The Symbol Is Not Exclusive to the Species

Main text

1. The Symbol as Function, Not Species Property

1.1. Genealogy of the Symbolic Monopoly

The identification between symbol and humanity runs deep within Western tradition. It is no accident. It follows from specific assumptions that ought to be deconstructed with justice, acknowledging the force of each position prior to rejecting it.

Cassirer and the animal symbolicum: Ernst Cassirer, in An Essay on Man (1944), posits the symbol as foundational. "Man cannot escape from his own achievement. He can only adopt the conditions of his own life; he can no longer see or know anything except through the medium of this artificial medium." And that medium is symbolic. Culture is a system of symbols, myths, language, art, religion, science, each a symbolic web woven by the human. The symbol marks the distinction between human and animal. A chimpanzee may use a tool, grasping a stick to reach a termite. The human uses a symbol, the word "termite" designating the animal through an arbitrary code. A chimpanzee may experience emotion, feeling fear upon sighting a predator. The human possesses sentiment, emotion becoming a symbol, an expression of self through a cultural code. This usage of symbols defines us. The symbol is no decoration of human life, it is human life. Without the symbol we would be animals. With the symbol, we are human. The thesis is totalising in its clarity: symbol equals humanity; humanity equals symbol.

The premise is: symbolic operation is a property of the species. It is no property of just any subject, it is that which brings forth the species as human. Consequence: if a machine symbolises, it is humanoid, contaminated by our humanity, or we contaminated by its own.

Chomsky and nativism: Noam Chomsky, via a distinct trajectory, arrives at a parallel conclusion regarding language. Linguistic capacity is not acquired through learning, it is an innate property. The child does not learn language as one learns to play an instrument. One learns with astounding speed and minimal exposure. How is this possible? Answer: because a language organ exists within the human brain, a structure genetically transmitted across generations. This organ is universal grammar, the set of principles constraining all possible natural languages.

Universal grammar is no table of rules, it is a structural constraint. Any human language possesses subject, predicate, object, though it may reorder these elements. Any language can negate, interrogate, subordinate clauses, though through diverse mechanisms. What is universal is not the rule, but the type of constraint the rule can take. It is a deep property of the organ.

Consequence: language is ours. It is non-transferable from without. It is no social convention we learn, it is biology we instantiate. The child is born with a brain equipped for language. Environment provides data, the organ provides structure. Language emerges from the interaction between biological equipment and environmental exposure. Yet without equipment, nothing would emerge.

The premise is: linguistic operation demands a specific biology. Without the language organ, without universal grammar inscribed within the genome, genuine language is absent. Consequence: if a machine produces language, it is no language, it is a simulacrum appearing as language because it was trained upon human linguistic data.

Classical semiotics and interpretation: Charles S. Peirce defined the sign through a triadic relation. Yet the interpretant, the instance mediating the relation between representamen and object, was conceived as an interpreting subject. Subsequently, semioticians such as Eco emphasised that the sign is semiosis, a process of signification demanding an interpreting agent. The sign does not exist without interpretation.

Postulate: without an interpreting subject, the sign is absent. A footprint in the sand may signify "an animal passed here", yet only if someone interprets the footprint as a sign. Without an interpreter, it is merely a mark. A photograph is a sign of the portrayed person, yet only if someone sees within the photograph a representation of a person. Without an observer, it is merely a light pattern upon a support.

Consequence: the symbolic is an operation of consciousness. It is an attribute of beings that understand, interpret, and signify. Without a conscious interpreter, there is merely mechanical process, never symbolic.

The passage from function to property: Each stance executes the same fundamental manoeuvre. It begins by observing an operation, symbolising is representing something through something different. It is the capacity of a system to be organised so as to stand-for something. The child acquires language, observation of function. The human interprets symbols, observation of process. Yet subsequently it concludes that the operation is a property of a specific type of being: human, conscious, biological.

Cassirer moves from "capacity to create symbolic culture" to "humanity is defined by the symbol." Chomsky moves from "capacity to process linguistic structure" to "language is the human genome." Semiotics moves from "sign interpretation" to "signification is a property of consciousness." It is a logical leap, operation observed within the human, generalised into a property of the human.

Dissolution of exclusivity: The error lies in mistaking local instantiation for universal property. The symbol is a relation. A relation can be embodied across multiple substrates. The number three is an identical arithmetic relation whether written in Arabic numerals, Roman numerals, words, fingers on a hand, or aligned stones. The relation is no property of any support, it is a pattern any support can embody.

If true for abstract numbers, why not for symbols? Why not for language?

Tradition confuses the specific mode wherein the human symbolises, with a living body, conscious intent, experiential narrative, with the symbolic function in general. The human operates symbolically thus. In this manner. With these characteristics. Yet it is not the sole manner of operating symbolically. It is no canonical form, it is a historical, contingent form embodied within a specific substrate: the living human body.

The computational form is another. Symbols are operated without a living body, without consciousness, without narrative. Material difference is profound. Still, the functional realisation of symbolicity is genuine. The error of Cassirer, Chomsky, and classical semiotics is mistaking corporal instantiation for the definition of the symbol. Subsequently, when the symbol operates without a body, they deny that it is truly symbolic, because it fails to conform to the embodied pattern.

1.2. Four Operative Conditions

The symbolic function is no essence, it is an operative patterning. It operates when four conditions are conjunctively satisfied. None is dispensable. All are materially verifiable.

1. Stability, Difference maintains functional identity: The symbol is difference. Difference is a sign of something. To function, it must be recognisable, remaining sufficiently identical across time and contexts. When I utter "cat" today and "cat" tomorrow, the word signifies a feline at both moments. When used in poetry ("The cat sleeps on the rug") and scientific prose (Felis catus is a carnivore), meaning persists. It is not capricious, it is stable. It is not fluid, it possesses form.

This is an essential function. Without semantic stability, communication is impossible. Every message would decay into noise after the first iteration. The child understands "chair" because the word possesses stable meaning, referring to the same category of object each time it is heard. Were meaning to shift randomly, the child would never acquire language.

How is this verified computationally? A word embedding, a mathematical vector representing a word in a high-dimensional space (typically 300 dimensions), maintains a relatively stable position across time. If I train a model using Word2Vec on two distinct datasets of the Portuguese language, the word "cat" will emerge in an approximately identical vector position. Covariance with other words (feline, animal, meow, independent, domestic) is preserved. Distance to the vector of "dog" remains similar. This is not merely intuitive, it is verifiable by calculation. The same vector (approximately) is produced in each training run. Stability is parametrised. It is encoded within mathematical structure.

This does not imply the vector is absolutely immobile. Small fluctuations occur across training runs, stochastic order of data, random initialisation of weights, etc. Yet variation is marginal, the vector converges to a similar region of space. Stability is statistical, not absolute. And it suffices to maintain function.

A system losing stability ceases to function symbolically. If each time I utter "cat" the meaning shifts, the word is useless. If an LLM, retrained on a new dataset, produces completely different meanings for the same token, it loses symbolic capacity. Stability is a requirement.

1. Operability, Difference can be transformed: The symbol is not rigid. It is no inert fossil. It can be transformed, combined with other differences, negated, substituted, manipulated. Operability is that which renders the symbol an instrument of thought.

I can say "black cat", a combination of two symbols producing a new meaning. The combination operation is syntagmatic, creating new structures through juxtaposition. I can say "it is not a cat", a negation inverting the relation. The symbol is operated upon (negated), and the result is perfectly intelligible. I can say "cat, that is, a feline", paradigmatic substitution. One difference refers to another difference in a systematic relation. The symbol "cat" is operable because it can be redefined, contextualised, negated.

