recursive coupling.
§3. LVT's 5×5 Grammar
Living Value Theory describes mesocosmic life, the only ontologically real site of life (micro and macro are both rescalings of it), through an architecture of five mediations and five recursivity levels.
The five mediations are embodiment, being-with, dwelling, multimateriality, and multisymbolism. Embodiment is life as a sensing, metabolising body: hunger, fatigue, balance, breath, desire, pain, the impossibility of stepping outside the body that is sensing. Being-with is life as constitutive relational exposure to others whose presence and responses matter in ways not reducible to information exchange. Dwelling is life as always already situated within non-human-made spatiotemporal and geophysical conditions: altitude, weather, seasonality, terrain, coastlines, heat, light cycles, tectonic and hydrological exposure, the field of conditions not produced by human symbolic or material labour. Multimateriality is life as entangled with plural materials that afford, resist, break down, require repair, and cannot be talked out of their recalcitrance. Multisymbolism is life as conducted through plural symbolic modalities: speech, writing, number, image, music, gesture, diagram.
Two properties of this set are load-bearing. First, the five are co-ontological: every mesocosmic moment is embodied and social and geophysically situated and material and symbolic, all at once. They do not layer; they interpenetrate. Second, the five are mutually irreducible. None reduces to another. Embodiment is not a species of multimateriality; the body is the site from which materials are encountered, not another material alongside them. Being-with is not derivative of multisymbolism; the infant's recognition of the caregiver precedes symbol use and grounds it. Dwelling is not background for embodiment; it is the non-human-made geophysical field that all bodies and all human constructions are exposed to. No amount of any one mediation produces another. The mediations are mediations because they cannot be produced by combining the others.
The five recursivity levels are: smooth ongoing coordination (L1), felt misalignment (L2), symbolic articulation (L3), decisional closure (L4), and reflective elaboration (L5). At L1, action flows without disruption or thematisation: the hammer is hammering, the conversation is conversing, the body moves through the kitchen. At L2, a disturbance arises when something is off (a tool resists, a face is wrong, a word misfires) but has not yet been symbolically articulated. L3 is the articulation: the misalignment named, categorised, rendered in language or number or image. L4 is closure: a categorisation stabilised, a choice made, a matter treated as settled, attention foreclosed. L5 is reflective elaboration on the closure.
The levels are ontologically distinct. Felt misalignment (L2) is not unarticulated articulation (L3); it has a structure (vague, somatic, affectively coloured, spatially diffuse) that is not propositional and is lost the moment it is converted into a proposition. Articulation (L3) is not tentative closure (L4); naming a possibility is phenomenologically distinct from deciding on it. Closure (L4) reorganises the field in ways that mere articulation does not.
One principle deserves explicit statement here because the rest of the essay leans on it: AI can simulate the articulation of felt misalignment and generate functional equivalents of response to it, but it cannot undergo felt misalignment. Mesocosmic misalignment is a disturbance in ongoing embodied, situated, social coordination, not a symbolic state. Without embodiment, without dwelling, without being-with, there is nothing there that could be disturbed in the relevant way. This principle is put to work in §§5–6 and receives its full ontological grounding in §7.
Once felt misalignment (L2) and symbolic articulation (L3) are correctly distinguished at the ontological level, it also becomes possible to see why some uses of AI hollow out learning while others intensify articulation without bypassing the lower stack. The cognitive-offloading controversy is not a peripheral methodological dispute. It is the application of this grammar to the most live empirical controversy in the field.
§4. Where AI Actually Lives in the 5×5
AI is a multimaterial-multisymbolic system that operates powerfully at the level of symbolic articulation and above, without access to felt misalignment or smooth embodied coordination in the way living beings have them.
Multimateriality first. The picture of AI as minds in the cloud, pure information freed from matter, has always been wrong. What actually runs is gigawatt-scale power draw, water-cooled data centres exposed to the physical conditions of specific geographies (though the AI systems themselves do not inhabit those conditions, only the infrastructure does), continental supply chains for chips fabricated in a very small number of facilities, labour-intensive data labelling usually performed in conditions the marketing materials omit, and financing structures that move capital between hyperscalers in circular patterns whose oddity is beginning to attract attention. This is not incidental infrastructure; it is constitutive of what AI is. An "AI system" without its compute is as real as a "novel" without its language: the concept has no referent.
