Abstract
N. Katherine Hayles's Bacteria to AI (2025) reaches many conclusions that Living Value Theory endorses: it decentres consciousness, locates stakes in living beings, and treats humans and technologies as coconstituted. Lacking the mesocosm, the five mediations, five levels of recursivity, three types of recursive counterparts, and differential temporalities, it makes cognition, information, meaning, analogue and digital, and Umwelt carry too much weight. This article keeps her empirical observations and reorganises them. Continuity from bacteria to humans lies in recursive mediation, and cognition is a late symbolic classification. Artificial intelligence is an externalisation and acceleration of multisymbolic activity running without zoetic architecture.
I. Introduction: A Theory Searching for Concepts It Almost Possesses
Reading N. Katherine Hayles’s Bacteria to AI: Human Futures with Our Nonhuman Symbionts (2025) is an unusual experience for anyone who works with Living Value Theory. Again and again she describes phenomena in terms that LVT would endorse almost word for word. She insists that human consciousness cannot define what counts as cognition, that bacteria engage actively and consequentially with their surroundings, that living beings have stakes in what happens to them in a way that rocks and tornadoes do not, and that human beings are constituted through their relations with other organisms and with technologies. She holds that symbolic abstraction is an evolutionary rarity, that material embodiment matters, that biological and technological processes have different histories, and that artificial intelligence should not simply be measured against human capacities. Her stated aim, announced in the book’s first sentence, is to combat anthropocentrism, and she pursues it through what she calls the integrated cognitive framework, which defines cognition broadly enough to include organisms without nervous systems and ties it to information, interpretation, and meaning (Hayles 2025, ch. 1).
LVT is a process ontology of how living beings coordinate with what surrounds them. Its generative principle is that life is recursive mediation in a mesocosm. A mesocosm is the lived field of beings, processes, materials, and conditions within which a living being actually coordinates; mediation names the differentiated ways in which that coordination proceeds, of which five are distinguished; and recursivity names the fact that living processes are altered by their own activity and by the responses that activity generates. From this principle follow distinctions that Hayles does not have: the five mediations and their division into zoetic and poietic kinds, five levels of recursivity, three types of recursive counterparts, recursive opening and closure, differential temporalities, and three kinds of fit, cosmic, mesocosmic, and intersymbolic. The question this article asks is why a framework whose empirical intuitions lie so close to LVT’s nevertheless produces so different an ontology.
The answer proposed here is that Hayles has many of the observations and none of LVT’s fundamental discriminations, so that the result is conceptual overloading rather than empirical error. A handful of terms, above all cognition, but also information, interpretation, meaning, analogue and digital, and Umwelt, have to carry weight that LVT distributes across a much larger architecture. The same pattern recurs in four moves. A feature of living or technical coordination is observed; the available vocabulary is too coarse to differentiate it; a category stabilised at L4, that is, an abstraction that travels beyond the coordination that generated it, is expanded to cover the undifferentiated cases; and the expanded category is then projected back onto the phenomenon as though it were a property of the phenomenon itself. Bacterial responsiveness becomes cognition, consequential differences become information, living responsiveness becomes interpretation and meaning, a forest becomes analogue computation, the operating constraints of a machine become an Umwelt, and the heterogeneous relations among people, microbes, and machines become cognitive assemblages.
What follows is a sympathetic reconstruction. LVT keeps almost all of Hayles’s empirical observations and reorganises their theoretical interpretation, because the fundamental continuity from bacteria to humans lies in recursive living through mediations, whereas cognition is a late and local classification of some of what such living can achieve. Human symbolic cognition is a late development within this history and no part of its foundation. Artificial intelligence, on the same account, is an unprecedented externalisation and acceleration of poietic, and especially multisymbolic, activity that runs without the zoetic architecture of life, and it takes no place as a further organism on a single cognitive continuum.
II. How Close Hayles Already Gets: A Convergence of Intuitions
An honest reading has to begin with how far Hayles has already travelled from the conventional frameworks she inherited, and the distance is considerable.
She begins by loosening the grip of consciousness. She chooses cognition over thought because she wants a concept that starts below conscious human reflection and scales upward, and she devotes much of her earlier work to the argument that nonconscious cognition is a precondition of conscious cognition rather than its poor relation (Hayles 2025, ch. 1). The integrated cognitive framework is built to include organisms without brains or centralised nervous systems, and she observes that most of the biosphere’s living is of this kind. LVT shares the underlying move and extends it. In LVT terms consciousness is not a helpful concept at all, because everything from L1 to L5 would count as consciousness on some usage, and it is precisely the differences between the levels that matter, differences the standard concept never allowed anyone to articulate. Hayles draws one line, between the conscious and the nonconscious, where LVT distinguishes five levels. The point that most recursively consequential living proceeds without reflection, verbalisation, explicit representation, or deliberation is common ground. The disagreement, which later sections develop, is whether all of this living has to be called cognition.