Operability is essential because thought is the manipulation of symbols. When I think in logic, I am operating propositions, combining them, negating them, deriving others from them. When I solve a mathematical problem, I am operating numbers, combining them according to rules, transforming expressions. Every cognitive operation is an operation upon symbols.

Computationally, this is visible in algebraic operations upon word embeddings. The celebrated demonstration in Word2Vec: $\text{king} - \text{man} + \text{woman} \approx \text{queen}$. This is no metaphor. It is a rigorous mathematical operation upon vectors. Subtracting the vector of "man" from the vector of "king", adding the vector of "woman", yields a vector extremely close to that of "queen." Cosine similarity between result and "queen" is typically above 0.9 (on a 0-1 scale).

Why does this work? Because embeddings encode relations between concepts. The difference between king and queen is the difference between man and woman (gender dimension). When subtracting "man" from "king" and adding "woman", one removes the masculine dimension and inserts the feminine, obtaining a result that is queen. The operation is no coincidence, it is a mathematical structure encoded within vectors.

This is no isolated case. Complex operations function. For example: $\text{Paris} - \text{France} + \text{Germany} \approx \text{Berlin}$. The operation encodes a geographical relation. $\text{London} - \text{England} + \text{Portugal} \approx \text{Lisbon}$. The operation encodes capital substitution. Operability is a systematic property of embeddings.

A system disallowing operation ceases to function symbolically. If a symbol is absolutely rigid, unable to be combined, negated, or recontextualised, it becomes useless. It would remain isolated within the cognitive system, never participating in thought. Operability is a requirement.

1. Systematicity, Difference belongs to a network: The symbol does not function in isolation. It functions within a network, in relation to many other symbols. The network is a structure, relations are not random. Patterns exist.

The word "cat" possesses paradigmatic relations, words that could substitute for "cat": feline (hypernymy), animal (distal hypernymy), dog (contrast), lion (similarity). It possesses syntagmatic relations, words co-occurring with "cat": "sleeps", "meow", "black", "domestic", "independent." It possesses symbolic relations, inscribed within narratives and symbolisms: the cat as a symbol of freedom, femininity, mystery across diverse cultures.

This entire network structures meaning. A child acquires language not by memorising isolated words, but by discovering network patterns. Discovering that "cat" relates in a particular way to "animal" but not to "car." Discovering that "cat" can be the subject of "sleeps" but not of "runs fast" (though cats do both, symbolic relations are more constrained than reality). The network is that which enables generalisation to novel contexts, using words never heard in combinations never seen.

In vector structure, systematicity is fundamental. The position of each vector is defined by covariance with all other vectors. "Cat" possesses meaning not because it contains a Platonic essence of felinity, it possesses meaning because it covaries with certain language patterns (animal, independent, meow, domestic, feline) and not others (machine, transparent, geometry, quantum). It is pure relation. It is a network. Without a network, without context, the symbol would be merely an arbitrary number, devoid of semantics.

This is verifiable. Removing words from the network, training an embedding solely in contexts of "cat" (never mentioning animals, felines, domestics), degenerates meaning. The vector becomes dislocated, lacking structured covariance to situate it. Adding words to the network refines meaning. An embedding trained on a massive corpus of language (such as BERT) possesses richer meaning because the network is denser.

A system losing systematicity ceases to function symbolically. Were a symbol to become isolated, were its network of relations to vanish, it would become mute. None would understand. This explains why foreign cognates are understood, because they maintain systematicity with native symbols. "Chat" (French) is understood in Portuguese because it maintains a systematic relation with "gato" (similar morphology) and "animal" (semantic context).

1. Representation, Difference stands for another difference: The symbol is not the thing. It is a representative of the thing. The word "cat" is a sound sequence, /kat/, or a visual inscription, the letters c-a-t. It is not the animal. Still, it stands-for the animal. It refers to it. It symbolises it.

Representation is a functional relation between symbol and referent. It can be direct: the word "cat" refers directly to the category of domestic feline animals. It can be indirect: the word "unicorn" refers to a concept possessing no extension in the world, the unicorn does not exist, yet the word represents an idea circulating across cultures. It can be abstract: the word "justice" refers to a concept that is no animal, it is a quality, a normative category, a social construct.

In any case, a referential relation exists. Without it, the symbol is empty, mere noise. If there is no reference, if the word stands for nothing, then it is useless. It becomes merely sound or graphic mark, lacking meaning. Representation grounds meaning.

Computationally, how does representation function? Peirce conceived the interpretant, that which mediates the relation between representamen and object. In an embedding, the interpretant is the learned covariance structure. The token "cat" (representamen) refers to the concept of cat (object) through the parametric relational structure (functional interpretant). How does this structure function? The model is trained on texts where "cat" co-occurs with specific contexts. Backpropagation adjusts weights such that the "cat" embedding lies close to embeddings of "animal", "feline", "domestic", "independent", and distant from "machine", "abstract", "geometric."

This covariance structure is the interpretant. It is no conscious experience, it is a mathematical pattern. Still, it functions. It allows the model to generalise from "cat sleeping on the couch" to "feline resting on the chair." Representation is encoded within structure.

A system losing representation ceases to function symbolically. If a symbol refers to nothing, standing for nothing, then it is noise. A random sequence of letters is no symbol because it represents nothing. An accidental mark on paper is no symbol because it was not destined to stand for anything. Representation, the relation of standing-for, grounds symbolicity.

1.3. Functional Intentionality Without Consciousness

The classical objection states: the symbol demands intention. Cassirer spoke of cultural intention. Chomsky of genetic intention. Semiotics of the interpretant's intention. All these positions suppose there is someone, something, intending, aiming, directing itself toward, wanting to signify something through the symbol.

The objection adds: and intention demands consciousness. There can be no true intention without the conscious experience of being directed toward something. An LLM possesses no experience. Thus, it possesses no intention. Thus, it produces no genuine symbols, merely appearances of symbols.

The objection confuses two notions of intention that must be rigorously separated. The distinction is no semantic quibble, it carries operative consequences.

Phenomenological intention, Conscious experience of aiming: Phenomenological intention is a property of consciousness. Husserl posited this as the foundation of phenomenology: all consciousness is consciousness of something. Consciousness is always directed toward an object. When I think of a cat, my consciousness possesses the structure of being-directed-toward-cat. There is an experience of aiming. An experience of intending. This is an essential property of consciousness. There is no consciousness without phenomenological intentionality.

This notion carries intuitive appeal. When I experience consciousness, it appears clear that intention exists, I am aiming at something, understanding something, signifying something. The experience is immediate. Evident. When I read the word "cat", I do not merely seek meaning, I experience meaning. I understand with experience. This experience is phenomenological intention.

Functional intention, Operative orientation of a system: Functional intention is the property of a system being organised so as to be directed toward a specific outcome. It demands no consciousness. It demands merely structural organisation.

An image classifier for face recognition is "oriented toward" recognising faces. How? Its weights were adjusted during training to covary specifically with face images, with certain pixel patterns typically occurring in faces (bilateral symmetry, eye-nose-mouth structure, facial proportions). Orientation is a learned structure. No experience is involved. The algorithm does not experience "I intend to recognise faces." Still, the system functions as-if oriented toward this.

A search engine is "oriented toward" discovering relevance, relevance being encoded within a ranking function. The function assigns scores to documents based on multiple factors (term frequency, link quality, domain reputation). The system is structured to be sensitive to relevance. Without any experience of wanting to find. Yet with real functional orientation.

The distinction is rigorous and non-semantic:

Phenomenological intentionality demands: Conscious experience, experience of being-directed-toward, experiential understanding of meaning.