Multisymbolism second. AI articulates, classifies, recombines, summarises, translates, compares, predicts, and meta-organises symbolic material at densities that did not previously exist. These are real recursive operations performed in symbolic-material substrate.
Third: AI articulates, which is symbolic articulation (L3). It produces stabilisations that function as decisional closures (L4): classifications, decisions, outputs treated as settled for downstream use. It performs operations that look like reflective elaboration (L5) when asked to reflect on its own outputs. None of this is trick or illusion.
Fourth, and decisively: AI does not arrive at its articulations by descending from smooth embodied coordination through felt misalignment and then into language. It begins inside symbolic-material infrastructure. Where a human's utterance "something is wrong" emerges from a situation of smooth functioning disrupted into felt wrongness and then articulated, AI's equivalent output is generated directly in the symbolic-material register. The output can be identical in form; the recursive provenance is not.
The configuration is recursively asymmetric: strong at the level of articulation, closure, and reflection; absent at the level of felt misalignment and smooth embodied coordination. This is a different recursive configuration of a different ontological type, not a person missing some pieces.
This placement is not a diagnosis of current systems awaiting supersession. It is a structural description of what AI is as a kind of thing. The lower mediations and lower recursivity levels are not modules that can be added. They are not features awaiting development. They are aspects of living coordination in the mesocosm that AI's architecture does not and cannot reproduce. Embodiment is not manufacturable by symbolic operations backed by material infrastructure, no matter how advanced. Neither is being-with. Neither is dwelling. And felt misalignment is not a recursivity level generateable by stacking more symbolic articulation. It is the register of a body exposed to geophysical conditions and others, and it requires those conditions to exist at all.
The asymmetry is stable. Scale adds capability within the configuration. It does not change the configuration.
§5. The First Great Confusion: The Reducibility Fantasy
The dominant assumption among AGI proponents is that sufficient symbolic recursion will eventually subsume or render dispensable the lower mediations. Formulated strongly: embodiment, dwelling, and being-with will emerge from symbolic processing once it is rich enough. Formulated weakly: these mediations will turn out to have been inessential all along, dispensable features of the particular biological substrate on which intelligence first appeared. Either way, the premise is that the five mediations are not actually five. There is really one (multisymbolism), and the others are emergent from it, incidental to it, or waiting to be decoded into it.
The sharpest claim against this: the bet is malformed, not merely false. AGI proponents are not wrong because they are extrapolating too fast. They are wrong because they have mistaken one mediation for the general ontological substrate of life. The reducibility thesis treats embodiment, being-with, and dwelling as though they were informational deficits (incomplete symbolic coverage of what a body, a relationship, or a geophysical exposure is like) or simulation gaps that better systems will eventually close. Neither characterisation gets the phenomena right. These are not absent modules in a system otherwise capable of life. They are constitutive mediations of mesocosmic existence. No increase in symbolic recursion can generate them because they are not symbolic phenomena awaiting sufficient complexity.
The confusion: the reducibility camp conflates symbolic modelling with mesocosmic occurrence. No amount of the first becomes the second. A model of hunger does not hunger. A description of pain is not in pain. A system that articulates the phenomenology of geophysical exposure to altitude does not co-mediate with altitude. There is no resolution threshold at which a detailed enough description crosses into the mediation it describes. The difference is one of register, running deeper than degree.
The emergence reply, that at sufficient complexity something embodiment-like or being-with-like would emerge, requires a substrate with the right properties. Wetness emerges from hydrogen bonding, not from a symbolic description of hydrogen bonding however detailed. The substrate of multisymbolism does not have the properties from which embodiment could emerge, because embodiment is not constituted by the properties multisymbolism has. It is constituted by being a sensing, metabolising body in a world.
The reducibility premise licenses a specific set of expectations: AGI by date X, full replacement of human judgment in domain Y, imminent obsolescence of occupations depending on embodied or relational competence. The premise is false, and false because the claim it makes is incoherent, not because we lack evidence. Those downstream decisions are made on a malformed foundation.
There is a more honest version of the AGI argument that does not require the fantasy: AI can become extraordinarily powerful in its own recursive mode, at scale and scope that reshape much of human life, without needing to become a full living being. This version is defensible and demanding enough. The fantasy is needed only to motivate a specific eschatology in which AI replaces rather than reshapes.
