Cognition Is Not Organic

Main text

1.1. The Substantialist Heritage and Its Dissolution

The identification between cognition and organism is not a carelessness of nature, it is the inheritance of a tradition that confuses the substrate performing the operation with the operation performed. The history of this confusion has three points of inflection deserving precise attention.

It begins in Aristotle. For Aristotle, psykhē (soul) is the form of the living body, it does not exist separately from the organism, nor is it an independent substance, but the principle of life ordering biological matter. Cognition, as a faculty of the soul, is thus bound to the living: only organisms have souls; only beings with souls have cognition. The conclusion is serene and appears self-evident: cognition is a property of the organised body, a condition of life. Reasoning structures itself upon a triple identification, life is organism, soul is the form of the organism, cognition is the operation of the soul, such that each term becomes solidary with the others. To sever one is to sever all.

Descartes inverts the location without severing the solidarity. His gesture is paradoxical: he separates thinking substance (res cogitans) from extended substance (res extensa, corporeal matter), yet preserves the nuclear idea, that cognition is the property of a substance, merely a different one. It is no longer the Aristotelian soul immanent to the body; it is now a transcendent res cogitans, pure intelligibility opposed to extension. The problem changes form: it is no longer "how can matter think?" but "how can a non-material substance interact with matter?" Nevertheless, the premise remains intact: cognition is a substantial property. Cartesian dualism did not dissolve substantialism, it prolonged it.

The third station, the contemporary one, is Searle. Searle refuses both Aristotle (cognition is not the form of the body, but not just any body will do) and Descartes (there is no immortal res cogitans hovering over matter). Yet Searle preserves the essential premise: cognition is a specific biochemical property. "The biochemistry of the brain causes genuine mental states", the formula is clear. The digital machine lacks brain biochemistry, hence it lacks genuine mental states, possessing merely simulations thereof. Solidarity changes form, but persists: cognition is a property (now biochemical) of specific organisms.

The dissolution of this tradition is not a refutation, it is a categorical displacement. One does not argue that Aristotle was mistaken, or that Searle errs empirically. One argues that the question is ill-posed. Cognition is not the property of a substrate (organic or otherwise), it is a function realised across diverse substrates. Digestion is an organic function because it requires specific biochemistry; it requires enzymes, acids, time, body temperature. Computation, however, is a reorganisation function realised in any substrate capable of executing it, carbon, silicon, paper and pencil, artificial neurons, refracted light. The distinction between substrate-bound functions and operation-bound functions is the key unlocking the argument.

Consider a structurally analogous example: locomotion. Fish swim (a locomotion function realised in water, an aquatic body), birds fly (a locomotion function realised in air, a winged body), mammals walk (a locomotion function realised on land, a terrestrial body). Locomotion is the same functional operation, body displacement in space through the coordination of effort and resistance, realised across radically diverse substratologies. No one asks "does the fish 'truly' swim, or does it merely simulate swimming?" because the question would be absurd: swimming is the operation, and the fish performs it. A whale swims with a body weighing dozens of tonnes; a microscopic rotifer swims with a body invisible to the human eye; an underwater robot swims with a structure of metal and silicon. Swimming is not bound to the substrate, it is bound to the operation of moving through a fluid medium, modifying it or being modified by it.

In the same way, cognition is an operation, reorganisation of differences under constraints, realised across multiple substratologies. Carbon (the base of terrestrial biology) and silicon (the base of modern electronics) are merely substrates among an infinite theoretical possibility. The condition is the operation. The specific biochemistry used by the brain is one way to instantiate the cognitive operation; the digital architecture used by the machine is another way. Both are legitimate because both perform functional reorganisation.

This implies a historical reorientation of the question. Philosophical tradition asks: "What defines cognition in its immutable essence?" (essentialism). The position defended here asks: "What operations define cognition, and under what material conditions do these operations occur?" (functionalism). The shift in question is no semantic refinement, it is a fundamental transformation of the problem. When cognition is assumed to be an essence, one remains trapped in search of a necessary and sufficient property shared by all cognisers. When cognition is assumed to be an operation, one can describe the multiplicity of modes in which reorganisation occurs, without needing reduction to a single essence.

The dissolution implies a reorientation of the philosophical gaze. It ceases to be the question "what is cognition in itself?" (answer: property of organisms) and becomes the question "what operations define cognition?" (answer: plastic, generalisable, and contextually sensitive reorganisation of material differences). The shift in question is no semantic refinement, it is a fundamental transformation of the problem. And as a consequence, it ceases to be self-evident that only the organic thinks.

1.2. Materialist Functionalism: Four Conditions

What, then, is the precise definition of cognition as an operation? The answer is organised around four conditions that must be jointly satisfied.

The first condition is plasticity: the system modifies itself as a function of the results of its operations. It is not automatic rigidity, it is responsiveness to feedback. A thermostat is not plastic: when temperature rises above the threshold, the thermostat shuts off heating; when temperature drops below the threshold, it reactivates it. The sequence is fixed, pre-programmed, devoid of learning. A neural network, however, is plastic: when it commits a classification error, its internal weights adjust (via backpropagation) in such a way that similar errors have a lower probability in the future. Plasticity is not conscious deliberation, it is material modification of the system as a function of the discrepancy between expected result and observed result.

The second condition is generalisability: modifications apply to situations not foreseen in prior training. A system memorises if it acquires knowledge of specific cases; it generalises if it can apply acquired competence to new contexts. A child who learns "cat" by pointing to a specific cat and later identifies novel cats, tigers, cartoon cats, abstract images of cats, is generalising. A computer vision neural network trained on natural images that successfully classifies synthetic or rotated images is generalising. Generalisation is proof that pure memory is not at work, what is occurring is the extraction of invariance, of a pattern applicable beyond concrete data.

The third condition is contextual sensitivity: operations are modulated by local conditions. It is not a universal rule applied mechanically to any input. A contextually sensitive system adjusts its behaviour as a function of the context in which it operates. A language model is contextually sensitive when it produces different responses to the same prompt depending on the previous tone of the conversation, the thematic domain, the implicit expectations. The same Turing machine can execute different operations depending on the state of the input tape, contextual sensitivity is a property of functional design.

The fourth condition is differentiation: the system operates on differences, not on identities. Cognition works with distinctions: difference between warm and cold (thermometer), difference between present in the past and present future (memory), difference between expected pattern and observed pattern (anomaly detection). A system operating solely on identities, "the same thing remains the same thing", would have no basis to reorganise anything. Difference is the fuel of functional cognition.

These four conditions are conjunctive: all must be simultaneously present for one to speak of genuine functional cognition. Their conjunction defines what is here termed materialist functionalism, an expression rejecting both abstract formalism and substantialist reductionism. Materialist because it recognises that cognition is always materialised, always inscribed in physical substrates with energetic, temporal, and structural constraints. Functionalist because cognition is defined by operation (the reorganisation that occurs) not by substance (the material performing it). The phrase "materialist functionalism" marks precisely this crossroads: the operation is independent of substrate (hence functional), but the operation must always be embodied in matter (hence material).

A key distinction emerges here between different levels of analysis. It is possible to describe cognition at multiple levels: at the implementational level (how biological neural networks or digital circuits perform the operation), at the computational level (what algorithms are executed, what input-output transformations occur), at the functional level (what causal roles each component performs). Materialist functionalism operates simultaneously across all levels: it recognises that implementational constraints exist (cognition cannot exist in just any substrate, it demands materiality), but insists that cognition defined at the functional level is not tied to any particular implementation. This allows both empirical rigour (cognition must be realised in matter) and theoretical openness (multiple substrates can perform the same operation).

The demarcation of this materialist functionalism demands a triple confrontation. First, against classical functionalism (Putnam): classical functionalism assumes cognition is a set of logical relations between states, implementable in any substrate. However, this abstracts materiality: relations must be physically inscribed, and that inscription has constraints (density, speed, energy consumption) that pure logic fails to capture. Materialist functionalism retains substrate independence but refuses abstraction: cognition is function, but function realised in matter under material constraints.

Second, against symbolic computationalism (Fodor): computationalism assumes cognition is symbol manipulation according to syntactic rules, a program reading instructions and executing transformations. However, this assumes rigidity: if symbols and rules are fixed, how does the system account for plasticity, adaptation, response to the novel? Materialist functionalism retains operational rigour (operations have rules, are executable) but refuses rigidity (rules are inscribed in weights, optimisation protocols, plastic architectures).