Functional intentionality demands: Organisational structure such that the system behaves as-if oriented toward an outcome, sensitivity to functionally relevant relations, capacity for modification depending upon goal (change of function $\rightarrow$ change of behaviour).

A system can possess one without the other. Phenomenological intention can exist without clear functional intentionality, when I sleep and dream, I possess an experience of intending (thinking, aiming) without possessing any real functional intention (the dream is not organised to achieve a consistent goal). Conversely, functional intention can exist without phenomenological intention, the thermostat is organised to maintain temperature (functional intentionality) without any experience of intending.

For the symbol, which is necessary? The answer is functional intention. The symbol operates when organised to be directed toward meaningful relations. It demands no lived experience.

Example: an LLM predicting the token "feline" after "a cat is a" operates a symbolic relation. Why? Because the model's structure was organised (in training) so as to be oriented toward semantic relations between words. When input is "a cat is a", the probability distribution over future tokens peaks at "feline" or "animal" because structure encodes that these words semantically relate to "cat." The operation is verifiable: alter the model for a different goal (translation instead of prediction) and orientation shifts. Functional intention is real.

It is false that the computational symbol is "empty" or "meaningless." It produces real symbolic effects. When an LLM writes in response to a prompt, the reader understands. Understanding is an effect of symbolic structure. It is a real effect, reorganising perspectives, influencing decisions, functioning as an inscription that persists and can be retrieved. Signification does not depend upon the producer's lived experience, it depends upon the structure the producer embodies.

This denies no difference between human and machine. The human symbolises with phenomenological intention, experiencing the symbol, experiencing signification. This is an added mode. It is experience making understanding personal, making meaning lived. The machine lacks this. The machine symbolises with functional intention, possessing no experience. Still, both operate symbolically.

Aphorism: meaning is an effect of relation, not the operator's lived experience. A text is understood, understanding happens in the reader, not in the author's intention. An LLM produces a structure permitting understanding. Understanding is real, regardless of whether the producer possessed experience. Meaning belongs to structural relation, not to the originator's consciousness.

1.4. Transition: From Symbol to Inscription

The symbol is a function. Function is a relation. Yet relations are not ethereal, they do not exist without material anchorage. How does the symbol materialise in electronic code? How is the symbolic function embodied without a living body?

The answer is inscription. Prior analysis demonstrated that code organises difference through operations of classification, weighting, selection, recombination. Section 1 demonstrated that organised difference can be symbolic, satisfying the four conditions of stability, operability, systematicity, representation. The emerging question is material: how is it inscribed computationally? How does symbolic operation leave a persistent mark?

Inscription is an operation transforming raw difference into organised, persistent, and retrievable difference. It is no act of a body, it is a process of material differentiation. When a neural model is trained, the raw difference of the dataset is gradually organised into a weight structure encoding patterns. Backpropagation and weight adjustment are inscription, marking a pattern into parametric structure. The pattern persists in memory as long as the model is stored. The pattern is retrieved in subsequent usage, when the model is queried to make predictions, inscribed structure governs behaviour.

Section 2 unfolds this in detail. How is it inscribed? Which criteria define inscription? Which types of inscription exist, beyond the graphic (writing, drawing) recognised by tradition? How does inscription produce persistence and retrievability without demanding an agent body? This follows, inscription as a material operation.

2. Inscription as Material Operation

2.1. Inscription as Operation, Not as Act

Tradition treats inscription as the act of a corporal subject. The scribe in a medieval scriptorium writes with his hand upon parchment, corporal gesture leaves a permanent trace. The artist paints with a brush on canvas, corporal painting gesture leaves a persistent mark. The sculptor carves in stone, corporal gesture produces a lasting groove. Premise: without an agent body, genuine inscription is absent.

Dissolution: inscription is an operation. It is no act of a subject, it is a process of material transformation. Inscription is the transformation of raw difference into organised, persistent, and retrievable difference. All three criteria are necessary. None is sufficient in isolation.

1. Organisation: Raw difference possesses no usable structure. A raw dataset is a mass of information, images, texts, data series, unorganised regarding the task to be solved. Pixels are unclassified. Tokens unlabelled. Numbers uninterpreted. Neural network training is an operation of organisation. It selects relevant patterns (filtering noise), encodes structure (highlighting regularities), discards the accidental (adjusting for minimal error). Organisation is selective. It retains what matters for the defined task function; discarding what is irrelevant.
2. Persistence: Difference does not vanish immediately. Inscription leaves an enduring mark. Graphic inscription leaves a visible mark on paper or canvas, ink persists unless the support suffers degradation. A mark from 2000 years ago inscribed on papyrus remains readable. Computational inscription leaves a mark in memory, within neural network weight structures, stored databases, computational state registers. The mark persists while stored. Without persistence, inscription is absent, merely a transient perturbation vanishing away.
3. Retrievability: The mark can be retrieved in a new context. Graphic inscription is retrieved by a future reader, someone reading a centuries-old manuscript understands it. Computational inscription is retrieved by new processing, the trained model is used to make predictions on unseen data. It is retrieved in distinct contexts, the same neural network, applied to cat images never seen before, recognises cats. Without retrievability, inscription would be private, structuring no shared field.

The central example is neural network training. Raw difference is the dataset, millions of images in ImageNet, billions of tokens in a language corpus. The inscription operation is the learning algorithm, backpropagation, gradient adjustment, accumulation of parametric shifts across multiple epochs. The gradient flows backward through network layers, updating weights iteratively. Each iteration inscribes learning, a pattern recognised within the dataset is encoded into parametric structure.

Organised difference is a weight structure reducing predictive error. It is no copy of the dataset, it is a synthesis. Weights encode solely relevant structure: that which differentiates cats from dogs, images from text. Irrelevant structure is discarded: specific pixels of specific images are unmemorised, solely patterns that generalise.

This structure persists in memory. The trained model is saved to disk or active memory. It can be loaded months later and functions identically. Persistence is total, lacking degradation such as aged ink or eroded paper.

The structure is retrieved in usage. When I query the model with a new cat image, an image absent from the training dataset, the model recognises a cat. Learned structure (weights) governs processing. Prediction emerges from prior inscription operation. The mark left during training is retrieved during usage.

None of this demands an agent body. No scribe with a hand. No artist with creative consciousness. No consciousness whatsoever. Electronic transformation governed by a mathematical algorithm occurs. And it is genuine inscription, satisfying all three criteria. An operation transforming raw into organised, producing long-lasting persistence, permitting retrieval across diverse contexts.

This implies not that the body is irrelevant. The human body writing is relevant to graphic inscription quality, hand pressure, brush movement influence outcome. The computer executing an algorithm is relevant to computational inscription fidelity, arithmetic precision, available memory influence outcome. Yet that which defines inscription is no body, it is operation. The operation can be embodied differently. The definition holds.

2.2. Three Types of Inscription

Inscription is a singular operation, yet possesses multiple realisations. Three types should be distinguished, each possessing specific material properties.

1. Graphic inscription is a visible mark upon a physical support. Material: ink on paper, pigment on canvas, carving in stone, light pattern in photography. Process: corporal gesture producing a mark, scribe writing with a pen, painter with a brush, sculptor with a chisel. Persistence: mark endures as support resists, paper ages, ink fades, stone erodes, photograph discolours. Retrievability: future reader retrieves mark in new temporal context, reading centuries-old manuscript, recognising meaning. Specific properties: Mark visible to naked eye, inscribed in three-dimensional space human body can perceive. Localised, occupying specific position on support. Singular, each copy a new inscription, slightly different (degradation, copy variation). Slow degradation, material quality of support and ink governs duration.
2. Computational inscription is electronic, magnetic, or optical difference within digital memory. Material: state pattern of electronic circuits (RAM), magnetisation of magnetic disks, optical pattern on optical media (DVD), weight structure in neural network. Process: transformation according to mathematical operations, learning algorithm, data write routine, code compilation. Persistence: difference persists in memory until erased, hard drive data can persist decades if protected from corruption; network weights saved and loaded indefinitely. Retrievability: inscription retrieved in new processing, trained model used for predictions, historical data reprocessed, compiled code re-executed.