She also treats life as a real discontinuity. She distinguishes material processes, which she calls agents, from living organisms, which she calls actors, and she grounds the distinction in the fact that living creatures have stakes in what happens to them while a tornado or an avalanche does not (Hayles 2025, ch. 1). She does not deny agency to the nonliving, however, and states emphatically that physicochemical processes have it, joins Jane Bennett and other neomaterialists on that point, and parts company with them only over interpretation and choice. LVT agrees strongly with the discontinuity and with the rejection of flattening. Its boundary between life and nonlife is located in metabolic stakes, in what depends on a coordination’s success for a being that must keep itself going, and it can say, as Hayles says, that a landslide has consequences and no stakes.
Her engagement with Lynn Margulis supplies one of the book’s strongest intuitions, that living entities are not self-contained units and that evolution proceeds through symbiosis and coconstitution (Hayles 2025, ch. 3). LVT translates this into the second of its mediations, multispecies being-with, in which coordination with other living beings, of one’s own kind and of others, is a constituent of any living mesocosm. Her technosymbiosis likewise refuses the picture of a completed human being who subsequently picks up external tools. In LVT the same refusal takes the form of the claim that multimaterial and multisymbolic deposits, once laid down, restructure the L1 coordination of everyone who must live among them, so that the human being who writes, walks along roads, and counts money is already a product of those deposits as well as their maker.
Hayles and LVT therefore inhabit nearly the same empirical universe. Their difference lies mainly in what each believes must be invoked to describe it.
III. The First Missing Concept: Recursivity before Cognition
The article’s first theoretical intervention concerns the concept on which Hayles’s framework rests.
Hayles needs a category broad enough to connect bacterial chemotaxis, cellular memory, animal behaviour, human reasoning, and computational processes, and her answer is cognition, defined as “a process of interpreting information in contexts that connect it with meaning” (Hayles 2025, ch. 1). Chapter 4 presses the claim hardest. It argues that cognition begins not with brains but at the level of the cell, and it draws on Michael Levin and Daniel Dennett’s phrase that cognition runs “all the way down” (Levin and Dennett 2020). The two principal cases are the CRISPR-Cas9 system, in which bacteria incorporate fragments of viral DNA into their own genomes and later use them to recognise and cut the same virus, which Hayles calls micromimesis, and the xenobots, assemblies of frog skin cells that acquire new behaviours once released from their place in the frog’s body.
The empirical phenomena do not force this conclusion. Consider the CRISPR case in its own terms. A bacterium encounters a virus; a fragment of viral genetic material becomes part of the bacterium’s own organisation; and on the next encounter with the same virus the consequences are different, because the bacterium is now differently organised. An encounter has altered the living process in a way that alters what a later encounter amounts to. This is recursivity, and it requires nothing further. The bacterium need not represent the virus, interpret a signal, know anything about it, or possess information as an independent thing that it processes. Hayles describes the event as micromimesis because the cell recontextualises the fragment and changes its function, but resemblance and recontextualisation are observers’ descriptions, L4 classifications laid over a sequence that can be stated more thinly: a consequential difference modifies a living process in ways that alter the conditions of subsequent differences.
The core sequence therefore runs from difference through living alteration and changed relation to further difference, where the cognitive account runs from stimulus through information, interpretation, and meaning to response. This is theoretically thinner than the cognitive sequence and, I think, empirically more exact, since it names only what the observations support. Hayles discusses the paper by Chris Fields and Michael Levin, which develops a description of meaning in terms of an organism’s own frames of reference that is explicitly nonrepresentational and does not rely on signs (Fields and Levin 2020; Hayles 2025, ch. 3). She notes the move and then keeps interpretation, information, and meaning at the centre of her framework. LVT takes the thinning one step further.
Hayles overcomes anthropocentrism by extending cognition downward, and a useful name for this strategy is cognitive inflation. She anticipates the objection that emphasising cognition may simply reinforce anthropocentrism, and replies that the sting is removed by insisting that all lifeforms share the capacity (Hayles 2025, ch. 1). LVT overcomes anthropocentrism by decentring cognition itself. It declines to ennoble bacteria by declaring them thinkers, and asks instead why thinking became the standard of ontological dignity in the first place. If dignity attaches to cognition, then the entire biosphere has to be promoted into the category in order to be respected, and the category is stretched until it fits a bacterium, a plant, a server farm, and a philosopher. If dignity attaches to living coordination, nothing needs to be stretched.