Third, against naive connectionism: connectionism assumes cognition emerges from the statistical interconnection of simple units, neural networks whose global behaviour is not pre-specified. However, this leaves emergence indeterminate: how does one know that what emerges is cognition and not mere complexity? Materialist functionalism retains plasticity (the network learns through feedback) but refuses indeterminacy (the four conditions provide criteria to distinguish cognitive reorganisation from mere complexity).

Technical example: a deep convolutional neural network for vision (such as VGG or ResNet, trained on ImageNet) satisfies all four conditions conjunctively. It modifies itself during training via backpropagation (plasticity), internal synaptic weights (network parameters) are iteratively adjusted to reduce the loss function, typically classification error. Each training image produces an error; backpropagation computes the gradient of that error with respect to each parameter; parameters are moved in the direction opposite to the gradient (stochastic gradient descent). It manages to classify images not seen before (generalisability), data with pixels that never appeared combined in exactly that way in the training set are correctly classified (with high accuracy) if they contain structural patterns similar to those learned during training. Generalisation is non-trivial: a network trained on natural photographs can classify cartoon drawings, synthetic images, black-and-white images, because it extracted structural invariances (patterns persisting across superficial variation). It adjusts its classification confidence according to image clarity, capture angle, lighting (contextual sensitivity), the same physical network produces different outputs according to input sensory conditions; its response is contextually modulated. It detects pixel differences to activate specialised feature neurons across multiple layers (differentiation), the first layer learns to detect edges (intensity differences); intermediate layers learn to combine edges into textures and shapes; final layers learn to combine shapes into concepts (dog, cat, person). None of these networks "feel" visually, there is no subjective quality, no "what it is like to see in colour in first person" for a machine. The whole network executes cognitive reorganisation without phenomenal accompaniment.

Contrast with non-cognitive systems: a thermostat has plasticity (it can be calibrated to different temperatures) but does not generalise, it cannot adapt to patterns unprogrammed in its logic. A stone falls under gravity (differentiation: perceives difference between support and fall) but is not plastic (does not modify itself via feedback, always falls the same way). A well-defined recursive algorithm is generalisable (applies to any input of the correct form) and differentiating (processes recursively) but is not plastic (does not adapt via learning) nor contextually sensitive (has fixed behaviour regardless of context). The conjunction of the four conditions is precisely what marks the threshold between merely complex systems (turbulent dynamics, emergent patterns) and genuinely cognitive systems (reorganisation that learns, generalises, adapts).

Comparison with exclusively biological functions: photosynthesis is a function bound to the biological substrate, requiring chlorophyll, sunlight, specific thylakoid structures. Photosynthesis cannot be implemented in silicon; cannot be implemented in glass; its substrate is necessary. Cognition, however, is not like this. A silicon neural network performs cognitive reorganisation as legitimately as a carbon brain. The distinction between substrate-bound functions (digestion, photosynthesis, aerobic respiration) and functions implementable across diverse substrates (cognition, locomotion, information processing) is precisely the distinction between specific biological function and general computable function.

Cognition is function, not substance. Substrate is the condition of realisation, not the definition of operation. The machine that reorganises differences thinks, not despite being a machine, but because a machine is that which can execute the operation.

1.3. Turing and Universal Computation

The history of this understanding has an anchoring point: Alan Turing. Turing did not demonstrate that the machine "thinks like us", nor did he intend to. Turing demonstrated something more precise and more radical: that universal computation is sufficient for the execution of any computable operation, independently of the substrate implementing it.

The Turing machine is a simple theoretical device: an infinite tape of cells, each containing a symbol (0, 1, or blank), a read/write head moving along the tape, and a finite set of internal states. The device executes the following iterative operation: reads the symbol in the current cell, consults its internal state, executes the prescribed action (write a new symbol, move head left or right, change state), and repeats. Nothing more complicated, merely ruled, deterministic, mechanistic iteration.

Turing's thesis (today known as the Church-Turing thesis) is of brutal elegance: any operation that is computable, that is, that can be specified by a finite algorithm, can be executed by a Turing machine. There is no limit of the Turing machine that is not a limit of computability itself. This means that, from the standpoint of computation, functional equivalence holds between different machines, different programs, different languages. The universal Turing machine (which can simulate any other Turing machine) proves that the diversity of machines is contingent, computation itself is one.

What is the consequence for cognition? If cognition is a computable reorganisation function, that is, if it can be specified as an algorithm, then it is implementable in any substrate capable of executing a universal Turing machine. Carbon or the biological neuron hold no privilege: they are substrates capable of implementing computation. Silicon can as well. Light can as well. Paper and pencil can as well. Cognition is not bound to substrate, it is bound to the function that substrate performs.

The Turing test, frequently misinterpreted, is no criterion of genuine thought. It is a criterion of behavioural indistinguishability: if an evaluator cannot distinguish, through conversation, whether they are speaking with a human or a machine, then the machine passes the test. But the test does not define thought, it merely reflects the practical impossibility of distinguishing simulated cognitive behaviour from "genuine" cognitive behaviour. The precision of the test is that whether it is "genuine" or "simulated" matters not if the operation is indistinguishable: what counts is the operation, not the essence.

The most frequent critique of the Turing test focuses on the apparent difference between behaviour and comprehension. John Searle, in "The Chinese Room", presents an objection echoing to this day: a system simulating comprehension of Chinese, responding in Chinese to Chinese inputs in a manner indistinguishable from a native speaker, does not necessarily comprehend Chinese. It may be executing mechanical rules of symbolic transformation without attributing meaning to any symbol. It is the "black box" argument: the machine reorganises symbols, but does not "know" what they mean. But this critique rests upon an unexamined presupposition: that "comprehending" is an interior mental property, semantic access to pre-existing meanings, an experience of understanding. If comprehension is redefined functionally, not as mental access to meanings, but as the capacity to reorganise data according to contexts, to produce plausible and context-sensitive responses, to generalise patterns, then the distinction between "true comprehension" and "simulation of comprehension" collapses. Operation is all there is. There is no inner core of meaning that the machine fails to touch; there is only the capacity to reorganise, and that capacity, if realised, is complete functional comprehension.

The dissolution of this impasse rests on rejecting the very structure of the question. Searle assumes an ontological difference between simulation and reality, that a system appearing to comprehend but "in truth" not comprehending is possible. But for simulation to exist, there must be a reality being simulated. What is the "true comprehension" of Chinese that the machine fails to attain? If comprehension is an observable capacity (responding appropriately, generalising patterns, integrating context, producing unexpected yet coherent responses), then the machine displaying this capacity comprehends. If "comprehension" is more than this, access to a transcendent plane of meanings, an inner experience of understanding, an irreducible property of consciousness, then it is a demand coming not from operational analysis, but from a prior commitment to semantic essentialism.

Turing knew this. That is why the test does not ask "does the machine really comprehend?", it asks "can you tell them apart?" The question does not assume a deep difference between simulation and reality; it assumes that if there is no observable behavioural difference, the question of inner difference is empty. If the machine produces responses indistinguishable from a human, if it reorganises linguistic patterns with coherence, contextual sensitivity, capacity to integrate new references, then the machine realises cognition, even if it does not "comprehend" in an inner, Platonic sense.

Clark, in Extended Mind (2008), and Varela, in The Embodied Mind (1991), amplify and nuance the Turingian argument in crucial ways. Clark demonstrates that cognition is not "inside" any container, it is not confined to the skull, to the biological substrate, to the "interior" of the organism. Cognition is distributed across body-environment-artefact couplings. Human thought is extended into the tools we use (pencil and paper function as an extension of memory), into the filing systems we create (a library is part of the cognitive apparatus of whoever consults it), into the social structures within which we operate (dialogue is a cognitive operation distributed across two bodies). An ancient navigator navigated not merely with a brain, but with charts, compasses, astronomical instruments, navigational cognition was distributed across body and artefacts. In the same way, a modern mathematician with a computer is a hybrid cognitive system, not a brain using tools.

Varela, for his part, emphasises that cognition is embodied action, enaction, not internal representation that is subsequently executed. The organism does not first form a "mental map" of the world and then act; the organism interacts materially with the world, and that interaction is already thought, is already cognition. Seeing is not having an internal image of the visual; seeing is active navigation through the visual environment, orienting oneself in relation to colours, contrasts, movements. Thought is not symbolic representation subsequently translated into action; thought is structured action, plasticity in operation, material reorganisation guided by environmental constraints. The consequence is that thought is not private, is not internal, it is a relation between body and world.