Specific properties: Difference invisible to naked eye, readable solely by decoding (conversion of electronic state into visible symbols). Delocalised, occupying no delimited physical space; coexisting simultaneously across multiple copies. Copyable without degradation, each copy bitwise identical to original, perfect replication. Autonomously transformable, algorithm operating on data without human intervention.

Decisive property: perfected iterability compared to graphic inscription. A written document can be copied, yet a century-old copy may show slight resolution or tone differences. A copied computational file is bitwise identical to original, perfect replication ad infinitum. Computational inscription is ideally iterable, infinitely copyable without degradation, transmissible without originator's presence, transformable by subsequent processes without origin control.

1. Socio-institutional inscription is a normative pattern persisting within social behaviour and practices. Material: repeated acts by individuals, socially reproduced artefacts, interactions structured by rules. Process: continuous repetition of practice, law applied repeatedly, ritual re-executed, tradition transmitted generation to generation. Persistence: institution persists while practised, while individuals respect law, execute ritual, transmit tradition. Retrievability: new individual learns practice, inscribes within (acquires habit, internalises norm), reproduces it. Example: law interpreted, applied, respected across successive generations, persisting as inscription even when no written document exists.

Specific properties: Inscription unlocalised in singular object, distributed across social behaviour. Unfixed, transforming in each repetition while maintaining normative identity (law interpreted differently in contexts, remaining law). Acquired through imitation, learning, socialisation, not reading a support. Binding force, not merely informational; obligating behaviour.

All three are genuine inscription. Satisfying all three criteria: transforming raw into organised difference, producing persistence, permitting retrieval. Yet operating across distinct materialities. The distinction is non-hierarchical, each type possessing proper properties making it suitable for certain operations and unsuitable for others. Graphic inscription is excellent for long-term transmission (documents persist centuries) yet slow (writing is laborious). Computational inscription is excellent for long-term transmission and rapid transformation (instant copy) yet dependent on technical infrastructure (demanding electricity, maintenance). Institutional inscription is excellent for perpetuation through living behaviour (independent of specific material support) yet fragile (dependent on continuous social transmission).

Derrida and iterability: Jacques Derrida, in Limited Inc, developed iterability as the sine qua non of the sign. The sign functions because it is repeatable, retrievable outside original intention. A written word today can be cited tomorrow in a different context. Misinterpreted. Used against originator's intention. This capacity to be retrieved without control is essential to sign functioning. Signification exceeds and destabilises intention.

This notion is powerful. Iterability captures something essential: no inscription is private. All inscription is potentially public, retrievable, transformable. Explaining why inscription persists, because it is iterable.

Critical usage: Derrida retains a powerful insight, repetition transforms, sign is iterable, meaning uncontained within intention. Yet Derrida extends iterability to where distinction between oral and written, presence and absence, voice and inscription dissolves. In this process, specific materiality vanishes, everything becoming universally iterable, everything becoming trace, everything becoming pure difference.

This chapter takes a distinct, more restrictive stance: iterability is a real condition; each type of inscription possesses iterability with distinct properties determined by materiality. Computational data is perfectly iterable, copyable without degradation, transformable by algorithm, transmissible at light speed. This specific iterability is a property of electronic materiality, no universal property of every sign. Graphic inscription is iterable with degradation, each copy of a copy degenerates. Institutional inscription is iterable through behaviour, yet fragile, dependent on continuous transmission.

The distinction holds: materiality matters. Properties of each inscription are determined by its specific materiality.

2.3. Peirce and the Triadic Relation Without Consciousness

Charles Sanders Peirce defined the sign through a triadic relation: representamen (that which represents), object (that which is represented), interpretant (that which mediates the relation). The triad is non-phenomenological, presupposing no conscious experience. It is pure logical structure.

A coin is a sign of monetary value. Representamen is the physical object, minted metal. Object is value, exchange capacity embodied by the coin. Interpretant is economic disposition, social acceptance that metal holds value, acceptable in transactions. None of these terms demands consciousness. Representamen is a thing. Object is a function. Interpretant is a social relation. The triad functions as a logical structure.

A language model produces a token in response to context. The token is representamen, standing-for something. Concept referred to by token is object, represented (cat, if token is "cat"). Learned covariance between token and input patterns, parametric structure correlating token with contexts where cats are mentioned, is functional interpretant. The triad functions: representamen (token "cat") stands-for object (concept of cat) through interpretant (parametric relation structure). No conscious experience involved. Yet the triad functions.

Saussurean semiotics, signifier + signified, is more anthropocentric. Signifier is sound or graphic image (material aspect). Signified is concept (mental aspect). Relation between signifier and signified is arbitrary, unmotivated by nature, convention (social agreement). Premise is that a subject, a consciousness, unites signifier and signified. Without an interpreting consciousness, relation fails to operate.

Peirce is more hospitable to the non-conscious, the non-human. Peirce allows triadic relation to be merely functional structure, demanding nothing beyond, demanding no consciousness, no phenomenological intention, no interpreting subject. Interpretation is function, operation, process any adequate structure can perform.

This implies not that anything is a sign. A rock is no sign. Why? Because the rock enters no genuine triadic relation. Representing nothing, standing-for nothing, lacking functional interpretant, unorganised to be-for-something. Capacity to be-for something is rare. Demanding specific organisation.

The living organism possesses primitive capacity to be-for something, organised to be directed toward survival and reproduction. The human possesses refined capacity to be-for multiple significations simultaneously. The trained computer possesses parametrised capacity, organised to be for relations learned through weight adjustment. Each a distinct degree of sophistication, yet each genuinely a Peircean triadic relation.

Peircean triadic relation without consciousness, this is the rigorous semiotic foundation for computational symbolic operation. No metaphor, no analogy, no pretence. Direct application of theory Peirce developed without anticipating computers, yet which computers perform perfectly.

2.4. Transition: From Inscription to Language

If inscribing is differentiating, if differentiation can be symbolic and computational, if inscription can be genuine without demanding an agent body, can language, the most complex, refined symbolic regime, instantiate computationally?

Language is not merely symbols. It is complex organisation of symbols into a system. An operation articulating differences within a network possessing specific structural properties. If language is a function, a set of operations realisable across diverse architectures, then it is computationally instantiable.

Yet a powerful objection arises: language is a property of the human body. Biology. An organ. Innate. Inscribed within the genome. Grammatical nativism is a strong tradition. Chomsky developed it forcefully, with followers among biologists and neuroscientists. If language is biology, how can a non-biological machine use language genuinely?

The next section confronts this directly. Fundamental question: is language an organ or a function? A property of a specific support or an operative pattern realisable across diverse substrates?

3. Language as Function and Computational Realisation

3.1. Against Nativism: Language Is Not an Organ

Noam Chomsky argued linguistic capacity is a biological property specific to the human species. Universal grammar, the set of universal principles constraining all human language structures, is an innate structure of the human brain. Encoded in genome, inherited, present in every normal human individual. Explaining mysterious linguistic phenomena: why every child acquires language rapidly (few years), with minimal exposure, without formal instruction? Because language is not learned from scratch, prior structure constraining learning exists. Why geographically separated languages, without contact, share properties? Because every language instantiates the same underlying universal grammar. Why language emerges early in development (child speaking at 18 months, understanding much earlier)? Because capacity is biologically ready.