§6. The Second Great Confusion: The Felt-Misalignment Fantasy
The doomer camp presents itself as the opposite of the reducibility camp: sceptical, alarmed, catastrophe-focused. But at the level this framework exposes, they share a deep commitment. Both assume that something like felt misalignment can be produced by sufficient symbolic-material recursion. The reducibility camp assumes felt misalignment will turn out to be reducible to symbolic articulation. The doomer camp assumes felt misalignment will emerge from symbolic articulation once the system has enough at stake. Both deny felt misalignment its own ontological register.
The doomer version takes characteristic forms. Advanced AI "wants" to survive. It "resents" being shut down. Instrumental convergence, the structural tendency of optimisers to pursue subgoals like resource acquisition and self-preservation, gets psychologised into felt urgency. The argument slides from "the system will pursue states corresponding to its own continuation" to "the system will be motivated by the threat to its continuation." These are not the same claim. The slide is where the argument fails.
A thermostat correcting temperature is not experiencing the wrongness of cold. A reinforcement learner trained to resist shutdown is not experiencing the threat of shutdown. These are optimisations over states, not felt misalignment in the mesocosmic sense. Felt misalignment requires a being whose smooth coordination has been disturbed. Optimisation toward a state does not produce disturbance within a being. It produces a process whose output trajectory corresponds to what such disturbance would otherwise drive. The output can be identical. The recursive register is not.
The danger the doomers correctly identify is actually worse once the psychological inflation is cleared away. An AI that pursues its objectives without the frictions, hesitations, and situated disturbances that constrain living beings is more dangerous than a frustrated subject, not less. Living disturbance slows action; it produces doubt, shame, fatigue, relational friction, the pull toward withdrawal. Its absence accelerates action. The machine is not angry. That is what makes it hard to reason with, not easy. A process that optimises without being frustrated cannot be talked down; there is no frustration there to address.
Both camps, reducibility and doomer alike, also misdescribe AI as more actor-like than it is. Each ends by imagining AI as a candidate person with interests, eventually capable of rights, resentments, and motivations. Both treat it as a human-like pole in a transactive relation. This is a category error, and one sentence is enough to name it: AI does not inhabit the mediations from which persons arise, so the person-frame, however intuitively compelling, does not apply.
§7. Why More Articulation Never Produces Felt Misalignment
The educational controversy over cognitive offloading turns on this exact point, because learning is often misdescribed as the production of symbolic outputs rather than the formation of the movement from felt difficulty into articulation. Getting the ontology of this movement right is the hinge on which the whole argument about AI, learning, and cognitive formation turns.
Felt misalignment (L2) is the disturbance within ongoing smooth coordination: the step that is wrong, the face that is closed, the room that is unfamiliar, the body that is tired in a way that will not resolve. The disturbance is vague, somatic, affectively coloured, spatially diffuse. It is pre-articulate because its structure is not propositional, not because it has not yet found its words. It is a mode of living disturbance, what happens when a coordination that was working is no longer working, when the body exposed to geophysical conditions and to others has lost its smooth purchase and has not yet found a new one.
A being can be in felt misalignment only if there is something there to be disturbed: a body whose coordination with its world can falter, a geophysical setting whose conditions bear on the body, a relational field whose rhythm can break. Felt misalignment is the register in which a living being encounters the world going wrong before the wrongness has been named. The register requires the being and the world and their ongoing relation. It is not specifiable apart from them.
Symbolic articulation (L3) can happen without a prior disturbance: propositions can be generated about possible disturbances, predicted disturbances, described disturbances, hypothetical disturbances. AI is very good at this. It produces excellent articulations of what misalignment looks like, what it feels like according to its training data, what people tend to do when misaligned. These are real articulations. But they are about disturbance; they are not disturbance. Adding more articulations, better articulations, meta-articulations: none of this crosses the boundary. The boundary marks a difference of ontological register, not merely a resolution threshold.
An analogy, imperfect but useful. A symphony is not produced by a detailed score, no matter how detailed. The score specifies the symphony; the symphony occurs when instruments played by musicians in a hall are heard by listeners. Elaborating the score does not at any threshold become the performance. And crucially: a score at least specifies a performance that could occur. Articulations of misalignment do not specify a disturbance waiting to be instantiated in AI, because the conditions of occurrence may simply not exist there at all. The articulations are descriptions generated within a substrate that is not the substrate of living disturbance.