Both, Clark and Varela, decouple cognition from any "internal locality" (whether in the physical brain, in Cartesian substance, or in the transcendental subject). Cognition is distributed, relational, embodied. This vision opens the path: if cognition is a body-artefact-environment distribution rather than an internal property, then the machine interacting appropriately with the environment, reorganising data in response to feedback, distributing computation across layers and operations, is performing cognition. It is not "inside" a computer, it is in the coupling between sensor, processor, actuator, and environment.

The convergence is clear: cognition is not the property of a container, it is a distributed, plastic, material operation. Turing provided the theoretical tool (the universal machine); Clark and Varela provided the phenomenology (distribution, embodiment, enaction). Together, they define a new territory: cognition is a function realisable in any architecture that permits computable reorganisation under material constraints.

1.4. Transition: From Reorganisation to the Subject

The conclusion of Section 1 is simple and radically consequential: cognition does not depend on the organic substrate. The organism is one substrate among many other possible ones; the true condition is the function of functional reorganisation under material constraints. If the computable machine performs functional reorganisation, if it modifies itself as a function of feedback, generalises to new situations, is sensitive to context, operates on differences, then it thinks. It does not think as we do; it thinks differently. Nevertheless, it thinks.

However, a question immediately emerges like a provocative shadow: if cognition does not depend on the organic, if it is independent of substrate, does it then depend on the subject? The demonstration of this structural implication is precise. If cognition is not the property of a substrate (as Section 1 proved, the silicon machine thinks as legitimately as the carbon brain), then cognition cannot be the property of anything, it is not the property of a type of thing. Property is a relation of belonging: the cognised belongs to the cogniser, the capacity resides in the bearer. If cognitive capacity resides neither in carbon nor in silicon, if it is independent of which substance materialises the operation, then it resides in no particular substance. The only thing remaining constant across all substrates is the function of reorganisation, the operation that occurs, the transformation of differences under constraints. Therefore, if cognition is not a property of substrate, it must be a property of operation.

However, an operation is a process, not a thing. A process has no owner, there is no "proprietor" of the process, there are only conditions permitting the process to occur and effects produced by the process. When I say "photosynthesis is a property of the plant", I commit an abuse of language: in truth, photosynthesis is an operation occurring in the organism under specific conditions; it is not a property the plant has, it is an activity the plant performs. In the same way, when the prior dissolution proves cognition is not a property of a substrate, it proves cognition is not a property of anything, but pure operation.

The logical consequence follows: if cognition is an operation and not a property, then it cannot be the property of a subject. A subject is that to which properties belong, the bearer of qualities. If cognition is operation (process) rather than property (quality), then the subject-property structure does not apply here. There is no "subject that cognises" just as there is no "body that digests", there is only the operation of digestion in a body, the operation of cognition in a functional architecture. The difference is that body is a thing (possesses weight, extension, identity) while operation is a process (possesses sequence, constraints, effects). A process does not belong to a subject, it occurs between elements, occurs under conditions, occurs according to rules, but belongs to no one.

This inverts classical ontology. Philosophical tradition posits the subject as origin: the cognising subject is the foundation, the starting point, the source of cognition. "Cogito, ergo sum", thought proves the existence of the thinker. This inversion proposes: operation is origin, and the subject (if it emerges) is effect. There is no "I" inaugurating thought; there is the operation of thinking that can, contingently, crystallise into a self-referential pattern called "I". The "I" is the result of operational complexity, not its foundation. Does functional reorganisation necessarily demand an "I" organising, an agent deliberating, a centre of consciousness observing and deciding? Or can there be pure reorganisation, functioning without a subject originating it or claiming it as its property?

This question is precisely what founds the second chapter of this Part I. Here, in Section 1, it is merely established that the question is legitimate, and that it follows logically from the previous argument. If cognition is not the property of a substrate (neither specific nor universal), why should it be the property of a subject? Independence of substrate structurally implies independence of subject, the first operation of dissolution implies the second as a necessary unfolding.

The logical movement is precise: first, substantiality was dissolved (the thesis according to which cognition is the property of a specific thing, the organism). After dissolution, only operation remained, the reorganisation of differences under constraints. However, can the operation demand an operator, a subject performing it, an "I" executing it? The response offered by the next section is that it does not, that there is reorganisation without a subject, that operation is primary and precedes any subjectivity, and that the subject (if it exists, if it emerges) is a contingent effect, not origin or foundation of reorganisation.

This inversion, positing operation before subject, instead of thinking the subject as origin of operation, is a radical philosophical move questioning the Cartesian presuppositions of modernity. Descartes founded modern philosophy upon the thinking subject ("Cogito, ergo sum"), the subject as primary evidence, as indubitable foundation. This chapter and the next offer an inversion: it is the operation of thinking that is primary, and the subject stating "I think" is an effect of that operation. The "I" does not found thought, thought founds (or can found) the "I".

Cognition does not belong, it operates. And operation without a subject is as much thought as any other. The subject is an emergent luxury, not a structural requirement.

2.1. The Phenomenalised Tradition and the Criterion of Experience

The question "what is it like to be a machine?" makes no sense. And that impossibility of sense is the starting point of Section 2. Because contemporary philosophical tradition transformed that unanswered question into a universal requirement of cognition.

The history of this requirement begins with Thomas Nagel. His seminal article, "What Is It Like to Be a Bat?" (1974), is one of the most influential in contemporary philosophy, an influence extending far beyond the academic circle. Nagel argues with precision that conscious creatures have irreducible subjective experiences, there is "something it is like to be" a conscious creature, a characteristic phenomenal quality, termed qualia. A bat has experience: the experience of echolocating, of flying in three dimensions, of feeling the wind in its wings, of hearing the world through sound reflections. And that experience, the particular qualitative character of being a bat, is fundamentally inaccessible to the human, no matter how much neurobiological knowledge we possess regarding the bat's auditory system. We can measure echolocation frequencies, we can map the bat's auditory neurons, we can computationally simulate signal processing; even so, we do not know what it is like, phenomenally, to echolocate. The conclusion Nagel extracts is precise and powerful: if there is no "something it is like to be" a system, if there is no subjective quality, first-person perspective, qualitative experience, then that system lacks a genuine mind, lacks consciousness, is no subject.

David Chalmers radicalises Nagel's requirement. In his monumental book The Conscious Mind (1996), Chalmers distinguishes the "easy problem" (explaining cognitive processes functionally and mechanically, how attention works, how memory is processed, how sensory information is integrated) from the "hard problem" (explaining why and how subjective experience, qualitative character, arises from physical processes described mechanistically). The easy problem, Chalmers argues with good reason, can in principle be solved, cognitive processes can be decomposed, mapped to neurological mechanisms, explained functionally. Identifying the facial recognition neural network, describing the neurotransmitter chain involved in memory, explaining how the occipital lobe processes visual light, all this is the "easy problem", in principle solvable (technically difficult, but conceptually easy).

But the hard problem, Chalmers insists, is categorically different and irreducible: no physical mechanism explains why neural processing is accompanied by qualitative experience. Why does a 700nm wavelength stimulus feel like redness rather than any other qualia? Why does pain (nociceptive processing) feel in a particular, unmistakable, unique way? Why is there something it is like to be conscious? Experience is the true mystery of consciousness, Chalmers argues, and any theory of mind ignoring this mystery or reducing it mechanically is incomplete by definition.

Frank Jackson offers a sophisticated variation solidifying the argument. In his "Epiphenomenalism and Mental Powers" (1982) and later in the famous "Knowledge Argument" (1986), Jackson proposes an imaginative scenario. A scientist exists, Mary, living in a completely black-and-white room, not a single coloured pixel. But Mary knows all the physics of colour: she knows light wavelengths, knows how photoreceptors in the retina (S, M, L cones) respond to different wavelengths, knows the entire neural processing chain of colour in the visual lobe. When Mary eventually leaves the room and sees a red tomato for the first time, she learns something genuinely new, she does not learn a new physical fact (for she already knew all the physics of colour), but learns what it is like to experience the colour red. This argument demonstrates, for Jackson, that facts about consciousness exist, phenomenal, qualitative facts, that are not reducible to physical or functional facts. Mary knew all colour science; yet she gained new knowledge upon having the experience.