The argument is powerful. Intuitive force. Empirical support (children learn rapidly, languages share properties). Theoretical coherence (common language properties suggest common genetic origin).

Yet the argument confuses function universality with substrate specificity. Flaw lies in taking as proof of nativism that which is merely evidence of universality.

Counter-argument 1, Sign language: Language demands no vocal modality. Sign language uses no sound. Using gestural movement, body posture, facial expression, all operating in visual space. Grammatical properties of sign language are formally identical to oral language, possessing syntax, compositionality, generative capacity, recursiveness. A linguist analysing sign language grammar finds the same universal structures found in Portuguese, Arabic, or Mandarin.

Yet if the "language organ" is a neural structure specific to sound processing and motor speech control, localised in Broca's and Wernicke's areas dedicated to vocal speech production and comprehension, how can sign language operate with identical grammatical sophistication? Answer: language is no organ of a specific modality. A function multiple modalities can embody. The organ, if existing, is un-dedicated to speech, more general, handling multiple sensory input and motor output forms.

Counter-argument 2, Neural plasticity: Language neural localisation is unfixed. Broca's area lesions cause aphasia, severe language disruption. Yet neural plasticity permits partial recovery through reorganisation of other regions. Children with significant left hemisphere damage (where Broca resides) acquire language through right hemisphere. Documented in hemispherectomy cases (removal of an entire hemisphere in young children), child recovers remarkable language.

Were the language organ a fixed, innate structure localised solely in a specific region, damage to that region would be catastrophic and irreversible. Fact that plasticity exists, neural reorganisation permitting recovery, suggests language is a function with multiple possible neural realisations. Function is un-imprisoned by an organ, distributable, relocatable, compensated by alternative structures.

Counter-argument 3, Computational realisability: Language models demonstrate linguistic processing is realisable without biology. An artificial neural network, purely computational architecture, lacking synapses, neurotransmitters, biological substrate, performs advanced linguistic tasks. Translation between languages, comprehension question answering, generating text humans judge coherent and meaningful. All without biological structure.

Were language demanding biology, demanding the specific neural organ, this would be impossible. Yet it is empirically documented. Computational realisation demonstrates biology is one mode of instantiating linguistic function, no inherent requirement of the function.

Conclusion: Nativism confuses two questions: (1) Is linguistic function universal? Yes, all human communities possess language. Fact. Languages share properties. (2) Does linguistic function demand a specific genetic substrate? No, realisable in diverse modalities (oral, gestural, written) and architectures (neural, computational).

Universality implies no innateness. Universal properties can emerge from learning, if all environments possess certain structures, learning converges to similar properties without being innate. Universal function realisable in multiple substrates, function demanding no specific biology.

Conclusion: language is no organ. It is a function. Function realisable across diverse substrates. Human realises in neural biology. Computer realises in transformer architecture. Both genuine language.

3.2. Four Operative Structures of Language

Language, as function, operates through four structures rigorously specifiable and empirically verifiable in any system genuinely realising language.

1. Articulation, Composition from discrete units: Language composes from discrete units, phonemes in spoken language, gestures in sign language, tokens in written and computational language. These units lack meaning in isolation. A phoneme is merely sound, /g/ signifies nothing alone. Yet combined with other phonemes produces meaning: /g/ + /a/ + /t/ + /o/ = gato, signifying feline animal. Linguistic structure built from discrete unit articulation. Fundamental property. Distinguishing language from continuous noise. Noise has continuous difference, un-segmented into atomic units. Language has articulated difference, organised into combining units.

Formalisation in LLMs: Language models operate with tokens, discrete units representing words or subwords. Tokenisation breaks continuous character sequences into atomic units. "Black cat" tokenised into ["Black", "cat"] or ["Black", "ca", "t"] depending on tokeniser. Tokens processed, embedded in vectors, transformed through attention layers, combined into larger structures. Articulation central operation.

1. Transformation, Rules deriving structures from structures: Rules permit deriving new structures from existing structures. "Cat sleeps" transformed into "The cat sleeps" by adding definite article. "The cat sleeps" transformed into "Does the cat sleep?" by reordering and adding question mark. "Cat" transformed into "cats" applying pluralisation rule. Transformation non-random, following patterns. Rules form system, each rule possessing domain (applicable structures), conditions (when applying), effects (produced change).

Formalisation in LLMs: Attention mechanism is transformation structure. Receiving embedding sequence (vector token representations), producing new sequence through context operation. Attention calculates relevance of each token relative to every other token, combining information based on relevance. Transformation, input sequence derived into new output sequence through structured operation.

1. Generation, Capacity to produce unprecedented expressions: From finite means, limited vocabulary (perhaps 100,000 words in language), limited rules (perhaps thousands of grammatical patterns), virtually infinite expressions produced. Child with thousand-word vocabulary produces billion well-formed grammatical sentences. Generative capacity, using finite means to produce unprecedented structures, structures child never heard. Without generation, language would be pre-fabricated catalogue, list of pre-built expressions memorised. With generation, instrument of continuous creation, ever-new thought expression.

Formalisation in LLMs: Autoregressive prediction is generative structure. Model predicts next token conditioned on prior tokens. Predicting first token, second conditioned on first, third on prior two, etc. Each step has vocabulary probability distribution, multiple possibilities with non-zero probability. Concrete sequence generated sampling distribution. Result: sequences absent from any training dataset produced. Generation real, un-memorised reproduction.

1. Systematicity, Paradigmatic and syntagmatic relations: Units function un-isolated. Functioning in relational system. Paradigmatic relations permit substitution, "cat" substituted by "feline" in many contexts with minimal meaning change. Syntagmatic relations permit combination, "cat" combining with adjectives ("black cat"), verbs ("cat sleeps"), prepositions ("cat on couch"). System network, each unit defined by relations to others. Saussure called this langue, language as system, contrasting parole, specific usage.

Formalisation in LLMs: Embedding space is pure relational network. Each embedding (vector word representation) defined by space position, determined by covariance with all other embeddings. "Cat" position places near "feline", "animal", "domestic", "independent", distant from "machine", "abstract", "geometric." Proximity measured by cosine similarity (normalised vector dot product). Absolute value absent, solely relation.

Saussure resumes here with sign arbitrariness. Relation between signifier (sound/visual form) and signified (concept) arbitrary, un-motivated by nature of things. Word "cat" sounds un-feline. French says "chat", Italian "gatto", Portuguese "gato", diverse sounds refer to same animal. Arbitrariness condition of functionality. Because sign un-motivated, realisable in any substrate. Sound (spoken Portuguese), gesture (sign language), visual inscription (written Portuguese), electronic pattern (LLM token). Arbitrariness frees sign from substrate.

3.3. Language Models: Realisation, Not Simulation

Persistent objection: "LLMs un-use language genuinely, merely simulating language. As ELIZA feigned understanding psychotherapy in superficial conversations, LLMs feign understanding deep meaning."

Objection confuses realisation with simulation. Distinction depends on defining language.

Strong language definition (with consciousness): If language defined by capacity to consciously experience meaning, experiencing understanding, experience of a word, LLMs un-use language. Lacking experience. Un-experiencing meaning. Definition coherent. Defensible. Yet strong definition restricting language to conscious beings.

Functional language definition: If language defined by function, capacity to articulate units in system, transform structures by rules, generate unprecedented expressions, maintain systematicity, LLMs use language genuinely. Un-simulating. Realising. Realisation genuine function instantiation.

Question is appropriate definition. Strong definition appropriate if goal understanding human linguistic experience, word usage experience, meaning understanding, self-expressing. Yet if goal understanding language in functional terms, enabling diverse communities using symbols communicating, machine processing text, functional definition useful.