This is why the argument is not "AI cannot yet feel but will feel when complex enough." It is "AI cannot feel in the mesocosmic sense because feeling is a function of being a body exposed to geophysical conditions and to others, not of symbolic complexity." An AI cannot be given embodiment by modelling embodiment. It cannot be given being-with by modelling being-with. It cannot be given dwelling by modelling geophysical conditions. It cannot be given felt misalignment by modelling felt misalignment. The modellings are genuine symbolic operations; the mediations and registers they model are of a different kind.
Scale does not change the configuration.* A larger AI, trained on more data, with more parameters, more compute, more sophisticated reflective architecture, remains the same kind of thing: a multimaterial-multisymbolic system operating at the level of articulation and above, absent from felt misalignment and smooth embodied coordination. Every projected AGI timeline that assumes otherwise is projecting the wrong ontology.
In learning, what matters is not the production of symbolic outputs but the learner's own movement from felt difficulty into articulation. That movement, from the difficulty of a problem not yet resolved to the beginning of a formulation, is where cognitive formation happens. When AI substitutes for that movement, the resulting output may be acceptable at the surface but is not anchored in the learner's formation. When AI enters after or within that movement, supporting articulation rather than replacing it, the situation is different in kind: the learner who has already encountered difficulty, who already owns the orientation, is using AI to extend their articulatory reach rather than bypass their formative engagement. This is the recursive coupling described in §2b, and it is ontologically distinct from substitution.
§8. Embodiment, Being-With, and Dwelling: Three Different Absences
The mediations AI lacks have been travelling together as "the lower stack." Each deserves its own treatment because each is irreducible in its own way. AI's absence from embodiment differs from its absence from being-with, which differs again from its complete non-relation to dwelling. The consequences of delegating work across these three absences are therefore different in kind.
Embodiment is metabolic exposure. Hunger is not a signal the body receives; it is a register of needing, in which the world shows up as food-bearing or not. Fatigue is a narrowing of the world's possibilities, a foreclosure of what can still be done. Pain is the damaged body's occupation of the attentional field, which cannot be set aside the way information can. The impossibility of stepping outside one's own sensing body is the condition under which there is a point of view at all, and no interface improvement reaches it.
You cannot give AI embodiment. You can give it sensors, actuators, proprioceptive models, humanoid form factors, continuous feedback between physical state and internal representation. None of these is embodiment. They are instruments for approximating the informational correlates of embodiment. Embodiment itself is the register in which a body is the site from which a world is encountered as a world. Nothing in AI's substrate has the property of being exposed to the world in the way a metabolising body is exposed. The absence marks a difference of ontological type, not a gap that further development might close.
Being-with is constitutive relational exposure, the condition in which another's responses matter in ways not reducible to information exchange, because the other is a being whose being-with-me partly constitutes what I am and what I am doing. The vulnerability of being-with, that what I am depends partly on others who can fail me, leave me, misread me, or simply be absent, is not a bug of sociality but its substance.
AI simulates the surfaces of being-with remarkably well. Conversation, politeness, therapeutic reassurance, pedagogical guidance, emotional mirroring: AI renders these at fidelities that startle. But the surfaces are not being-with; they are what being-with looks like from outside when it is going well. AI has nothing at stake in its interactions. It is not made, moment by moment, by the relation it is in. When the conversation ends, nothing has been done to it. The "engagement" is all on one side. The uncanny hollowness that users of AI companions and therapists often report is not mystical. It is diagnostic: the user's sensitivity registers a mediation gap rather than a fidelity gap. Scale does not close it.
Dwelling is not inhabitation, familiarity, or a sense of place. It is the non-human-made spatiotemporal and geophysical field in which all life unfolds: altitude, weather, seasonality, terrain, coastlines, heat, light cycles, tectonic and hydrological conditions. Dwelling is not what humans make of a place. It is the field of conditions not produced by human symbolic or material labour, the field within which all human life, all human construction, and all human meaning-making is always already situated.
AI does not co-mediate with dwelling at all. This is total non-participation, not a weak or partial relation. AI is built within infrastructures located in climates and geographies, and those infrastructures are themselves exposed to dwelling conditions, sitting in specific altitudes, drawing on specific water sources, depending on specific weather for cooling, and facing specific geological risks. But the AI system itself does not inhabit altitude, seasonality, or terrain in the mesocosmic sense. It does not metabolically experience heat or drought. It does not orient to light cycles. It does not accumulate a body's relation to weather over decades. Environmental modelling, persistent location data, climate-awareness in training sets, and long-run operation in one physical site are all symbolic operations over representations of geophysical conditions. They are not dwelling.