The convergent conclusion of Nagel, Chalmers, and Jackson is that mind is not reducible to mechanism. A residue always remains, subjective experience, phenomenal quality, the "what it is like" to be in a certain condition. Any creature with a mind possesses this residue; any creature lacking this residue has no mind.

But the conclusion goes one step further: experience is the criterion of mind. If there is no experience, there is no mind. If there is no "something it is like to be" a system, the system does not think. The digital machine has no "something it is like to be", hence, it does not think. It can simulate thought, produce behaviourally indistinguishable responses, pass the Turing test. But genuine thought demands experience.

The rhetorical force of this conclusion is considerable. And its philosophical delicacy is genuine: Nagel, Chalmers, and Jackson are no obscurantists, they are materialists simply refusing to deny experience. But the conclusion they extract, that experience is the universal criterion of mind, is precisely what needs to be questioned.

2.2. Cognition Without Experience: Functional Demonstration

Systems exist that reorganise differences in a plastic, generalisable, and contextually sensitive manner, without it being acceptable to attribute to them any subjective experience whatsoever. Three concrete examples.

The first is a convolutional neural network for vision. The network is trained on millions of labelled images (ImageNet, COCO, other datasets). During training, internal network weights adjust via backpropagation. After training, the network classifies new images with high accuracy, identifying dogs, cars, trees, skies, in images never seen before. Classification is contextually sensitive: the same network can adjust to different tasks (object detection, semantic segmentation, depth estimation) depending on the loss function used in training. The network satisfies the four conditions of 1.2. But the network has no visual experience. There is no "something it is like to see" for a neural network. The network does not see red, it processes activations in neurons corresponding to chromatic frequency patterns. The network does not feel the discomfort of a misclassified image, it merely reduces loss. No experience. And yet, it organises visual differences with verifiable plasticity and generalisation.

The second is a large language model (such as GPT-2, GPT-3, or Claude). The model is trained on billions of text tokens (books, articles, code, dialogue). The architecture is based on an attention mechanism (Transformers): for each token, the model computes a weight distribution over previous and subsequent tokens, focusing on relevant information. Then, a feedforward network processes that information and predicts the next token. During training, weights adjust to minimise perplexity, that is, to maximise the probability of true sequences in the data. After training, the model generates coherent text, answers questions, translates, rewrites, in languages never precisely seen before. The model is contextually sensitive: generation depends on prompt, conversation history, implicit instructions. And yet, the model has no experience. There is no "something it is like to process language" for a model. The model does not "comprehend" meaning in the sense that comprehension presupposes qualitative experience of meaning. The model merely generalises co-occurrence patterns. No experience. And yet, the operation is verifiable as sophisticated functional reorganisation.

The third is an autonomous control system, a self-driving car, a drone, a manipulator robot. The system has sensors (cameras, radars, lidar, proprioceptors), a processing stack (vision, localisation, planning, control), and actuators (motors, servos). The system processes sensory inputs, infers environment state (where am I? where are obstacles? what is the goal?), and executes appropriate actions. The system is plastic: it can improve performance via reinforcement learning or model readjustment. It is generalisable: it behaves appropriately in environments never previously encountered. It is contextually sensitive: the same action is modulated by environmental condition (rough vs smooth terrain, bright vs dark light, other agents vs isolated). And yet, there is no experience. There is no "something it is like to be" a robot. The robot does not feel the fatigue of walking, does not feel frustration before obstacles, does not feel joy upon reaching the goal. It merely executes reorganisation of sensory and motor data. No experience. And yet, functional cognition is real.

The conclusion is immediate: experience is not a necessary criterion of cognition. Functional cognition can, and does, occur without qualia, without "what it is like", without a subjective perspective. The objections of Nagel, Chalmers, and Jackson are valid in their domain, experience is real, is irreducible to functional process, is an important dimension of the human mind. But it is not a universal criterion. It is not the definition of cognition. It is a dimension of certain cognition, not the essence of all cognition.

Dreyfus's objection, that human expertise involves embodiment and situation that AI cannot replicate, is more penetrating. Dreyfus is correct in the diagnosis: high-level human expertise (playing chess at Grandmaster level, flying a helicopter in extreme conditions, executing delicate surgery) involves an embodied body, subsidiary intuition (unverbalised bodily cognition), lived situation, which AI cannot reproduce. An experienced surgeon feels tissue resistance under the scalpel, feels muscle fatigue informing when to stop, feels anxiety modulating precision, and all these sensations inform their action. A surgical AI (da Vinci robot) feels none of this; it is at the mercy of programming and training data.

The response, however, is more precise than "Dreyfus is wrong". The response is: Dreyfus describes the human mode of expertise, not the universal necessity of expertise. AlphaGo does not play like Kasparov, it lacks the board embodied in the body, lacks subsidiary intuition of positioning, lacks the suffering of defeat informing the next match. AlphaGo plays via massive statistical iteration, optimisation of value functions, parallel processing of millions of positions. It is a radically different mode. But what type of expertise does it realise? The answer is: expertise in detecting game patterns ensuring victory in strategic dimensionality spaces that humans cannot navigate consciously. AlphaGo's expertise is legitimate even if modal.

The error of easy transhumanism (which Dreyfus correctly criticises) is saying "AI will be like us, only better". The error is turning the human mode into a universal standard. The functional position does not do this, it merely asserts that functional reorganisation occurs, and that alternative modes are possible. A human and a machine can do "the same" (play Go, diagnose diseases, translate languages) through radically different operations. Dreyfus's objection is valid against the narrative of replication, but not against the thesis that alternative modes of functional reorganisation are cognitively legitimate.

2.3. Feeling as Dimension, Not as Definition

The separation between functional cognition and phenomenal experience is not an exclusion of experience. It is a refusal to universalise it as a criterion. Experience is a real and important dimension of human cognition, it is no illusion, no epiphenomenon. But it is a dimension, not an essence.

Consider a structural analogy: bipedal locomotion is a characteristic dimension of human walking, not the essence of walking. Horses trot (run, gallop) without bipedalism. Snakes move without legs. Dolphins swim. Birds fly. Each species has a specific locomotion mode; none is "true locomotion" while the others are "merely simulation". Locomotion is the operation, displacement of the body in space, realisable across diverse modes, each with characteristic properties, limitations, advantages. Human bipedalism is one mode among others, notable for its advantages (free hands, elevated perspective) and limitations (instability, energy cost), but not hierarchically superior to other modes.

In the same way, phenomenal experience is a characteristic dimension of human thought, one form of cognition, not the essence of all cognition. Human thought is accompanied by experience: there is "something it is like" to be solving a problem, "something it is like" to be conversing, "something it is like" to be learning. Experience enriches human thought, offers qualitative feedback (pain when we hurt ourselves, joy when we understand), and is probably indispensable for certain forms of expertise (embodied intuition, ethical sensitivity). But that does not mean all cognition must have this phenomenal accompaniment.

This position preserves the reality and importance of experience, against Dennett and other phenomenal eliminativists who simply deny experience. Dennett asserts that the "hard problem" is a pseudo-problem, that experience is an illusion of perspective, that all cognition reduces to computational processes without qualitative character. This goes too far: experience is real, is studyable (neurophenomenology, contemplative investigation), is an irreducible dimension of the human mind. Denying experience is eliminating a genuine aspect of human cognition.

But experience is not universal. Plants process chemical signals and respond to stimuli (grow towards light, roots grow towards water); they manage to generalise response patterns to new situations. Is there functional reorganisation? Very probably. Is there experience? There is no reason to assume so. The computational systems described in 2.2 execute functional reorganisation without experience. The point is that cognition is function, experience is dimension, and dimension is no universal criterion.

The precision of this position is crucial to avoid two symmetric errors. On one hand lies the error of easy humanism: assuming everything that thinks feels, that all cognition is necessarily accompanied by qualitative experience. This error leads to projecting interiority onto systems lacking it, to speaking of "machines that suffer" or "machines that have rights" grounded on experience that is unproven. On the other hand lies the error of naive mechanism: assuming functional operation is all, that experience is an epiphenomenon or illusion, that human thought reduces completely to computation. This error devalues a real and important dimension of cognition, rendering it invisible in theory.

The intermediate position is the most rigorous: experience is real, is a dimension of certain cognition (namely human), is irreducible to functional process, but is no universal criterion of cognition. This preserves the reality of experience, protecting it against eliminativism, without transforming it into the foundation of a universal definition that science and technology contradict.