Once recursivity is primary, information and meaning fall into place as special descriptions rather than foundations. A chemical concentration can matter metabolically to a bacterium without first becoming an informational object, and it can alter the organism’s course without being interpreted. Hayles holds that information flows between organisms and environments and that meaning is created, modulated, and transformed along those flows (Hayles 2025, ch. 3). In LVT, meaning is an achievement of particular recursive configurations, a late L3 and above stabilisation that some living beings accomplish with symbols, whereas the difference that matters to a bacterium is metabolic and requires no meaning to matter. Information is an abstraction generalised from many cases, and when it presents itself as naming something directly present in a cell it becomes what LVT calls a disguised L4. That diagnosis anticipates the next section, because Hayles needs cognition so badly only because she lacks an account of how recursive processes are differently mediated.
IV. The Missing Architecture of Mediation and Mesocosm
Hayles is deeply committed to mediation, and this is one of the book’s ironies. In her discussion of Karen Barad’s agential realism (Barad 2007) she recognises that reality is never available without situated modes of access, and she coins the term “unmediated flux” for reality as such, which scientific apparatuses can disconfirm models of but never directly present (Hayles 2025, ch. 1). Her own account of emergence, which she calls intermediation, treats each level of organisation as a medium whose patterns lock in the dynamics of the level below (Hayles 2025, ch. 2). LVT can accommodate much of this. Something close to the unmediated flux survives in LVT as the cosmos, which is indifferent to any organism and exceeds every mesocosm, and her constrained constructivism, in which a theory may be disconfirmed by what it cannot accommodate, is an account of what LVT calls cosmic fit. But mediation in Hayles remains general. LVT asks what kind of mediation.
LVT distinguishes five mediations. Three are zoetic, belonging to living process as such. Multisensorial embodiment is coordination with a living being’s own ongoing organisation, carried out through every sensory register at once. Multispecies being-with is coordination with other living beings who respond in turn, of whatever kind. Multiversal dwelling is coordination with the nonrecursive found conditions within which life occurs, the terrain, climate, season, and astronomical regularities that no one made and that do not answer back. Two are poietic, arising when recursive life acquires the capacity to deposit its coordination into durable nonrecursive form. Multimateriality is the deposit of coordination into material organisation, as in a path, a chair, or a bridge. Multisymbolization is its deposit into symbolic form, as in a name, a law, or a calendar, so that a distinction once drawn persists across absence and death. The three zoetic mediations, operating through the first two levels of recursivity, constitute a complete living mesocosm, and this is the mediational world of the great majority of living beings. The poietic mediations are transformations of what the zoetic ones can already do, and their absence in most life marks no deficiency.
Read through this architecture, what Hayles calls cellular cognition can be decomposed. A bacterium responding to a chemical gradient is coordinating with its own organisation through several sensory channels at once, which is multisensorial embodiment. Its relations to its neighbours, its hosts, and the viruses that attack it belong to multispecies being-with, and LVT leaves open whether any of these relations reach interrecursivity in the strict sense, each party responding to the other’s response, tentatively reserving that achievement for animals. The gradients, surfaces, temperatures, and flows within which it moves are multiversal dwelling. A biofilm is a different matter. Its secreted matrix is a deposit that persists beyond the activity that produced it and shapes the coordination that follows, and it belongs to the lowest rungs of the evolutionary stratigraphy of material deposition, where each generation rebuilds much the same structure from its own secretions. Projective fabrication, in which things are made for coordinations that have not yet occurred, on the model of earlier made things, and carried across absence by symbols, marks the human threshold and the arrival of multimateriality and multisymbolization as mediations in their own right. No symbolic cognition has to be posited anywhere in this reading.
Humans, by the same architecture, are best understood as living within all five mediations, the two poietic ones being a transformation of the zoetic three rather than a larger quantity of the same thing, so that human life is no bacterial life with more cognition added. A theorem, a constitution, a novel, a university, a currency, or a large language model cannot be understood by turning up a scalar dial labelled cognition, because each is a deposit of material and symbolic coordination that later life must proceed within. This allows LVT to explain both the continuity and the discontinuity that Hayles wants to explain. The continuity lies in recursive mediation, and the discontinuity lies in the arrival of poiesis.
The concept that Hayles needs most urgently, and lacks, is the mesocosm. A mesocosm is the lived field of available beings, processes, materials, and conditions within which a being coordinates. It is the world as lived rather than the world in itself or the world as represented, it is always specific and layered with the history of earlier coordination, and it comes in as many kinds as there are forms of life. It does not require phenomenological consciousness, it is more exact than a universal environment, and it makes scale relational, since what counts as near, large, or consequential depends on the coordination at hand.