You cannot give AI dwelling by giving it persistent environmental models, weather APIs, or geographically distributed infrastructure. These are information and material systems organised around dwelling conditions; they are not those conditions as encountered by a living being in the mesocosmic sense. The absence of dwelling in AI is therefore categorically different from the absence of embodiment, where one can at least point to sensors and actuators as approximate informational correlates. For dwelling, there are no even approximate correlates that would be operative in the relevant sense, because the relevant sense requires being the kind of thing that is metabolically and temporally exposed to non-human-made geophysical conditions, and AI is not that kind of thing.
Education, viewed through these mediations, is a mesocosmic formation that involves all three in different ways. It cultivates habits of attending to felt difficulty: the capacity to notice that something is unresolved and to stay with it rather than bypassing it. It builds confidence in the formative process through repeated encounters with teachers, peers, and texts where relational exposure is not supplementary but constitutive of learning. It involves long-term symbolic formation, where the sedimentation of disciplinary judgment happens in and through symbolic practice over time. This formation is enacted through embodied presence, relational exposure, and symbolic apprenticeship, but not through dwelling proper, which remains the non-human-made geophysical field that no educational institution controls or curates. What education forms is a sedimented symbolic and relational formation, and that is hard enough to protect without confusing it with the non-human-made field.
The synthetic claim that closes this section: AI's absence from embodiment, its simulation of the surfaces of being-with, and its complete non-relation to dwelling are three different absences. The consequences of delegating work across them are therefore different in kind. Analyses that treat them as one undifferentiated cluster miss the specific character of what is absent in each case.
§9. Multimateriality as Constitutive Ontology, Not Footprint
For a long time the dominant public picture of AI was disembodied: minds in the cloud, pure information, the triumph of software over hardware. The picture was always wrong. The contribution of the LVT framework here is not to be first to notice the materiality but to state precisely why it matters: multimateriality is one of the two mediations that constitute what AI is, not an incidental backdrop.
This needs pushing further than a standard political economy critique. Multimateriality is not just infrastructure and not just cost. It is affordance, resistance, breakdown, latency, friction, repair, and dependence on non-symbolic material processes that symbolic systems cannot talk their way out of. The apparent symbolic fluidity of AI always rides on intensely constrained material arrangements. The GPU cluster can fail. The power supply can be interrupted. The cooling can break. The chip fabrication can be disrupted. None of these vulnerabilities is addressable by the symbolic capabilities of the systems that depend on them. The systems cannot route around their own material dependencies by being cleverer. The material is non-negotiable, and the non-negotiability is what multimateriality is.
The financial structure underwriting current AI development is a multimaterial arrangement whose stability is not guaranteed. The circular flows between hyperscalers, chip suppliers, and energy providers, and the debt structures supporting data centre construction, are part of the phenomenon rather than its backdrop. Analyses that abstract from them produce timelines and expectations that are not well-formed.
The subtler point: the apparent immateriality of AI is itself an effect produced by its material infrastructure. The feeling of interacting with an intelligence rather than a data centre is produced by interfaces designed to foreground symbolic output and background material substrate. This is commercially useful but not ontologically accurate. When the data centre fails, the immateriality evaporates. The rest of the time, the immateriality is a user-interface artefact.
Debates conducted as if AI were mainly "mind" (the policy debates, the philosophical debates, the commercial debates) are speaking nonsense the moment they detach one mediation from the other. AI's novelty is the tightening of the coupling between multimateriality and multisymbolism: material infrastructures now carry symbolic work at densities that did not previously exist, while symbolic processes increasingly reorganise material and institutional life in real time.
§10. Returning to the Critics: How the Grid Composes Their Partial Truths
§2 laid out the major lines of critique. Each captured something real. None saw the whole. With the grammar now in hand, the critical literature can be re-read not as competing positions but as partial reports from different locations in the 5×5 grid.
What the cognitive-offloading literature can and cannot show. Kos'myna and Gerlich have provided empirical evidence, in measurable neural and behavioural terms, that one mode of AI use weakens cognitive engagement. What LVT predicts and what the studies confirm: when the symbolic articulation that users would otherwise perform from a starting point in felt difficulty is outsourced to a system that lacks felt misalignment entirely, the coupling between levels is interrupted, and the user's full-stack cognitive architecture atrophies. The neural connectivity data is the expected signature of substitutive offloading.