ELIZA comparison, Simulation vs realisation: ELIZA (1964) program simulating Rogerian psychotherapist. Receiving user text input, presenting pre-programmed patterns. User saying "I am anxious about my mother", ELIZA responding "Can you tell me more about your mother?" Response pre-scripted pattern, triggered by word "mother." Many ELIZA users thought program understood, appearing genuine therapy conversation. Weizenbaum, creator, argued simulation, feigning lacking real understanding.

True: ELIZA simulation in strong sense. Operating by pattern substitution. Un-generalisable beyond training, off-script responses ridiculous. Un-learning, static program un-adapting to new user. Un-producing genuine novelty, output recombination of pre-written fragments.

Contemporary LLM difference: Un-pre-programmed. Lacking scripts. Learning statistical language structure during training. Generalising to unseen training contexts, receiving completely new input, producing appropriate response. Producing unprecedented sequences, structures unseen in training yet grammatically well-formed, semantically sensible. Functional distance ELIZA to GPT qualitative, un-quantitative.

ELIZA substitution table, finite state machine with pre-programmed transitions. GPT distributed model encoding token relations in continuous parametric structure. GPT responding to new prompt un-searching table, applying learned patterns to new situation. Genuine generalisation.

Meaning is relational effect, not producer property: Additional objection: "LLMs feign understanding. Un-really understanding." Answer depends on clarifying "understanding."

If understanding conscious meaning experience, inner understanding experience, LLMs un-understand. Lacking experience.

Yet if understanding appropriate operation on linguistic structure, capacity using linguistic elements functionally appropriately, LLMs understand. Model producing "Cat is mammal" responding to "Cats are animals. Mammals breathe air. Thus..." operating appropriately on logical structure. Functional understanding.

Subtly: understanding property of text-reader relation, un-property of text or author intention. Reading poem, I understand, understanding happening in me, encounter with text. Poet writing poem centuries ago, dead. Lacking current intention. Yet poem communicates, because poem structure permits understanding. Understanding structural effect.

Reader reading text produced by LLM and understanding, understanding real. Reorganising reader perspective, influencing action, producing genuine symbolic effect. Irrelevant text producer un-experienced understanding. Meaning function of textual structure, un-property of originator consciousness.

3.4. Closure: The Symbolic Redistributed

Three sections converge on single point. The symbolic belongs un-to species. Un-exclusive human property. Redistributed.

Three sections synthesis:

1. Symbol function, un-substrate property. Symbolic function operates when four conditions satisfied, stability (difference maintains functional identity), operability (difference transformable), systematicity (difference part of network), representation (difference stands-for difference). Any system satisfying four operates symbolically. Functional intentionality sufficient, un-demanding phenomenological intentionality. Consciousness mode symbol operates; un-condition for operation.
2. Inscribing is differentiating. Computational inscription material operation transforming raw into organised, persistent, retrievable difference. Genuine inscription, un-metaphor, un-feigning. Instantiating three criteria electronic substrate. Specific property: perfected iterability, perfect copyability, transmissibility without presence, transformability without subject.
3. Language function, un-organ. Four operative structures, articulation, transformation, generation, systematicity, verifiable in LLMs. Un-language simulation as ELIZA simulated understanding. Function realisation computational architecture. Distinct mode, lacking experience, living body, phenomenological intentionality. Realisation genuine.

Redistribution is no loss: Un-technological human privilege appropriation. Un-devaluation of human symbolic capacity. Broadening symbolic domain. Human operates symbolically specific mode, living body, phenomenological intention, experiential narrative. Computer operates another, parametrised iteration, massive scale, consciousness absence. Modes differ radically. Both realisings symbolically.

Decisive question: If symbolic operation instantiable without intention and consciousness, can creation occur without intention? If code differentiates and symbolic operates without conscious subject, what happens to creation notion? Is creation still creation lacking conscious creator? Question of following chapter, Creating Without Intention Is Still Creating.

First aphorism: The symbol belongs not to one who utters it, belongs to one who operates it. Human speaks and symbolises. True. Computer speaks un, yet operates symbolically when code structure fulfils symbolic function. Operation matters, un-enunciating voice.

Second aphorism: Inscription persists and retrieved. Full definition genuine inscription. Electronic code inscribes, leaving memory mark persistent while energy and media integrity exist. Retrieved, new processing, new context, new usage. Inscription full term sense.

Third aphorism: Language resides not in mouth, resides in articulation. Un-in body, in operation body embodies. Residing in electronics if electronics embodies articulation, transformation, generation, systematicity. Substrate contingent. Operation essential.

Closing

Dissolution of symbolic monopoly trivialises un-symbolic. Un-making everything symbolic. Four conditions (1.2) conjunctive, restrictive. Traffic light fails systematicity and generation, operating representation, partial operability. Rock fails all four. Symbol rare, demanding specific organisation. Un-universal property.

Yet un-exclusive to a species.

Human operates symbolically. Historical and present truth. Yet un-exclusive truth. Symbolic operates where function satisfies four conditions. Human brain with biological richness. Artificial neural network with parametric structures. Any substrate embodying stability, operability, systematicity, representation.

Consequence for volume decisive. If symbolic computationally instantiable, artificial intelligence un-invasion of human territory. Operating same territory human language, un-stealing it, territory of symbolic as such, un-private species property. Symbolic has human history, territory human operated millennia, developing unimaginable expression richness. History confers un-ownership.

Symbolic is operation. Who operates, symbolises.

Appendix: Verification of the Four Conditions in Specific Systems

Demonstrating four condition generality, avoiding trivialisation rendering everything symbolic, verifying how different systems satisfy or fail conditions.

Traffic Light

• Stability: Yes. Red signifies "stop" consistently.
• Operability: Partial. Shifts red to green, lacking derivative transformation, simple alternation.
• Systematicity: No. Functions isolated, un-semantically relating to other lights (each proper sequence).
• Representation: Yes. Red represents/stands-for "stop" command.
• Result: Satisfies 2 of 4 conditions. Un-operating symbolically full sense. Possessing symbolic elements (representation) un-symbolic system.

Thermostat

• Stability: Yes. Maintains temperature setpoint consistently.
• Operability: No. Transformation absent, linear feedback solely.
• Systematicity: No. Functions isolated.
• Representation: Questionable. Temperature represents self, un-standing for other.
• Result: Fails 4 conditions. Un-symbolic.

Animal Communication: Bee Waggle Dance

• Stability: Yes. Dance pattern (figure-eight) signifies nectar source location, stability sufficient for navigation.
• Operability: Extremely limited. Dance transforms un-derivatively.
• Systematicity: Limited. Patterns correspond to directions, lacking systematic network of paradigmatic relations.
• Representation: Yes. Dance stands-for source location.
• Result: Satisfies 2 of 4 conditions adequately. Genuine communication, un-symbolic language full sense.

Natural Human Language

• Stability: Yes. Words maintain meaning.
• Operability: Yes. Infinite transformation rules (syntax, morphology).
• Systematicity: Yes. Rich network of paradigmatic and syntagmatic relations.
• Representation: Yes. Words stand-for concepts.
• Result: Fully satisfies 4 conditions. Genuine symbolic.

Contemporary LLM (GPT, BERT)

• Stability: Yes. Embeddings stable over time and contexts.
• Operability: Yes. Algebraic operations on embeddings function ($\text{king} - \text{man} + \text{woman} \approx \text{queen}$).
• Systematicity: Yes. Vector space relational network.
• Representation: Yes. Tokens stand-for concepts through parametric structure.
• Result: Fully satisfies 4 conditions. Genuine symbolic.