The mesocosm also bears on Hayles’s proposal to reconstruct the Umwelt of GPT-3. She extends Jakob von Uexküll’s term (von Uexküll 2010), which she renders as “world horizon,” to computers and neural nets, arguing that the world a machine can take in and act on is fixed by its architecture, inputs, and outputs, and she describes the equation of this Umwelt with a “mind” as “heuristic, not a literal description” (Hayles 2025, ch. 6). LVT can preserve her insight and decline her terminology. What a system can take in and produce is fixed by engineered multimaterial infrastructure and by an immense multisymbolic archive that human beings deposited, and these conditions are real and worth reconstructing. But a mesocosm is a lived field, and a model has no metabolic stakes in the field it runs within. The conditions of its operation belong to human mesocosms, among the engineered grounds that people coordinate with and within. The heuristic is acknowledged, and it gives out exactly where the question of what the system is begins.
Hayles wants a concept that lets radically different entities inhabit differently constituted worlds. The mesocosm supplies this without requiring those entities to share cognition, phenomenology, or life.
V. The Missing Vertical Dimension: Levels of Recursivity
Hayles’s second great difficulty is to explain how bacteria and humans can be continuous without being simple variants of one cognitive capacity. She wants a framework that is integrated without being flat, and she says so, invoking Cary Wolfe’s “jagged ontologies” as a model of a densely textured conceptual landscape that never resolves into a smooth surface (Hayles 2025, ch. 1). The aspiration is exactly right, and what the framework lacks is a structured account of where the jags fall. LVT supplies one in five levels of recursivity.
L1 is recursive mediation proceeding without becoming separately salient. A healthy body does not have to be monitored to keep working, and a coordination that holds does not announce itself as holding. Most bacterial activity belongs here, and so does an enormous share of human action. L2 occurs when ongoing mediation becomes recursively salient through disturbance, discrepancy, attraction, or interruption, so that something feels off or newly compelling before it can be named. Neither level requires symbols, and both are fully available to the zoetic mediations. L3 is the level at which a difference can be articulated for the first time in symbolic form, as this specific question put into words for this interaction: what is happening, what should be done, why that failed. L4 is the stabilisation of such articulations into portable grounds that travel beyond the coordination that produced them, in the form of classifications, doctrines, disciplines, laws, and institutional categories. L5 is the recursive repositioning of those stabilised grounds themselves, in critique, genealogy, and theory. Each level depends on the ones beneath it without being reducible to them, and the dependency is one of what can exist without what, with no ranking of worth. A being with only L1 and L2 available lives a complete life.
The levels make the status of Hayles’s own vocabulary visible. Cognition is an L4 category, and so are information, interpretation, intelligence, organism, analogue, and digital. Each was stabilised through human symbolic practice, in particular disciplines, for particular purposes, and each travels. Protective and productive L4 abstraction is indispensable, and the diagnosis concerns only what happens when a category stabilised at L4 is applied to processes that extend far beneath L4. LVT calls an abstraction that presents itself as naming something directly present, when it names something generalised from many cases, a disguised L4. Cognition applied to a bacterium has this structure. It reads a human L4 abstraction back into L1 coordination as though the abstraction had been there all along. It also exhibits two of the forms of mesocosmic misfit that LVT distinguishes. It is a recursivity misfit, because a category belonging to a late level is operating at an early one while presenting itself as belonging there, and it is a mediational misfit, because a single term collapses several mediations into one.
The critique is reflexive in an unusual way, because Hayles herself repeatedly performs L5 work. She questions the inherited architecture of humanism, cognition, agency, and technics, and she repositions stabilised categories as objects of inquiry in precisely the manner that L5 names. She performs this high-level recursive critique without having a theory of recursive levels, and the absence has predictable consequences. When a critique at L5 lacks the vocabulary to say where in the structure a given category belongs, it has to use whichever stabilised category lies closest to hand, and cognition lies close to hand.
The levels also yield a better account of continuity than the cognitive ladder. On the ladder, bacteria possess little cognition, humans more, and artificial systems perhaps more again, so that every comparison reduces to a quantity on a single axis. On the LVT account, recursive processes become differently organised across mediations and across levels, and there is no single axis along which the organisms and systems of interest could be ranked. Two beings can differ in which mediations they live through, in which levels are available to them, and in which types of counterpart they coordinate with, and none of these differences reduces to a difference of degree in anything. This becomes decisive for the discussion of artificial intelligence below.
VI. Symbiosis, Counterpart Types, and the Biological Mandate
Three further LVT concepts bear on the middle chapters of the book.