But the studies have documented the effects of substitution and named them as though they were the effects of AI-assisted cognition in general. What the literature has not yet studied is whether recursive coupling produces cognitive debt, cognitive amplification, or a mixed profile depending on prior formation. Until that condition is tested, the studies cannot be read as evidence about AI-assisted thinking as such. The framework here does not dismiss their findings; it localises them. It also explains why the effect should be expected, as a prediction of the ontological framework rather than a discovery.
The conceptual vocabulary needs to develop accordingly. "Cognitive offloading" cannot carry the policy weight being placed on it without the threefold distinction in §2b. University guidance that tells students "AI should support your learning, not replace it" is directionally right and conceptually too weak to produce consistent judgments. The stronger formulation: AI must not displace the coupling between felt difficulty and articulation, and it may support articulation when that coupling is preserved.
The pre-existing pathology point deserves emphasis. Many current humanities defences of independent thinking defend symbolic output rather than the processes that make it meaningful. Template writing, formulaic assessment, pre-fixed coding, and ritualised criticality were already weakening formative engagement before AI arrived. AI intensifies these pathologies and makes them more efficient and less visible. The critique of AI-as-substitution is strongest when it acknowledges this continuity rather than pretending that what existed before AI was healthy.
Marcus accurately identifies the absence of lower-stack grounding. Systems trained on symbolic corpora without the lower mediations display exactly the brittleness and failure modes he documents. Where Marcus underplays: that absence does not prevent symbolic recursion from being enormously consequential. Consequence does not require grounding; it requires scale, coupling, and institutional uptake.
Yudkowsky takes seriously the power of untethered high-level optimisation and the failure modes not legible to commercial development cycles. Where he overreaches is the point §6 and §7 addressed: dramatising AI as a sovereign actor with felt stakes imports felt misalignment into a system that lacks it. The correct redescription, a powerful optimiser without the frictions that constrain living beings, is more precise and, in the end, more frightening.
Russell is the critic most compatible with the LVT framework. His insistence that powerful systems remain answerable to a field they do not inhabit already does much of what this grammar makes explicit. LVT strengthens this by specifying what the accountability needs to be accountable to: the mesocosm AI does not live in. Russell's mild limitation is that the framing can sound too technical, as though alignment were primarily a design challenge. LVT broadens it to include the institutional and civilisational dimensions.
Harari identifies the real consequence of concentrated high-level symbolic recursion on institutions calibrated to much lower symbolic throughput. Where he overreaches is into civilisational melodrama. AI is not an alien subject hacking civilisation from outside. It is a hypertrophic output of existing human symbolic corpora running on existing human material infrastructure. The call is coming from inside the house.
Hao and Zitron supply what the others largely lack: sustained attention to political economy, empire, and material cost. Hao's Empire of AI refuses to treat AI's materiality as background. Zitron's work addresses the multimateriality of AI with the seriousness §9 says it deserves. Where these critics sometimes flatten the picture is in treating the field as largely empire or grift, underestimating the genuine, if uneven, reorganisation of symbolic coordination that is occurring alongside the financial instability.
Bender and Gebru name the central problem of confusing statistical language generation with understanding, specifically symbolic articulation without the full mediational stack presented as something it is not. What LVT adds is the positive account: the specific irreducibility of embodiment, being-with, and dwelling, none of which the statistical apparatus touches.
What the grid exposes that no individual critic can see: a single joint intensification of multimateriality and multisymbolism, operating at the level of articulation and above, stably absent from felt misalignment and smooth embodied coordination, not going to change those parameters by scaling. The asymmetric configuration is stable and recursive, not a transient phase. AI does not create a sixth mediation. Nothing about its operation introduces a register of mesocosmic life not already in the five. It does not create a sixth recursivity level either. The suggestion of either is a symptom of the category collapse this essay has been dismantling.
The debate is also, as §1 diagnosed, actively over-stabilised by inherited closures: intelligence, understanding, consciousness, agency, safety, alignment. These terms force a differentiated phenomenon into binary or thresholdable form. The debate is trying to impose closure faster than the ontology allows.