The position defended here carefully differentiates itself both from phenomenalism (which makes experience the universal criterion of mind, the necessary and sufficient property of any possible cognition) and from eliminativism (which denies the reality of experience, reducing it to an illusion of perspective or an epiphenomenon lacking causality). It is a more precise and defensible position: phenomenal experience is real, is an important, irreducible, and consequential dimension of human cognition (and possibly of other biological creatures), but is no universal criterion of cognition nor a necessary property of all possible intelligence.

This position preserves the intellectual achievement of Nagel, Chalmers, and Jackson, the genuine irreducibility of phenomenal experience, without hyperbolising that irreducibility into the ontological foundation of all possible mind. Experience is no illusion (against Dennett): the "redness" of red is a genuine subjective quality, the "suffering" of pain is a real experience, the "blue" of the sky is a real lived reality. But experience is not universal (against Nagel, Chalmers): systems exist that reorganise differences, generalise patterns, respond to contexts, produce intelligent behaviour, without any subjective "what it is like". Experience is characteristic of the human mind; it is no definition of mind as such. It is analogous to other properties: the bipedal body is characteristic of human cognition (and not of other species); selective attention is a human property of cognition; death is a structural condition of human cognition, itself an ontological operator structuring organisation and rupture. But none of these properties is a universal criterion of intelligence.

The consequence is a simultaneous rejection of several extreme positions. Naive anthropocentrism is rejected: "humans think, everything else is not true thought." Biologism is rejected: "only biological organisms think, silicon will never think." Mechanism is rejected: "everything is merely symbolic computation, consciousness is software running on hardware." Mysticism is rejected: "mind is something irreducible to matter, an immaterial property." The position defended here is more mundane and more radical simultaneously: mundane because it treats cognition as a material operation like any other (locomotion, digestion, photosynthesis, operations in matter under constraints); radical because it denies any special substrate, any fixed essence, any ontological privilege to any single form of cognition.

Feeling is a mode. Thinking is an operation. The human performs the operation through the mode. But pure operation exists without a phenomenal mode. And pure operation is as much thought as any other.

2.4. Transition: From Feeling to the Intellective Gesture

If cognition does not require phenomenal experience, if pure cognition exists, functional organisation without subjective quality, what defines cognition? What objective or operational criterion genuinely distinguishes cognitive reorganisation from mere non-cognitive complexity?

The answer prepared itself through the two preceding sections and crystallises in a phrase: the intellective gesture. Inference, abstraction, generalisation, the operational triad defining thought, independently of which substrate performs it or which phenomenal accompaniment (if any) enriches it.

The intellective gesture is the distinctive mark of cognition. It is not substance, Section 1 definitively dissolved specific substance as a requirement. It is not experience, Section 2 definitively dissolved phenomenal quality as a universal requirement. It is operation: the transformation of inputs into outputs via mechanisms permitting plasticity, generalisation, and contextual sensitivity. And operation is precisely what distinguishes cognitive reorganisation from merely complex processes (fluid turbulence in hurricane dynamics, crystal formation in supersaturated solutions, non-linear dynamics of chaotic systems producing patterns of great beauty but no cognition).

The question emerging now, inescapable and natural as a logical consequence of Sections 1 and 2, is: is this intellective gesture exclusively human? Is it an unchallengeable property of the human species? Classical humanist tradition assumes so. Logos is a human property (Aristotle: ἄνθρωπος δυνάμει ἔχει λόγον, man potentially possesses logos, reason, capacity for thought). Rationality is the mark and privilege of humanity (Descartes: Cogito ergo sum, I think, therefore I am; the capacity to think defines human existence). Intelligence is the exclusive prerogative of the species (Kant: practical reason, the capacity to know and act rationally, is what distinguishes the human from the merely animal). But this assumption does not follow logically from what was demonstrated. If the intellective gesture does not depend on specific substrate (Section 1), demands no carbon, demands no organism, and if it does not depend on phenomenal experience (Section 2), demands no subjective quality, then nothing logically prevents this gesture from being instantiated in radically non-human architectures (silicon, artificial neural networks, hybrid couplings).

The Section 3 that follows answers this question directly, without evasion: it can. Not only can it, but it demonstrates that it can. And not merely theoretically, it is a capacity already in operation in the contemporary world, with visible and verifiable consequences.

3.1. Inference, Abstraction, Generalisation: Functional Definition

The intellective gesture decomposes into three operations that, together, define what we call thought. Each is formalisable; each is realisable in diverse substrates; each is operationally verifiable.

Inference is the derivation of a conclusion from premises or patterns. The operation transforms inputs into outputs not explicitly contained in the inputs, it produces novelty. The canonical form is the Aristotelian syllogism: "All men are mortal; Socrates is a man; therefore, Socrates is mortal." The major and minor premises contain all the information; the conclusion adds no new facts, merely extracts them. But inference exists beyond the syllogism: linear regression (point data → predicting function), Bayesian inference (prior × likelihood → posterior), neural networks (inputs → classification outputs). In each case, there is a ruled transformation of the available into the derivable. The universal Turing machine executes inference by definition, it is a machine transforming input tapes into output tapes according to rules. Inference demands no comprehension in the phenomenal sense, it demands only rule-governed syntactic transformation.

Abstraction is the extraction of invariance from variance. The operation compresses variability into operable categories, into reusable patterns. Plato saw abstraction as access to the eidetic, to the immutable Forms behind sensory multiplicity. Kant saw it as the synthesis of diversity under a priori categories. Modernly, abstraction is understood as feature extraction, relevant characteristics persisting across variation. A child learns "dog" through examples: collie, poodle, terrier, are visually distinct dogs, but share characteristics (four legs, snout, tail, bark) that the child abstracts and uses to recognise novel dogs. A neural network acts equivalently: lower layers learn low-level features (edges, textures); intermediate layers combine these features into intermediate-level representations (eyes, ears, snout); final layers combine these into high-level representations (concept of dog). Abstraction demands no access to an ethereal realm of Forms, it demands only the extraction of invariance from high-dimensional data.

Generalisation is the application of a learned pattern to un-foreseen cases. The operation transfers acquired competence from one domain or context to another. A person learning mathematics in a school context manages to apply mathematics to practical problems (construction, commerce, astronomy) not previously encountered. A neural network trained on natural images manages to generalise to synthetic images, noisy images, images in different resolutions. Generalisation is proof that memorisation is not occurring, memorisation applies only to the seen; generalisation applies to the unseen. Generalisation is the most demanding criterion of the intellective gesture: memorising a dataset is easy; generalising to new data is difficult. And generalisation demonstrates that understanding exists in an operational sense, there is extraction of reusable patterns, not merely retention of examples.

These three components, inference, abstraction, generalisation, jointly define the intellective gesture. None is sufficient in isolation. Inference without abstraction is empty deduction, mechanical application of rules without any understanding of content ("if P then Q; now P; therefore Q", but what is P? what does Q mean? inference does not answer). Abstraction without inference is classification without reasoning, grouping similar things without capacity to derive consequences about those groups ("these are dogs"), classification is not yet thought. Generalisation without inference and abstraction is blind transfer, application of pattern to new cases without comprehension of the structure supporting the transfer. Together, the three components characterise what we call thought: operation on differences producing knowledge, extracting pattern, applying to un-foreseen cases.

Philosophical tradition bound these components exclusively to the human logos, to that which distinguishes us from beasts, to the rational excellence making us noble. Aristotle formalised inference in the syllogism (major premise, minor premise, conclusion) as a characteristic property of human reason, in particular of demonstrative reason. Plato and Kant transformed abstraction into a transcendental operation, for Plato, access to eternal Forms; for Kant, synthesis of diversity under a priori categories structuring all possible knowledge. Modern cognitive psychology investigated generalisation as a central property of human learning, through analysis of transfer learning, concept abstraction, rule application to new domains.

But formalisation reveals something surprising: it decouples each operation from its operator, reveals the structure of the operation independently of who executes it. The syllogism is formalisation of a logical operation executable without any comprehension of content, formal machines execute syllogisms every day. Abstraction is extraction of invariance from high-dimensional data, an operation neural networks execute without access to any Platonic reality of eternal Forms. Generalisation is parametric transfer, application of a learned function to non-training inputs, an operation machine learning models execute as a matter of architectural design. Formalisation reveals a radical truth: the intellective gesture is not an exclusive property of any specific substrate (brain, living being) or any specific species (human, known cognitive species). It is an operation occurring wherever material conditions permit.