The first is the distinction between three types of recursive counterparts. A counterpart is the aspect of what a living being coordinates with, and LVT asks whether that aspect alters its course in response to being engaged. Nonrecursive counterparts do not so alter: a river bank, a building, a tool, a temperature gradient, a server. Governance, measurement, and intervention can address them without the intervention becoming part of a responsive loop. Selfrecursive counterparts alter in relation to their own ongoing state, as a heartbeat or a healing bone does, and observation or intervention may enter the very regulation it addresses. Interrecursive counterparts are living beings whose responses respond to one’s own responses, so that any intervention becomes part of the interaction it seeks to govern. Recursivity is categorical and is settled aspect by aspect, since a single entity can present both recursive and nonrecursive aspects, whereas stability, predictability, controllability, and durability are gradable properties that vary independently of it.
Hayles’s cognitive assemblages homogenise their participants by making cognition the common denominator, and her own distinction between agents and actors only partly corrects this. Her distinction between agents, which are material processes, and actors, which are cognitive entities, tracks something like the difference between nonrecursive and recursive aspects, but it draws the line through interpretation, and the line therefore falls in the wrong place for computers. In her scheme a computer is an actor. In LVT its hardware, its software, and its trained weights are nonrecursive deposits, built by human coordination into durable form and run by human beings who maintain the power and the infrastructure it requires. Treating them as actors in the same sense as the living beings with whom they are assembled conceals the very asymmetry that matters most for assessing risk and responsibility.
The second concept follows. Symbiosis, once differentiated, is no longer simply mutual cognition. A host and its microbiome coordinate through different mediations, with different stakes and different temporalities, and the relation between a person and a pathogen, a fungus, or a gut community cannot be reduced to a shared exchange of information. The same holds for the human relation to the infrastructures that sustain urban life, which depends on nonrecursive deposits of enormous reach and the recursive people who maintain them.
The third concerns what Hayles calls the biological mandate. She holds that every living being, even a unicellular one, will act so as to continue its existence, that the need to survive and reproduce is universal, and she names the desire to survive the biological mandate (Hayles 2025, ch. 1). She is alert to the dangers of teleology, and in her discussion of Gaia she records Lovelock’s decision to disown the implication that Earth’s systems aim at preserving life (Hayles 2025, ch. 3). Yet the mandate retains a purposive grammar. LVT removes it. Living beings need not hold continued existence as an aim, since what the evidence supports is that organisms undergo locally consequential differences, and that coordination which holds produces continuation. Survival is therefore an outcome of recursive viability and no represented goal. What LVT means by stakes is a fact about the mesocosm, namely what would be lost, damaged, or made unavailable if a coordination fails, and not an organism’s attitude toward its own continuation. This avoids introducing future-directed purpose into the basic definition of life.
The point has consequences for the second of Hayles’s great inversions, in which organisms move from survival and reproduction toward design and purpose while computers begin with assigned design and purpose and may progress toward survival and reproduction. If the mandate is a desire, a machine might one day acquire it. If stakes are metabolic facts about what depends on a living coordination’s success, then the relevant question about any artefact is whether it maintains its own organisation against dissolution, and none of the computational systems Hayles discusses does. Their continuation depends on human maintenance of energy supply, hardware, and software, and what depends on that maintenance is borne by humans.
VII. Analogue and Digital: An L4 Classification Mistaken for Ontology
This section is the article’s sharpest critical intervention, and it begins with Hayles’s sympathetic treatment of the work of Jennifer Gabrys on forests (Gabrys 2022). Gabrys shows how digital sensors are deployed to measure and intervene in forest processes, so that the forest becomes internal to computational devices that interpret it in digital terms. She and her collaborators then invert the relation, so that the forest’s own processes are construed as analogue computational media calculating, for instance, moisture and nutrients in the soil, “analogue vegetal computation” that is ontologically prior to its digital sensing (Hayles 2025, ch. 1). Hayles takes this as a vivid case of what she calls reversible internality, in which each of two systems can be framed as internal to the other, and she is careful to present the choice of frame as consequential, ethical as well as intellectual, rather than as a discovery about what the forest is.
That caution is to her credit, and it brings her closer to LVT than the passage first suggests. LVT adds that the two frames share a premise worth examining, namely that the vocabulary of computation, analogue or digital, is the right one in which to describe a forest at all. A forest is itself an L4 classification of a complex cluster of the three zoetic mediations, the living bodies, the other beings that answer back, and the found conditions of water, light, slope, and season, so that to call it analogue is to pile one classification on another. Neither analogue nor digital exists as such in the mesocosm. A transistor is not intrinsically digital, since its physics is continuous and its digital character is an achievement of thresholds imposed in its use. An electric current is no more digital than a temperature gradient is analogue, since both labels are L4 classifications of particular arrangements of material carrier and symbolic practice, valuable for engineering and misleading as ontology.