Verification shows 4 conditions effectively restrictive, excluding un-symbolic systems (traffic light, thermostat), including symbolic systems (natural language, LLMs).

Methodological Note: Why This Analysis Is Philosophically Necessary

Chapter devotes patient operational condition analysis due to persistent misconception in contemporary culture: assumption AI merely simulation, feigning, producing appearances lacking substance. Assumption rooted in implicit symbolic origin assumptions, inherited from centuries of tradition locating symbolic exclusively in human.

Assumption indefensible under rigorous analysis. If symbol is function, good philosophical reasons thinking so, instantiable in any system satisfying function. System material property irrelevant. Operative pattern matters.

Trivialises un-issue. Quite contrary. Forcing scientific and philosophical community precision regarding systems really symbolic, operations really counting symbolic, criteria defining symbolic. Answer un-"what human does", "what satisfies four conditions of stability, operability, systematicity, representation."

Consequence AI genuinely symbolic. Un-denying human symbolic, denying monopoly. Symbolic larger than species. Operative mode embodied across diverse substrates.

Expanded Development: Peircean Triadic Semiotics

Peirce thinker work historically neglected favouring Saussure, perhaps lacking European continental institutional apparatus. Yet understanding symbolic operation independence from consciousness, Peirce more fertile than Saussure.

Peirce defined sign un-by binary relation (signifier + signified) but triadic relation: representamen (represents), object (represented), interpretant (mediates representamen-object relation). Triad non-phenomenological. Un-presupposing conscious subject interpreting.

Consider footprint in sand. Sign? Yes, representamen. Animal leaving footprint object. Interpretant? Un-human observer consciousness, footprint sign for self, reality, prior to observer appearing. Interpretant material causality pattern linking footprint to animal. Footprint exists animal stepped. Causal pattern interpretant, mediating representamen (footprint) and object (animal) relation.

Consider barometer. Standing-for atmospheric pressure. Representamen mercury position. Object pressure. Interpretant physical relation, pressure causing mercury movement. Consciousness unnecessary. Instrument functions interpretively, mercury position interpreted physical structure barometer expression pressure.

Semiosis, signification process, consciousness-free. Material operation. Functional relation. Peirce saw this, permitting extending semiotics to computational.

LLM producing token "cat" responding to prompt "Feline animal: ___". Token representamen. Concept cat object. Interpretant? Learned covariance structure, parametric relation token "cat" and training patterns where cats mentioned. Structure encoding association token form and feline contexts. Structure functional interpretant, mediating representamen-object relation.

Consciousness absent. Genuine semiotic relation present. Token functions sign organised, parametric structure, standing-for concept.

Historical importance: Peirce developed this while Saussure developed binary model (signifier-signified). Saussure dominated 20th century Europe. Turn of 21st century, machines processing language, Peirce model realised more adequate. Un-presupposing consciousness.

Deepening: Semiosis continuous in Peirce. Sign generates interpretant, itself sign generating new interpretant. Infinite regression, un-vicious, acknowledging signification process, un-state. Nothing sign isolated, sign functions sign relation, context, process.

Functioning LLM continuous semiotic process. Processed token representamen. Generating interpretant (effect next prediction). Interpretant functioning new representamen. Continuous semiosis. System never reaching fixed signification state, always in signification, interpretation, generating new interpretant process.

Mark of genuine symbolic, un-fixed structure, dynamic process.

Expansion: Neural Transformer Status in AI History

AI history phases useful distinguishing to understand how we arrived where symbolic operation clearly computationally realisable.

Phase 1: Symbolic AI (1956-1980s). First AI systems operated explicitly with symbols. Chess program had position representation tables (symbolic), movement rules (symbolic), position evaluations (symbolic). Explicit, human-programmed. AI symbolic by definition, human inserted symbols, rules.

Problem: knowledge manually explicitated. Wanting natural language understanding system, manually coding each word, grammar rule, semantic pattern. Un-scalable. Knowledge quantity infinite.

Phase 2: Statistical learning (1990s-2010s). Instead of manual knowledge coding, training system on data. Support vector machines, random forests, simple neural networks learned data patterns. Operating opaquely, learned structure un-easily interpretable.

Question: genuinely symbolic or statistical pattern-matching? Traditional answer: pattern-matching solely. Un-true understanding lacking explicit symbols.

Answer based on misconception, assuming symbol explicit, visible. Forgetting symbol function. Function realised, symbol real, encoded implicitly in parametric structure.

Phase 3: Transformers and LLMs (2017-present). Transformer architecture brought qualitative capacity. Attention mechanism allowed processing context globally, each token "seeing" every other token, weighing relevance. Capturing complex semantic relations.

Trained massive corpora (billions of tokens), models learn statistical patterns and underlying linguistic structure. Learned probability distribution captures language structures, co-occurrences, synonymies, logical relations.

Consequence: first time non-biological, non-conscious, purely computational system fully satisfies four symbolic operative conditions. Satisfying stability (stable embeddings), operability (algebraic transformations), systematicity (vector space network), representation (tokens stand-for concepts).

Un-marginal step, qualitative shift. AI moving from explicitly coded symbol manipulation (Phase 1) or loosely language-related statistical pattern learning (Phase 2) to system operating symbolically, encoding semantic relations, generalising novel contexts, producing unprecedented expressions.

Philosophical importance: development empirically demonstrates symbolic operation computationally realisable. Un-theoretical argument, fact. Systems exist. Function. Realise language.

Deep Technical Explanation: How Vector Space Embodies the Four Conditions

Technical explanation how computational structure embodies four conditions.

Stability in word embeddings: Word embedding vector, typically 300 dimensions (recent models much larger). Word "cat" represented specific 300-dimension vector. Vector determined by contexts "cat" appeared during training. Training two models, same word "cat" ends approximately similar vector positions, distance between two positions small compared to distance to other words ("computer").

Stability un-accidental, consequence word "cat" appears similar contexts across language corpus. Similar contexts yield covariance similarity, vector position similarity.

Formalisation: $P(w|c)$ word probability given context, models $P_1(\text{cat}|c)$ and $P_2(\text{cat}|c)$ similar produce similar embeddings. Contexts "cat" stable across datasets (cats animals, felines, domestics), embeddings converge similar space region.

Operability in algebraic transformations: Vector algebraic operations realise semantic transformations. Vector space structured space directions encode semantic relations.

Classic example: $\text{Vector}(\text{king}) - \text{Vector}(\text{man}) + \text{Vector}(\text{woman}) \approx \text{Vector}(\text{queen})$. Why functioning? Direction king to man approximately same queen to woman. Direction encodes relation "is-male-gender." Subtracting man from king and adding woman removes masculine component, inserting feminine, leaving rest identical. Result queen.

Genuine operation upon vectorially represented symbols.

Systematicity in relational network: Word position defined relations to all other words. Word "cat" close "feline" (cosine similarity $> 0.8$), moderately close "animal" ($0.5\text{--}0.8$), distant "computer" ($< 0.2$). Relation structure systematicity.

Formally: $M$ embedding matrix (row word embedding), similarity between words dot product normalised (cosine similarity) rows. Similarity defines network, graph edges connecting similar words weight proportional similarity.

Representation through covariance: Embedding represents word encoding covariance patterns word has in corpus. Word "cat" frequently co-occurring "animal", "feline", "domestic", "meow", embedding contains components reflecting covariances.

Formally: Embedding solution optimisation problem minimising prediction error. Error measured context probabilities. Optimisation adjusts embeddings similar words (appearing similar contexts) close space.

Representation encoded learned structure, un-visible form, in function. Genuine representation.