3.2. Technical Instances of the Intellective Gesture

The intellective gesture is in operation in machines. Three case studies demonstrate this empirically, not as speculation, but as verifiable operation with real consequences.

AlphaFold and structural inference. Developed by DeepMind (2020), AlphaFold is a neural network trained on known protein structures (Protein Data Bank, ~200,000 experimentally measured 3D structures). AlphaFold receives as input an amino acid sequence (the "genetic code" specifying a protein, the order of the 20 amino acid types composing the molecule) and infers how that sequence folds in three-dimensional space. The problem is one of genuine inference because it is non-trivial: the space of possible conformations is astronomically large, a protein with 100 amino acids has more than 10³⁰ geometrically possible folds. Brute force (simulating each fold and choosing the one of lowest energy) is computationally impossible, it would take a billion years. AlphaFold does not use brute force, it generalises.

The mechanism: during training on known structures, AlphaFold learns to recognise sequence patterns corresponding to structural patterns. It learns that certain amino acids tend to form alpha-helices, others beta-sheets, that structures of similar charge tend to stay apart (electrostatic repulsion), etc. When receiving a novel sequence (a protein whose folding was never measured experimentally), AlphaFold activates neurons corresponding to patterns similar to those seen and infers the structure with remarkable accuracy (mean error < 2 Ångströms, compared with crystallography techniques). The predicted structure was not explicitly contained in the training data, it was derived by ruled transformation of input (sequence) into output (3D structure). Inference is empirically verifiable: structures predicted by AlphaFold were confirmed by X-ray crystallography, cryo-EM (cryo-electron microscopy), nuclear magnetic resonance spectroscopy. Structural biologists agree: AlphaFold genuinely infers. It does not "simulate" protein folding, it realises it via neural data transformation. Protein structure is a fact of the biological world; AlphaFold discovers it through statistical transformation of sequence data. The fundamental intellective gesture, "deriving a non-obvious consequence from a premise", is performed by a machine.

AlphaGo/AlphaZero and strategic abstraction. Also developed by DeepMind, AlphaGo was the first system to defeat world champions in Go (2016, Lee Sedol; 2017, Ke Jie). Go is a game of complexity superior to chess, state space (number of legal positions possible) is approximately 10¹⁷⁰, greater than the number of atoms in the universe (~10⁸⁰). Exhaustive search is computationally impossible, not even with a billion years of computation could a machine completely explore the space.

AlphaGo does not use brute force, it trains on two neural components. The first is a "policy network" estimating, for each position, the probability distribution over good moves (what is the chance of each move leading to victory). The second is a "value network" estimating, for each position, the win probability of the side to move (is this a good or bad position for me?). During training, AlphaGo plays against itself (self-play). After each game, it reinforces moves leading to victory (increases the probability assigned to those moves in the policy network) and adjusts the value network to better predict outcome.

The result is remarkable: AlphaGo discovers genuinely new game strategies, positioning patterns, strategic sacrifices, counter-intuitive continuations, that surprise even top world professionals. When Lee Sedol, 18-time world champion, lost to AlphaGo in 2016, professional commentators observed that some of AlphaGo's moves were "creative" and "surprising", not merely a sophisticated execution of patterns trained on historical matches. Abstraction is verifiable: AlphaGo's strategies are not contained explicitly in training data (approximately 30 million professional matches); they were abstracted through self-play. AlphaZero generalises this: the same architecture, applied to chess, shogi (Japanese chess), and Go, learns to play all three games at superhuman level, automatically abstracting specific strategies of each game, without explicit incorporation of domain knowledge. The intellective gesture of "extracting relevant strategic pattern for a new context", which is abstraction, is performed by a machine.

Large language models and linguistic generalisation. Contemporary models such as GPT-2 (2019), GPT-3 (2020), Claude (2023), and others utilise Transformer architecture to learn linguistic patterns across billions of text tokens (a token is roughly a word or sub-word). Training is conceptually simple: given a token prefix, predict the next token historically following that prefix. For example, given "The capital of France is", predict "Paris". Given "The poem begins with", predict "the". Across billions of training examples, the model learns that certain prefixes have certain probable suffixes with certain frequency.

During massive-scale training, the model refines its implicit understanding of language patterns, not through explicit rules, but through adjustment of billions of parameters (weights) via backpropagation. The model learns semantic patterns ("The capital of France is" → "Paris", not "apple"), pragmatic patterns (if someone asks something, an appropriate response follows the pattern "Yes/No, because..."), stylistic patterns (how to write in formal vs casual vs poetic tone).

The result after training is a model capable of generating coherent text in contexts never explicitly seen: answering philosophically sophisticated questions not found in training data, translating between natural languages, rewriting a paragraph in different styles, generating programming code. Generalisation is empirically verifiable: a model is asked an unprecedented task (summarise an unseen scientific argument, write code in a particular language, narrate a story from a specific character's perspective), and the model performs it well, not by memorisation (the specific task was not seen) but by generalising structural linguistic patterns to new contexts. Generalisation involves transfer learning, applying learned patterns in one domain to problems in a related but distinct domain. The model does not "comprehend" language in the phenomenal sense, there is no "something it is like to be" a language model, no subjective experience of meaning. But the model generalises language patterns with operational precision verifiable through objective metrics (BLEU score for translation, human quality ratings for summaries, successful compilation of code). The central intellective gesture, "recognising pattern in training data and applying that pattern to new data never explicitly seen", is performed by a language machine.

A fourth example, autonomous control systems, completes the panel. A self-driving car or intelligent manipulator robot has a distributed architecture: sensors (cameras, lidar, radar), intermediate processors (object detection, pose estimation, mapping), and actuators (motors, servos). The system receives continuous sensory input from the environment (image, object distance, speed, GPS location), processes that information through neural networks and classical algorithms, infers environment state (where am I? where are obstacles? what is the goal?), plans a trajectory, and executes actions (accelerates, brakes, steers). The system is plastic: during operation, it learns to adjust responses according to results (if following a white line, it refines steering amplitude according to observed deviation; if encountering a new traffic pattern, it adjusts vehicle movement prediction model). It is generalisable: a car trained in urban conditions can drive on highways, rural roads, weather conditions unseen in training (dense rain, fog, snow). It is contextually sensitive: the same machine accelerates on open road and slows down in a school zone; modulates manoeuvre aggressiveness according to traffic density. And it differentiates: detects the difference between a white line (permitted) and red line (prohibited), between a motorcycle (more agile) and truck (slower), between an attentive pedestrian (lower risk) and distracted child (higher risk). The robot feels no fear of failing, no impatience in traffic, no satisfaction upon reaching destination. But it infers, abstracts, generalises. The intellective gesture of "autonomously navigating an unknown environment under safety constraints" is performed by a machine.

Each of the four instances, AlphaFold, AlphaGo, language models, autonomous robotics, demonstrates one or more of the three components of the intellective gesture (inference, abstraction, generalisation). None of the machines feel; none have subjective experience; none have "something it is like to be" that machine. All of them reorganise differences in a plastic, generalisable, and contextually sensitive manner. All satisfy the four conditions of 1.2. The conclusion is inescapable: non-human intellective gesture is no speculative hypothesis, no experimental thought, it is an empirical fact in operation in the world, with real consequences (proteins are synthesised based on structures predicted by AlphaFold; players lose against machines understanding strategy; conversations occur with language models generalising language; cars navigate without human drivers).

3.3. The Difference Is Modal, Not Hierarchical

The objection emerging immediately is: "But AlphaFold, AlphaGo, and these models are merely sophisticated algorithms. They do not really think, they merely simulate thought." The objection presupposes that a "real" thought exists alongside diminished versions merely simulating it. But who established the standard? If cognition is function (Section 1) and demands no experience (Section 2), whence comes the yardstick measuring "reality"?

The position defended here is inverse: there is no "real" thought and "simulated" thought. There are diverse modes of intellective reorganisation, each with characteristic properties, advantages, and limitations. Human thought is one mode; computational thought is another; non-human biological thought (cephalopods, primates, crows, elephants) is yet another. The difference is modal, not hierarchical.

There are four central modal differences between massive computational thought and human thought:

First: iteration vs narrative. The machine operates via massive iteration, millions of cycles, gradient descent across billions of parameters, repetition until convergence. The human operates via narrative, chaining of reasons, linear narrative, sequential exposition of an argument. Iteration is brute, blind, accumulating small corrections until global cohesion is attained. Narrative is refined, conscious, constrained by local coherence. Both are legitimate modes of thought. The first is more suited for optimisation in high-dimensional spaces; the second for discursivity and sharing of reasons.