The proper distinction is between two kinds of symbolisation. Analogue symbolisation preserves selected continuous differences in a material carrier as systematically corresponding differences in what is symbolised: the height of a mercury column, the angle of a clock’s hands, the modulation of a groove in vinyl, the contours of a map. Digital symbolisation suppresses or ignores continuous material variation so that materially different states count as equivalent instances of one discrete symbol, so that every voltage within a range counts as a one and every voltage within another range counts as a zero. Digitality is therefore an achievement of multimaterial stabilisation in the service of multisymbolization, and it presupposes engineered carriers whose differences have been hardened into reliable thresholds. Neither time nor space is the defining property of the analogue. The deeper property is the preservation of a continuously structured relation, which may be spatial, temporal, metric, intensive, or topological.
The distinction leads at once to the question of temporality, where Hayles’s intuitions are again acute and again underdifferentiated. Chapter 5 takes up the argument of Giuseppe Longo, Maël Montévil, and Stuart Kauffman (2012) that physicochemical processes follow trajectories that can be mapped in advance in phase spaces, whereas biological evolution leaps into adjacent possibilities that no prestated space contains. Hayles accepts the broad distinction but, persuaded by the work on mineral evolution, proposes that temporality is better seen as a spectrum, anchored at one end in celestial mechanics and at the other in the mutable exchanges of the microbial world, and she builds a chapter around the claim that living and mineral processes have different temporalities (Hayles 2025, ch. 5). The intuition is important. LVT replaces the spectrum between two ends with a plurality of temporalities generated within every living configuration. Milliseconds of sensory integration, seconds of movement, minutes of interaction, the hours of metabolic rhythm, the days of waking and sleeping, the years of a biography, the generations of an institution, and the millennia of a symbolic tradition all operate at once, and each mediation has its own temporal bandwidth, a zone in which most of its coordination settles, surrounded by wide variation. Embodiment is bounded by metabolism, dwelling ranges across the widest span from the passing cloud to the tide and the season, and being-with has the most open temporality of the three, since its counterpart responds to one’s own responses. There are not two temporal regimes but many, and they interact recursively because they belong to different mediations.
The difference matters for the relation between symbol and process. An analogue symbolisation can preserve a continuous trajectory. A pen trace on a thermograph drum records the climb and fall of the temperature in the shape of the line, and a reader can see how fast it changed. A digital symbol can abstract from the duration and continuity of what it symbolises. A reading of seventeen degrees does not inherit the history by which seventeen degrees was reached, because the symbol has been made to count the same whenever and however it is taken up. The machine therefore does not escape time. Its symbols can escape the temporality of what they symbolise, and this deserves to be stated as a general proposition. Digitalisation is a technique for partially liberating multisymbolic recursion from the temporalities of the processes symbolised. It is the more consequential because the poietic mediations are in every aspect technologies of time, multimateriality by holding forms steady across spans no body could hold them, and multisymbolization by resting on the premise that a symbol remains the same whenever it is taken up. Digitalisation intensifies the second of these. It is also the thread that connects the analogue and digital distinction to the problem of artificial intelligence, to which the argument now turns.
VIII. Artificial Intelligence without the Cognitive Continuum
This is the section in which the reconstruction pays off, because artificial intelligence is the case that most exposes the cost of building on cognition.
Hayles sees correctly that present systems do things that older theories said machines could not do. In chapter 7 she reports the performance of GPT-4 on examinations designed for humans, from a simulated bar examination to the SAT and the Graduate Record Examination, and the improvements over GPT-3 in mathematical reasoning, comprehension, and coding (Hayles 2025, ch. 7). She is equally clear that machine-generated texts should not be treated as human ones. Her chapter 6 rejects what she calls the null strategy, according to which a text is a text whoever or whatever produced it, and insists that it matters whether a text was made by a human or a machine (Hayles 2025, ch. 6). She also notices the importance of difference in embodiment and evolutionary history, concluding in chapter 9 that a symbiosis between human and artificial intelligence will be productive only if the very different embodiments of the partners are understood. These are the right observations. The difficulty is that, because cognition is her common denominator, the central question must then be how much cognition, understanding, meaning, or protosentience the machine possesses. The Microsoft evaluation she discusses lists the absence of inner dialogue and of the ability to revise earlier output, and she connects these with the conditions of an emergent sense of self (Hayles 2025, ch. 7). The question is posed on a single axis, and everything depends on where along it the system is placed.
LVT changes the question by profiling a system according to the mediations and levels in which it participates. A large language model participates in extraordinary multisymbolic operations, since its trained weights condense an immense archive of human symbolic activity into a form that can be run. It depends on engineered multimaterial infrastructure, of chips, data centres, cables, and power supply. It takes part in exchanges with living readers who prompt it, interpret what it produces, and decide what to do with the result. It does not possess metabolic embodiment, it is not engaged in the maintenance of its own living organisation against dissolution, it has no being-with in the sense of coordination with other living beings who answer back in turn, and it bears no stake in the conditions it runs within. The profile is extremely uneven, with very high multisymbolic reach and none of the zoetic architecture.