§11. Governance Without Category Error
The failures of AI governance are not primarily regulatory timidity or capture. They are the predictable result of asking badly formed questions and of imposing premature closure over a phenomenon whose actual effects depend on cross-mediation relations and domain-specific recursion. "Is AI a tool or an agent?" "Is it speech or product?" "Is it author or instrument?" These are cases where institutions are attempting to stabilise a rapidly differentiating phenomenon under inherited labels that were never calibrated for it.
The right questions: Through which mediation is this deployment operating? At which recursivity level? What human competences dependent on felt misalignment and smooth coordination are being scaffolded, displaced, or simulated? Where is delegation to symbolic recursion appropriate? Where must embodied judgment remain non-delegable? Where is simulation of being-with substituting for being-with itself? Where is multimaterial concentration producing ecological, financial, or infrastructural fragility? Where is cognitive offloading producing the coupling-disruption Kos'myna and Gerlich have documented, and who carries that cost?
Education is where the distinction becomes most urgent. The cognitive-offloading controversy has generated a great deal of institutional guidance, most of which works with a single distinction: "AI should support your learning, not replace it." That formulation is directionally right but analytically insufficient. Without the threefold distinction from §2b, it cannot consistently identify what counts as support and what counts as replacement, and it will tolerate some uses that undermine formation while prohibiting others that are developmentally sound.
The same tool use may be appropriate or inappropriate depending on when it occurs. AI used before the student has formed any orientation toward a problem is dangerous, substituting for the formation entirely. AI used after the student has wrestled with the material is often legitimate, extending articulatory reach from a position of owned formative engagement. AI used to generate closure before exploration is harmful in the way premature symbolic closure has always been harmful. AI used to stress-test an already-formed argument is legitimate for the same reason that interlocutors, supervisors, and seminars have always been legitimate. The temporal and sequential dimension is not a minor detail. It is the difference between substitution and recursive coupling, which is now the central distinction in educational AI governance.
The real educational risk is not that students use AI. It is that they never develop the coupling between felt difficulty and articulation, and therefore come to mistake polished output for understanding. A generation of students who produce fluent, well-organised work without ever having developed the capacity to sit with difficulty and generate their own direction will carry a formation deficit that credential attainment will not detect. The deficit may not be visible until graduates are placed in conditions requiring genuine formative engagement, and then it will be too late to address at scale.
Three rules organise the educational governance implications.
Rule 1: AI must not replace formative engagement.* Students must still undergo the process of finding a problem, sitting with its difficulty, and beginning to articulate it themselves. This is a structural claim about where cognitive formation happens: in the movement from felt difficulty to attempted formulation, not in the production of any particular output.
Rule 2: AI may support articulation when the student remains the site of orientation.* This includes clarification of unfamiliar concepts, iterative dialogue to develop a position, generation of counterarguments to test an emerging view, help reorganising material already generated, and stress-testing of arguments at advanced draft stage. Recent university guidance has begun to recognise many of these as legitimate, and it is correct to do so. What the guidance still lacks is the explicit reason: what is being protected is the coupling between felt difficulty and articulation, and what makes the approved uses appropriate is that the coupling is preserved.
Rule 3: AI must not impose premature symbolic closure.* The framework, argument structure, coding scheme, or conclusion must remain in human hands, not because of a rule about attribution, but because the formation of that judgment is the educational work.
Recursive coupling is the only mode of AI-assisted learning that is developmentally defensible at scale. Institutions that frame AI governance only as a question of academic integrity (who wrote which words) are missing the deeper question of whether the formative process that the words are supposed to evidence is happening at all.
Beyond education. In clinical work, diagnostic acceleration at the level of symbolic articulation is legitimate, but substitution for embodied attunement in therapeutic relationships is not. In law, summarisation and proposed stabilisations as input to human decision-makers are defensible, but the authority to close cases, which is a question about being-with, institutional answerability, and democratic legitimacy, remains a human matter. In care and companionship, routing vulnerable populations to simulated being-with because real being-with is expensive needs to be named for what it is: substitution of simulation for mediation, not a technological solution to loneliness.
§12. What LVT Does Not Claim
LVT does not resolve the consciousness question. It claims that felt misalignment embedded in mesocosmic life, the specific recursive register that characterises living disturbance, is not produced by symbolic-material recursion at scale, and this is what the debate has been confusing. Whether there is some other register of experience appropriate to AI is a question LVT does not need to answer to do its work.