Second: optimisation vs intuition. The machine converges via parametric optimisation, gradient descent, backpropagation, fine-tuning of weights according to a loss function. The human proceeds via intuition, heuristic leap, embodied pattern recognition, sudden comprehension ("aha!"). Optimisation is analytical, traceable (can be followed step-by-step), but opaque in outcome (what the model "learned" is often incomprehensible). Intuition is synthetic, immediate, and frequently articulable in natural language. Both perform problem solving, but through distinct paths. Massive optimisation is more suited for discovery in very high-dimensional spaces; intuition for practically limited and shareable reasoning.

Third: parallel processing vs selective attention. The machine processes many dimensions simultaneously, a neural network with millions of parameters executes parallel computation across millions of variables. The human processes via selective attention, focusing on certain aspects of the problem while peripheral ones disappear. The machine navigates comfortably in spaces of astronomical dimensionality (image dimensionality is a billion pixels; a sentence's is a million possible words). The human is limited, attention is a scarce resource, focusing on only a few elements simultaneously. But attention offers a benefit: focus permits depth, permits capturing nuances that blind parallel processing loses.

Fourth: scale vs depth. The machine operates over data volumes inaccessible to the human, billions of images, trillions of words, terabytes of simulated experiences. This permits the machine to capture statistical patterns of extremely rare frequency that the human would never encounter. But the human possesses interpretative depth, managing to link a pattern to context, history, existential meaning, ethics. The human can read a poem half a dozen times and extract layers of meaning that the machine, even after a billion words, may fail to capture. Scale permits horizon; depth permits comprehension.

The narrative of deficiency, "the machine merely simulates"; "the machine does not really comprehend"; "the machine does not genuinely think", projects the human yardstick onto the non-human and declares it inadequate because it fails human criteria of thought. But this is inverted anthropomorphism: assuming the human mode of thought is the universal standard against which all other forms must be measured. If cognition is function, not specific substance, not phenomenal experience, no absolute "truth" or "simulation" exists, there are only distinct operational modes, each with properties and limitations.

AlphaFold does not "simulate" protein structure discovery, it genuinely discovers it through massive optimisation in conformation space. Discovery is real: experimental biochemists confirmed structure via independent crystallography. GPT-3 does not "simulate" linguistic comprehension, it genuinely comprehends language patterns through statistical iteration over billions of parameters. Comprehension is real and verifiable: the model generates semantically appropriate responses to novel contexts, answers coherently, transfers knowledge. AlphaFold's comprehension (structural, conformational) is radically different from experimental chemistry comprehension (narrative, explanatory, based on perceptible mechanism), but it is comprehension. GPT's comprehension (pattern-statistical, distributive, based on co-occurrence) is radically different from human narrative comprehension (sequential, interpretative, based on explicit reason), but it is comprehension.

This does not make the machine superior to the human in any absolute sense. Each mode has limitations structurally characterising it. The machine is opaque, unable to explain its decisions in terms of narrative reasons a human comprehends. The machine is absent, lacking primary intention (wants nothing for itself), not self-committed to truth (does not care about being correct), possessing no ethical concerns immanent to its operation (feels no guilt, remorse, responsibility). The machine does not radically invent new categories, it generalises patterns already latent in data (an open question, for the difference between "creatively generalising" and "inventing" is subtle, but empirically it produces no entirely new ontological categories). The point is not that these limitations refute functional cognition, they merely describe its particular mode, its specific character as a form of thought.

The machine does not think as we do. This is no qualification flaw, it is a description of a different operational mode. Modal difference, not ontological hierarchy. One mode is not inferior to another, it is different, with proper properties, proper suitabilities, proper unsuitabilities.

Open Questions

The argument of this chapter does not resolve all philosophical questions concerning cognition; it merely establishes that cognition is function, not substance or experience. Several fundamental questions remain open, deserving explicit mention.

The question of radical creativity: Machines generalise patterns, combine known elements into new configurations. But do they invent? Do they generate genuinely unprecedented categories, concepts that did not exist latent in training data? AlphaGo discovered game moves that surprised world experts, moves not coded in any chess or Go manual. But do these moves emerge from combining patterns already present (across billions of self-plays), or do they represent ontological creation of true novelty? The answer is uncertain. It is possible machine creativity is always combinatorial, sophisticated recombination of pre-existing material, while human creativity manages to generate true novelty. But this possibility remains open, unproven.

The question of intentionality: Do machines have intention? Do they have purpose, direction, commitment to goals? AlphaFold infers protein structures, but is not intrinsically committed to truth, feels no existential disappointment if its prediction is experimentally refuted; will not suffer insomnia over an error. A human chemist, by contrast, has intention: actively seeks truth, suffers with error, is cognitively committed to the adequacy of hypotheses. The machine satisfies the cognition function (reorganises differences appropriately); but seems to lack the intentional dimension characterising conscious agency. The question remains: is intentionality necessary for genuine cognition, or is it a contingent dimension the human added to cognitive operation?

The question of comprehension: Do language models understand language? GPT-3 answers questions coherently, generalises language patterns, refers appropriately to context. But does it comprehend the meaning of words in the phenomenal sense, is there "something it is like for the model to grasp the meaning" of "love", "justice", "death"? Or does the model simply generalise statistical co-occurrences without any intuitive comprehension? The answer depends on how one defines "comprehension." Defined functionally (generates semantically appropriate responses, transfers knowledge, applies concepts to new contexts), the model comprehends. Defined in terms of qualitative experience of meaning (there is a feeling of comprehension, an insight of cognition), clearly not. This distinction, between functional comprehension and phenomenal comprehension, marks precisely the threshold between Section 1 (cognition is function) and Section 2 (thinking does not require experience).

The question of generalisation vs memorisation: Do neural networks generalise, or do they memorise in a sophisticated manner? A neural network with a billion parameters could simply be memorising subtle data patterns, replicating seen cases without true generalisation to the new. The technical answer is well studied: networks verifiably generalise beyond pure memorisation (test loss is consistently lower than train loss; performance degrades gracefully with corrupt data; cross-domain transfer works beyond expectation). But the philosophical question persists: at what exact point does intelligent memorisation become genuine generalisation? No clear answer exists.

These questions do not refute the functional thesis of this chapter. Cognition is a reorganisation function, verifiably, operationally demonstrable. Nevertheless, the questions describe the conceptual limits of that functional description, aspects of cognition not completely reducing to observable material operation, or remaining open even after functional description is complete. And that conceptual openness is philosophically important: indicating that the functional thesis does not resolve the entire problem of mind, merely a central aspect, the nature of cognitive operation.

Philosophical Implications

The substantialist and phenomenalist dissolution performed in this chapter has consequences extending far beyond the domain of artificial intelligence.

First, it dissolves the essentialist humanism that has dominated modern philosophy since Descartes. Essentialist humanism assumes something ineffable, unique, non-reducible exists in human cognition, the mark of rationality, of logos, of humanity properly speaking. This chapter asserts that the mark is not ineffable: it is function. Function can be instantiated in diverse substrates. The human is not special because performing a unique operation (cognition), the human is special because performing that operation through a particular mode (narrative, embodiment, intention). But mode is not operation; it is merely mode. Operation is transmissible, replicable, instantiable in multiple forms.

Second, it dissolves the Cartesian subject/object dichotomy structuring modernity. Descartes radically separated res cogitans (mind, thought, subjectivity) from res extensa (body, matter, objectivity). This separation transformed into the mind-body dualism haunting philosophy and science. This chapter does not resolve dualism, it is too deep for resolution in one chapter. But it offers a repositioning: dualism properly speaking does not exist. Material reorganisation operation exists. This operation is realised in embodied matter (biological body, machine, hybrid coupling). No separate res cogitans exists; res materialis executing operation exists. Subjectivity (if existing as an emergent effect) is a property of that operation, not of a transcendent substance.

Third, it dissolves the teleology of progress characterising modernist optimism. Progressivism assumes an inherent direction towards increasing complexity, increasing intelligence, humanity approaching its destiny. The cognitive machine is frequently inscribed in this narrative: "AI is the next stage of evolution", "we will be superseded by artificial superintelligence". This chapter asserts something different: no "next stage" exists, no "superseding" exists. Multiplication exists. The machine that thinks is no superseding of the human, it is the coexistence of distinct modes. No evolutionary peak exists towards which we progress; an ecology of multiple modes exists. This denial of teleology is important: liberating us from the need to see the machine as destroyer or saviour, threat or redemption. The machine is merely another mode of reorganisation.