The apparent paradox of contemporary AI then disappears. A language model can outperform most humans on some symbolic tasks while being far less zoetically organised than a bacterium, and there is no contradiction in this, since the two comparisons are made along different dimensions. The contradiction appears only when every capacity is projected onto one axis called intelligence or cognition, which forces a ranking of a bacterium and a language model in which one of them has to come out ahead. This is the point at which the absence of a vertical dimension becomes costly, since without the levels there is no way to say that a system produces symbolic output at a rate and fluency formerly confined to L3 and L4 articulation by living beings while lacking the L1 and L2 coordination on which, by LVT’s dependency logic, such articulation depends. The outputs of such a system can be taken up at any level by those who read them. The system, as a run deposit, does not stand at those levels itself.
The three kinds of fit sharpen the point. Standardised examinations, coding tests, and language benchmarks overwhelmingly assess intersymbolic fit, the coherence of a response with a body of other symbolisations, namely the answer key, the examiners’ rubric, and the standard of a profession’s written knowledge. Hayles reports that GPT-4 placed in the upper tenth of bar examination takers, in the ninety-third percentile of the SAT reading and writing examination, and in the ninety-ninth percentile of the Graduate Record Examination’s verbal section, and she notes its results on sommelier qualifications (Hayles 2025, ch. 7). The sommelier tests are instructive. They measure command of the symbolic knowledge of wine, which is a real and difficult achievement, and they say very little about the mesocosmic coordination of a sommelier at work, who tastes, reads a table, and serves in a room with particular people. Benchmarks of this kind tell us a great deal about intersymbolic fit and little about embodied mesocosmic coordination, and that describes what they measure without faulting them. This reading reinterprets Hayles’s account of GPT performance without disputing any of her figures.
The historical event is therefore of a different kind from the one the cognitive continuum suggests. Rather than the next stage of cognitive evolution, in which a further type of mind has arrived, it is better described as follows: multisymbolization, which evolved within lives organised by all five mediations, has deposited its accumulated products, written languages, archives, and corpora, in a form that can now be condensed into technical systems and run on demand, without the zoetic architecture from which human symbolism originally emerged and within which it is normally taken up. What has been externalised is the running of multisymbolic deposits. This formulation is more radical than either of the familiar positions, that such systems really think or that they merely predict tokens, because both of those ask whether the system has cognition, whereas the LVT question concerns which mediations and levels are operative and where uptake occurs. The recursion that deposits require lies in their uptake, in reading, checking, interpreting, and deciding, and in the deployed systems Hayles discusses the weights do not change in response to a conversation. A language model’s outputs are nonrecursive symbolic outputs, and responsibility for what is done with them rests with the living readers who take them up. That the exchange feels like a conversation reflects the recursive discernment of those readers, who attribute answering back to the system as they do to any counterpart, and whose attributions will be settled, as attributions always are, by the feedback that follows.
A further novelty is acceleration, and it is the point on which the analysis of analogue and digital symbolisation bears most directly. Symbolic operations can now occur at rates detached from human reading, from embodied learning, from biographical maturation, and from institutional rhythms. A system can produce in seconds what a living writer would take days to draft. What is being decoupled is the temporality of multisymbolic operation from the zoetic temporalities in which symbolic life has always been embedded, and contemporary AI is better described in these terms than as digital cognition. The decoupling is not symmetrical, because production runs at machine speed while uptake still runs at metabolic speed. Reading, verifying, and acting on what has been produced remain tasks for beings who must eat and sleep, and the burden of reopening and checking a symbolic output falls on those with the least capacity to match its rate. This is a concrete form of the temporal mismatch between mediations that LVT regards as a frequent source of coordination failure, and it is perhaps the clearest example of what the cognitive framework, for all its breadth, has no vocabulary to say.
IX. From Cognitive Assemblages to Differentiated Planetary Systems
Hayles closes the book with political and ecological ambitions, and the conceptual issues recur there on a larger scale.
Cognitive assemblages get a good deal right. Hayles observes that contemporary life depends on infrastructures that combine human beings, computational systems, and nonhuman organisms, and that computational media have become the controllers on which the rest of the infrastructure relies, so that if a pipe fails the water can be rerouted, whereas if the controllers fail the entire facility stops (Hayles 2025, ch. 1). LVT supports this relational intuition. The point is that the dependencies of contemporary mesocosms are distributed across humans, microbes, and machines, and that no account that begins from an isolated individual can describe them.