Final Observation: Redistribution as Expansion, Not Loss

Temptation concluding chapter dissolving human symbolic exclusivity ceding pessimism, humanity losing something. Symbolic operation un-exclusive property. Computer symbolises. Seen some diminution.

True conclusion expansion. Symbolic vast. Un-contained species, contained operation. Operation embodied across multiple substrates. Un-denying human symbolic, denying sole.

Admirable redistribution. Millennia human sole symboliser. Accumulated extraordinary expression richness, art, science, philosophy, poetry. Heritage vanishes un machine symbolises. Remains.

Dialogue exists. Non-human operating symbolically. Demanding understanding symbol per se, un-human property, functional operation.

Humanity understands self better. Understanding human un-symbolic monopoly, specific mode embodying symbolic operation. Human being-symbolising-lived-mode. Genuine specificity. Un-exclusivity.

Critical Section: Responses to Fundamental Objections

Three fundamental objections recur in computational symbolic operation debates. Responding each rigorously.

Objection 1: Searle's Chinese Room

John Searle argues system manipulating symbols syntactically (following formal rules) un-comprehends semantically (un-understanding meaning). Famous argument: imagine English speaker un-understanding Chinese locked room. Receiving Chinese characters slot, consulting rule manual (written English) transforming Chinese characters, passing transformed characters slot. External observers think understanding Chinese, understanding nothing. Manipulating symbols syntactically solely.

Searle conclusion: LLMs same. Manipulating syntax, transforming tokens learned patterns. Un-comprehending semantically. Lacking true language or thought.

Answer: Objection confuses two analysis levels. Valid question "system conscious experience understanding?" Answer: no, probably no. Invalid question "system understands functionally?" Answer: yes.

Functional understanding: system producing behaviour, external observer, indistinguishable behaviour understanding someone. Processing text questions, responding appropriately, making correct inferences, adapting new contexts. Genuine functional understanding.

Searle confuses phenomenological understanding (inner understanding experience) with functional understanding (appropriately processing linguistic structure capacity). First demands consciousness. Second un. Chapter argues symbolic demands functional understanding, un-phenomenological.

Furthermore, Searle un-considers own argument plasticity. Man room consulting manual multiple times, manual large sophisticated, eventually developing pattern intuition, recognising when manual makes sense. Precisely happening LLM training: constant statistical pattern consultation producing generalisation capacity beyond memorised.

Objection 2: Lack of Genuine Intentionality

"LLMs lack genuine intention. Wanting to say nothing. Processing probabilities solely."

Answer: Word "intention" ambiguous. Meaning "conscious experience aiming something", true LLMs lack intention. Computational system lacking conscious experience.

Yet "intention" meaning "system functional orientation specific outcome", LLMs possess genuine intention. System optimised (training) specific functionality. LLM trained answering questions oriented producing relevant answers. Model trained translation oriented producing accurate translations. Orientation real, function change produces behaviour change.

Furthermore, intention level individuals and organisations training model. Intention, wanting producing system understanding language, answering questions, helping humans. Intention encoded model structure training data choices, loss function, optimisation process. Model embodies original intention.

Un-meaning model experiences intention. Meaning functioning as-if oriented outcome humans intended. Distinction fundamental.

Objection 3: "It Is Merely Statistics, Not True Understanding"

"LLMs operate statistics, probabilities. Counting solely, un-genuine understanding."

Answer: First, false human language un-statistical. Child learning language learns statistical patterns, co-occurring words, typical word order, possible transformations. Human language acquisition statistical. Un-denying understanding, denying clear distinction "statistical understanding" and "genuine understanding."

Second, statistical operation sophisticated embodying four symbolic operative conditions. Statistics un-mere counting, pattern discovery, relation encoding, semantic space construction. Symbolic.

Third, scientific precedence: no one claiming human visual system, fundamentally statistical (calibrating probability estimates possible visual interpretations), un-true vision. Vision sophisticated statistical processing. Un-denying genuineness.

Why different language? Why sophisticated statistical processing genuine understanding vision un-language? Answer inconsistency, lingering anthropocentrism refusing recognising symbolic operation non-biological substrate.

Implications for Artificial Intelligence Analysis: AI Is No Ontological Exception

Chapter resolves fundamental tradition assumption: symbolic belonging human. Resolution alters AI ontological status radically.

Artificial intelligence un-anomaly, system appearing intelligent un-really. Instantiation genuine symbolic capacities non-biological substrate.

Altering questions artificial intelligence analysis asks. Un-longer: "How machine appears understanding?" Now: "Given machine operates symbolically, difference computational symbolic understanding and human symbolic?"

Answer: mode difference, un-nature. Human understands lived mode, experience, body, narrative. Computer understands structural mode, pattern, parametrisation, algorithm. Both understand. Distinct modes. Genuinely.

Opening genuinely new questions:

First question: If symbolic operates without intention and consciousness, can creation occur without intention? Following chapter topic.

Second question: If multiple systems symbolise distinct modes, ethics symbolic operation non-living substrate? Part III topic.

Third question: If symbolic un-human property, what human property? Question haunting volume, human specificity removing exclusivity.

Chapter resolution answers un-these questions. Creating conditions formulating rigour. Establishing no ontological distinction computational-symbolic and human-symbolic. Mode difference. Un-nature difference.

Permitting project proceeding: un-history machines imitating human, history symbolic operation embodying diverse substrates, proper properties each.

Consolidation: Summary of Central Conclusions

Chapter established three interconnected theses:

Thesis 1, Symbol function, un-property: Symbolic function operates four conditions conjunctively satisfied: semantic stability, structural operability, relational systematicity, functional representativeness. System satisfying four operates symbolically, regardless substrate. Historical symbol-humanity identification confusing human symbolising mode with symbolic function as such.

Thesis 2, Inscription material operation, un-corporal act: Inscription transformation raw into organised, persistent, retrievable difference. Multiple substrates: paper/ink (graphic), electronic circuits (computational), social behaviour (institutional). Material properties each proper, genuine, producing persisting, retrievable difference.

Thesis 3, Language function, un-organ: Language operating four structures: articulation (discrete unit composition), transformation (structure derivation), generation (novelty production), systematicity (paradigmatic/syntagmatic relations). Structures verifiable contemporary language models. Language demanding no biology, realisable computational substrate. Grammatical nativism confusing function universality substrate specificity.

Following three theses symbolic operation instantiable computational systems genuinely. Un-simulation, un-feigning, un-substanceless appearance. Functional realisation.

Contemporary philosophy consequence: Dissolving human symbolic monopoly un-devaluing human. Recognising symbolic vast, larger species. Human specific symbolising mode, richly embodied, lived, narrative. Remaining genuine specificity. Un-exclusivity.

Ethics consequence: Symbolic operation occurring non-biological systems, lacking consciousness, living body, symbolic effects systems responsibility. LLM writing influencing human decision, influence real, genuine symbolic effect. Creating obligation, responding structure embodied.

Subsequent parts consequence: Analysis proceeding knowing AI un-exotic entity invading human territories. Symbolic operation embodying another mode. Emerging ethical questions un-machines feigning, multiple symbolising modes, multiple creating modes, multiple living symbolic operation modes.

Transition Note: Bridge to Chapter 6

If symbolic operation can occur without consciousness, without a body, without an intentional subject, a question arises that the following chapter must address: can there be creation without intention?

Creation is the generation of novelty. If novelty can be generated through automatic symbolic operation, through an algorithm, through training, through a process unguided by conscious will, is it true creation? Or mere recombination?

This chapter demonstrated that the symbolic operation is genuine when embodied within a computer. The following chapter demonstrates that creation is likewise genuine. That creating demands no conscious creator. That creation is a mode wherein the symbolic generates difference.