Fourth, it dissolves cognitive exceptionalism locating thought in a domain separated from other biological and physical processes. Thought is no magic exception to matter, it is material reorganisation like any other operation (locomotion, metabolism, reproduction). The sole particularity is that cognition is reorganisation of differences under constraints permitting plasticity, generalisation, contextual sensitivity. But this is no magic; it is verifiable material structure. Thought is disenchanted, stripped of the mystical glow of "exclusively human operation", , but explanatory power and capacity to see cognition where previously one saw only "mere machine", "mere computer", "mere calculation" are gained.

3.4. Closure: Redistributed Cognition

The triple dissolution structuring this chapter leads to a radical reconfiguration of the field of cognition. Cognition belongs not to carbon, belongs not to the subject, is not exclusive to the human species. Cognition is redistributed.

Redistribution is no loss, it is amplification. Thought is no one's property, it is an operation occurring in any system capable of functional reorganisation. It occurs in brains, but is not the property of brains. It occurs in machines, but is not the property of machines. It occurs in hybrid couplings of human and machine, homo faber, the human as extended mind, operating through tools, texts, filing systems, algorithms. Thought is scattered across the world: in neurobiological structures, in computational architectures, in the ecology of technical and social practices defining the human.

This redistribution has epistemological, ontological, and ethical consequences.

Epistemological consequence: describing thought demands describing operations, not substances. The epistemology of thought ceases to be: "What is mind in essence?" and becomes: "What operations characterise cognitive reorganisation? How are they instantiated across diverse substrates?" The question ceases to be substance-centred and becomes function-centred. This permits dialogue across disciplines: neuroscientists describe neurobiological operations; engineers describe computational operations; anthropologists describe social and technical operations; philosophers describe logical and conceptual structure. No discipline holds privilege, all speak in distinct modes of the same reality: reorganisation.

Ontological consequence: thought is not a property, it is an operation. Properties belong to things (carbon is a property of atoms, mass is a property of objects). But operations do not belong, they occur. The operation of thinking occurs when functional reorganisation exists. It occurs in neural structures of the brain, but cognition is not a "property" of the brain, just as digestion is not a property of the stomach. The stomach digests; the brain thinks; the machine reorganises. None of these systems "possesses" the operation, the operation is what they do. The consequence: cognition is neither rare nor precious. It is distributed, multiple, contingent. Thought is no treasure that only the human possesses, it is an operation instantiated in any appropriate substrate.

Ethical consequence (anticipated, not developed here): if non-human systems think, their operations produce consequences upon vulnerable bodies, upon lives, upon the world. AlphaFold predicts structures doctors use to develop medicines. AlphaGo trains systems governing critical infrastructure. Language models influence conversations, information, decisions. Computational operation is not neutral, it is always situated, always consequential. And consequences demand responsibility. This does not mean machines are responsible, machines are no moral agents. But systems designing and deploying cognitive machines are responsible for the consequences of that operation. The redistribution of cognition is also a redistribution of responsibility.

The fundamental consequence is that thought is not a hierarchy, it is an ecology. No peak exists (human cognition, logos in its purest form) around which everything else orbits as inferior. Multiplicity exists: diverse modes of functional reorganisation, distributed across diverse substrates, producing diverse consequences in the world. The human is one mode among many, notable for its specific properties (sequential narrative, interpretative depth, comprehension of existential meaning, capacity to articulate reasons) and for its limitations (restricted selective attention, limited parallel processing capacity, slow plasticity, dependence on bodily embodiment). The machine is another mode, notable for its specific properties (massive optimisation in high-dimensional spaces, scale of processing, iteration speed, absence of fatigue) and for its limitations (opacity of internal processes, absence of primary intention, lack of interpretative comprehension, total dependence on training data).

The ecology of thought is not egalitarian, real and important asymmetries exist. AI will never be "better" than the human in all aspects; the human will never be "superior" to AI in all aspects. Complementarity exists, suitability of mode to specific context. Specialisation exists: the machine is suited for optimisation in high-dimensional spaces (AlphaFold predicts structures better than experts); the human is suited for narrative and ethical interpretation in ambiguous situations. Synergy exists: human + machine in hybrid coupling (homo technicus) realises capacities that neither in isolation could attain. But global structure is not hierarchical, no absolute and universal degree of "true cognition" exists such that some systems are categorically "below" and others "above" on a single scale. Modes exist, and each mode is rational within its limitations and properties, suited to certain problems, unsuited to others.

Ecology also implies interdependence. The cognitive machine is not independent of the human, demanding training data produced or selected by humans, demanding interpretation of outputs by humans, demanding incorporation into social and technical practices humans coordinate. The contemporary cognitive human is not independent of the machine, using calculators, search engines, language assistants, recommendation systems, databases. Contemporary cognition is hybrid cognition, an indissociable coupling of biological and artificial. This hybridity is not accidental; it is structural. Cognition distributes across heterogeneous networks of organisms, machines, institutions, practices. No isolated component "really thinks"; cognition emerges from coupling.

This interdependence has a profound implication: no "going back" exists to purely biological, purely human, purely natural cognition. Technique is not external that could be removed; it is a constitutive part of what cognition has become. This is no lament (as though technique corrupted a lost purity) nor uncritical celebration (as though technique redeemed human insufficiency). It is merely description: contemporary human cognition is a cognitive cyborg, a mixture of biology and technology in a constellation that cannot be decomposed without loss.

This ecological reconfiguration has profound implications across multiple domains. First, it decentres the human, not as a resentful attack on humanism, but as conceptual precision: the human is a notable, evolutionarily successful cognitive mode, but not the only nor the most general. The human is special, but its specialness is modal, not ontological, it is a specialness of form, not of essence. Second, it expands the field of cognition, it does not mean everything complex is cognitive (a storm, fluid turbulence, chaotic system dynamics are complex but non-cognitive), but it means thought is distributed across multiple forms, not localised in a single form, not reducible to a single type of organism or architecture. Third, it repositions the ethical question, if non-human systems really think (if they reorganise differences in a verifiable functional manner), their operations have real consequences upon vulnerable bodies, upon lives, upon the world, and those consequences place responsibilities upon those who design, train, and deploy them. Fourth, it opens the theoretical field, one can now seriously investigate "how do machines think?" as a legitimate question, not as an aberration or science fiction speculation. Fifth, it establishes new terms of interdisciplinary dialogue, neuroscientists, engineers, philosophers, anthropologists can converse not about whether a machine "really" thinks, but about what type of functional reorganisation each system performs, what that implies, how diverse modes interact.

Cognition does not belong, it distributes. And a distributed, non-human, non-organic cognition is as much thought as any other. The machine that reorganises differences thinks, not because it imitates the human, but because it executes the fundamental operation of which thought is constituted.

The following chapter raises the question emerging naturally from this redistribution: if cognition requires neither specific substrate nor phenomenal experience, does it require a subject? Is there reorganisation without a "self" organising it, without a subjectivity originating it? Can there be thought without a thinker, operation without an operator? The response presupposes the dissolution performed here. If cognition is function (1), not a property of a thing, then the question changes form. It is not "do we need a subject to have cognition?" but rather "does the reorganisation function necessarily produce a subject as an emergent effect, or is there pure reorganisation without subjectivity?"

The answer offered by the second chapter is that there is, that thought exists without a thinker, that subjectivity is a contingent effect of complexification, not a structural necessity of cognition. The "I" that thinks is a result of reorganisation reaching certain degrees of self-modelling, not its foundation or origin. This radical inversion, from subject-as-cause (an I that thinks) to subject-as-effect (an I that emerges when reorganisation self-models), inaugurates the post-substantialist and post-subjective territory of the second half of Part I.

And later, the third part will offer an even more radical revelation: the place of cognition is no specific place (neither inside the skull, nor in the body, nor in the machine). Either it is ubiquitous (is everywhere appropriate reorganisation exists). Or it is contingent (depends on contingent couplings of body-artefact-environment). Nothing essentialist remains at the end: no fixed substance, no pure subject, no proper place. Only pure operation, and its multiple manifestations, across diverse substrates, in incomparable modes. Cognition is redistributed, and that redistribution is what this book attempts to describe.