What gets flattened is the difference among the participants. A person, a bacterium, a language model, a pipe, an electrical grid, a government, and a forest do not participate in the same way. A person lives through all five mediations and coordinates with interrecursive counterparts. A bacterium lives through the three zoetic mediations. A language model is a run multisymbolic deposit supported by multimaterial infrastructure. A pipe and a grid are nonrecursive multimaterial deposits that others must maintain. A government is a stabilisation of interrecursivity, composed of nonrecursive deposits, statutes, buildings, and records, and the recursive people who read, apply, and contest them. A forest is an L4 classification of a cluster of zoetic mediations. To call the whole arrangement cognitive conceals these differences, and LVT instead asks which mediations are active, which entities are recursively alive, which counterpart relations obtain, which levels of recursion operate, where metabolic stakes fall, and which temporalities must be synchronised. Hayles herself senses the difficulty when she resists the idea of planetary cognition on the ground that cognition requires a cognizer with a bounded membrane, and the planet has none (Hayles 2025, ch. 3). The resistance is well founded, and the counterpart types provide the vocabulary that makes it systematic.
The climate crisis then appears as a failure of alignment across mediations, levels, and temporalities. Agreement at L4 that climate change is dangerous coexists with multimaterial infrastructures still organised around fossil fuels, with embodied habits laid down at L1, with institutional inertia, with interrecursive conflicts among those who bear the costs and those who control the decisions, and with economic temporalities measured in quarters and election cycles against a planetary process measured in centuries. The problem is therefore one of misalignment across all of these at once, and inadequate information or insufficient cognition forms only a small part of it. This provides a more precise version of the planetary problem that Hayles wants to solve, and it keeps her own insistence that analysis must end in remedies and stop short of critique alone (Hayles 2025, ch. 1). Remediation, in LVT’s sense, is the repair of coordination, and it requires knowing which coordination has failed.
X. Conclusion: Hayles as an Independent Test of LVT
The opening paradox can now be resolved. Hayles gets an extraordinary amount right. She sees the discontinuity between life and nonlife, the biological stakes of living beings, the nonconscious character of most living organisation, the importance of symbiosis, the coconstitution of humans and technologies, the significance of embodiment, the distributed character of contemporary infrastructures, the inadequacy of human exceptionalism, the strangeness of artificial intelligence, and the importance of different temporal regimes. The shortage lies in conceptual resolution rather than in insight.
The substitutions can be summarised briefly. Where Hayles has cognition, LVT begins with recursivity. Where she has generalised information and interpretation, LVT differentiates five mediations. Where she has the Umwelts of organisms and machines, LVT has mesocosms, which belong to living beings, and treats the operating conditions of machines as part of human mesocosms. Where she has variation in cognitive capacity, LVT has five levels of recursivity. Where she has cognitive assemblages, LVT distinguishes nonrecursive, selfrecursive, and interrecursive counterparts. Where she has biological and computational regimes or the analogue and the digital, LVT examines differential temporalities and the mediational architectures that generate them.
Bacteria to AI can therefore be read as an unusually compelling demonstration of why a framework of this kind is needed. Hayles has already abandoned most of the assumptions that LVT opposes. She rejects the primacy of consciousness, anthropocentrism, bounded individuality, and simple human exceptionalism. Yet without the mesocosm, the five mediations, the levels of recursivity, and differential temporalities, she repeatedly has to convert L4 classifications into ontological principles. The result is a theory that sees the right world but cannot yet differentiate its processes finely enough. LVT allows her empirical world, retained intact, finally to fall into place.
Selected References
Barad, Karen. 2007. Meeting the Universe Halfway: Quantum Physics and the Entanglement of Matter and Meaning. Durham, NC: Duke University Press.
Fields, Chris, and Michael Levin. 2020. “How Do Living Systems Create Meaning?” Philosophies 5 (4): 36.
Gabrys, Jennifer. 2022. “Smart Forests Atlas.” www.jennifergabrys.net/2022/11/smart-forests-atlas.
Hayles, N. Katherine. 2025. Bacteria to AI: Human Futures with Our Nonhuman Symbionts. Chicago: University of Chicago Press.
Levin, Michael, and Daniel Dennett. 2020. “Cognition All the Way Down: How to Understand Cells, Tissues and Organisms as Agents with Agendas.” Aeon, 14 October.
Longo, Giuseppe, Maël Montévil, and Stuart Kauffman. 2012. “No Entailing Laws, but Enablement in the Evolution of the Biosphere.” Proceedings of the 14th Annual Conference on Genetic and Evolutionary Computation.
von Uexküll, Jakob. 2010. A Foray into the Worlds of Animals and Humans with a Theory of Meaning. Translated by Joseph D. O’Neil. Minneapolis: University of Minnesota Press.