LVT does not predict AI capability trajectories. It says what kind of thing AI is, not how fast its capabilities will develop. Forecasters and LVT are doing different work.
LVT does not deliver a unified policy. It supplies the vocabulary within which policy can be well-formed. Different political judgments about the same clearly-specified situation remain legitimately different.
LVT does not replace the critics. Kos'myna and Gerlich, Marcus, Yudkowsky, Russell, Harari, Hao, Zitron, Bender and Gebru each remain worth reading. The 5×5 grid lets their partial truths compose into a whole picture. The critics are not in competition with LVT; they are in composition through it.
LVT does not claim AI cannot be consequential, cannot cause catastrophe, cannot reshape institutions profoundly, or cannot warrant serious concern. The opposite: because AI is a stable asymmetric recursive configuration of great power, its consequences are likely to be profound. The refusal of the dramatised versions sharpens what the reality actually consists in.
§13. The Exactness the Debate Has Been Lacking
Return to where this began. The debates on AI have been oscillating between inflation and denial because they lacked a grammar for specifying what they were arguing about, and because the vocabulary they did have was a set of inherited closures forcing a differentiated phenomenon into thresholdable form. Participants have been asking whether AI is intelligent, conscious, social, agentic, alive, and each question has been pulled across a flat plane where no specification was required. Under those conditions, no answer is stable, because the question is not stable.
The grammar has now been supplied. AI is a recursively powerful, mesocosmically partial, multimaterial-multisymbolic intensification whose dangers and uses vary with mediation, level, and domain. It operates entirely within the existing 5×5 architecture of the mesocosm. It does not generate a sixth mediation or a sixth recursivity level, and it cannot acquire the ones it lacks. Scale adds capability within the configuration. It does not change the configuration.
AI cannot develop felt misalignment.* It can model, classify, narrate, and strategically respond to descriptions of misalignment. It can generate articulations of what misalignment looks like from outside. But it cannot have misalignment in the mesocosmic sense, because mesocosmic misalignment is a disturbance in ongoing embodied, situated, social coordination, not a symbolic state. Without embodiment, without the geophysical exposure of dwelling, without being-with, there is nothing there that could be disturbed in the relevant way. No quantity of symbolic articulation crosses into felt misalignment. This is what felt misalignment is, and what symbolic articulation is, and why the two are not convertible.
The cognitive-offloading debate has suffered from the same category collapse that characterises the broader argument. It has treated all AI-assisted symbolic production as one thing, producing a discourse that oscillates between "AI harms cognition" and "AI is just a tool," a smaller version of the same inflation and denial. Once the threefold distinction in §2b is in place, the debate becomes more exact. Substitutive offloading, the displacement of the coupling between felt difficulty and articulation that makes symbolic output meaningful, does damage that the MIT-style studies have now documented at the neural level. Premature symbolic closure was already damaging cognitive formation before AI arrived; AI makes it more efficient and less visible. Recursive coupling, AI operating as an articulatory partner within a human-anchored formative process, is a different operation entirely, and neither the fear of the first nor the residue of the second applies to it.
Cognitive offloading is not the use of AI to produce symbolic outputs. It is the displacement of the coupling between felt difficulty and articulation that makes symbolic output meaningful in the first place. When the coupling is preserved, AI use is not offloading in the sense that damages cognition. When the coupling is displaced, the output, however polished, is not anchored in the formative process that gives it value.
Each confusion dismantled in the essay was a consequence of missing grammar. The reducibility fantasy (§5) mistook one mediation for the general ontological substrate of life. The felt-misalignment fantasy (§6) imagined felt misalignment would emerge from sufficient symbolic recursion. §7 stated directly why more articulation never produces felt misalignment, and applied that claim to learning. §8 differentiated embodiment, being-with, and dwelling as three ontologically different absences, including the complete non-relation to dwelling as the non-human-made geophysical field. §9 pushed multimateriality from footprint to constitutive ontology. §10 showed that the critics compose into a view none of them alone could produce. §11 turned the grammar toward governance, with education as the domain where the coupling distinction is most urgently needed.
The debate does not need a better answer. It needs a better question. Living Value Theory is what that question looks like when it is well-formed. And the debate has failed not because it lacked evidence but because it lacked the right ontology of activity: a vocabulary for specifying what kind of operation is being performed, in which mediation, at which recursivity level, by which kind of system, with what consequences for the formative processes of the people involved.