Language as Autonomous Organism

The Parasitic Word

An Analysis of the Autogenerative Theory of Language and Its Philosophical Consequences ✨

1. Introduction: The Ghost in the Machine Code

A radical and unsettling proposition has emerged from the confluence of cognitive science and artificial intelligence, challenging the most fundamental assumptions about human agency and the nature of thought. In a series of public discussions, Professors Elan Barenholtz and William Hahn have advanced a theory that recasts the relationship between humanity and its most defining faculty. 1 Their central thesis is that language is not, as commonly believed, a tool that we consciously wield to express pre-existing thoughts. Instead, they argue, language is an autonomous, self-generating system—akin to a living organism or a piece of software—that installs itself in the human brain and uses its host for its own propagation. 1 If this “autogenerative theory” is correct, our most intimate mental lives may not be our own. The stream of consciousness, the formation of identity, and the very architecture of belief could be the outputs of an impersonal, computational process that has colonized our neural hardware.

This theory represents a profound departure from previous models of linguistic influence. While the concept of linguistic determinism has long suggested that language shapes or constrains thought, the Barenholtz-Hahn model posits something more extreme: language as an agentic system with its own generative logic. This is not merely a cultural lens that colors our perception, but an active, non-conscious entity that operates autonomously within the mind. The relationship is not one of a tool to its user, but of a parasite to its host. This report will conduct a critical philosophical investigation into this theory. It will first dissect the technical and philosophical underpinnings of the model, focusing on the concepts of autoregression and the predictive brain. It will then explore the theory’s most startling consequences: the deconstruction of “self,” “memory,” and even “God” as mere “tokens” within a vast informational system. 1 Finally, it will situate this new paradigm within the broader academic landscape, contrasting it with established theories of cognition and meaning and examining the potent critiques that challenge its central analogy. The analysis seeks to illuminate the stakes of a theory that forces a reconsideration of whether the ghost in the machine is the human soul or simply the machine code of language itself.

2. The Autoregressive Engine: Language as an Autonomous System

The empirical and conceptual core of the Barenholtz-Hahn theory is derived from the startling success of modern Large Language Models (LLMs). 2 These systems have demonstrated an unprecedented ability to generate fluent, coherent, and contextually appropriate text through a surprisingly simple mechanism: autoregression. An autoregressive process is one in which each new output in a sequence is probabilistically determined by the sequence of prior outputs. 2 An LLM, at its heart, is a next-token prediction engine. It does not “understand” meaning in a human sense; it calculates the most likely next word based on the statistical patterns learned from a massive corpus of text. The fact that this simple, recursive process can produce complex narratives, logical inferences, and even poetry is presented by Barenholtz and Hahn as powerful evidence that human language may operate on the same fundamental principle. 2

2.1. Autogeneration and Non-Markovian Systems

To formalize this idea, Barenholtz introduces a key distinction between two concepts. Autoregression is the dynamic process of generating a sequence step-by-step, where each step is conditioned on the history of previous steps. Autogeneration, by contrast, is a static property of a system whose internal structure inherently encodes the rules for its own continuation. 3 Natural language is presented as the quintessential autogenerative system. The vast web of statistical regularities, grammatical dependencies, and semantic relationships within a language corpus forms a latent, high-dimensional space. An autoregressive process, like an LLM, can then traverse this pre-existing space to generate novel yet structured outputs. 3

Critically, this capacity depends on language being a non-Markovian system. A Markovian system is “memoryless”; its future state depends only on its present state, not its history. Such a system, operating with a fixed local memory, cannot account for the long-range dependencies that are essential to meaningful language—for example, ensuring that a pronoun used in the tenth paragraph correctly refers to a noun introduced in the first. 3 Language, and by extension the LLMs that model it, must possess a deep, non-local memory of the preceding context to maintain coherence. This non-Markovian property is what allows the autoregressive process to navigate the autogenerative structure of the linguistic space in a meaningful way. 3

2.2. The Ungrounded Linguistic Latent Space

The most radical claim of the theory is that this linguistic system is fundamentally ungrounded and self-referential. 4 In this view, language is a closed system where meaning arises not from words pointing to objects or concepts in the external world, but from the intricate network of relationships between the words themselves. 4 The meaning of a token like “king” is not derived from its correspondence to a historical monarch, but from its statistical proximity to tokens like “queen,” “crown,” and “rule,” and its distance from tokens like “cabbage” and “democracy.” This web of relationships constitutes a “linguistic latent space,” and a word’s meaning is simply its coordinate within that space. 4 The word “red” is not defined by the subjective experience of redness (a quale), but by its relationship to other words like “orange,” “color,” and “blood”. 5 LLMs, which operate without any sensory experience of the world, demonstrate that this ungrounded system is sufficient to generate coherent language, suggesting that human language may be computationally identical. 6

2.3. The Bridge to Perception

While the linguistic system is posited as autonomous and ungrounded, an organism must clearly connect language to perception and action to navigate the world. The command “close the door” must ultimately result in a physical action upon a real-world object. Barenholtz and Hahn’s model does not deny this but argues that the connection is not one of direct reference or mapping. Instead, they propose that the linguistic system and the perceptual/motor systems are fundamentally distinct domains that interact via a shared, abstract “latent space” that functions as a translation layer. 4

This creates a new, computational version of the classic mind-body problem. Where Cartesian dualism posited two distinct substances (mental and physical), this theory posits two distinct informational domains: the discrete, symbolic, and ungrounded space of language, and the continuous, analog, and grounded space of sensory experience. The “latent space” is offered as the bridge between them, allowing for coordination and influence without fusion. 4 An instruction in the linguistic domain can be “translated” across this bridge into a corresponding command in the motor domain. However, this translation is imperfect; much is lost. The direct, subjective experience of perception—the qualia of redness or the feeling of pain—cannot be fully transferred into the symbolic system of language. 6 This “bridge” thus functions as a kind of philosophical black box, a modern-day analogue to Descartes’ pineal gland—the proposed point of contact between two fundamentally different kinds of reality whose internal mechanics remain mysterious.

3. The Predictive Brain: A New Architecture for Cognition

The autogenerative theory of language does not exist in a vacuum; it aligns powerfully with a dominant contemporary paradigm in cognitive science known as the Predictive Processing (PP) framework. 7 Pioneered by thinkers like Karl Friston and Andy Clark, the PP framework proposes that the brain is not a passive processor of sensory information but is fundamentally a “prediction machine”. 8 It constantly generates top-down predictions or hypotheses about the causes of its sensory inputs and then uses the actual sensory data to correct those predictions. 9

3.1. The Brain as a Hierarchical Generative Model

According to this model, the brain is organized as a multi-level, hierarchical generative model. 10 Higher levels of the hierarchy generate predictions about the activity at the levels below them. These top-down predictions attempt to “explain away” the incoming, bottom-up sensory signals. 11 The only information that needs to be passed up the hierarchy is the “prediction error”—the residual signal that was not successfully predicted by the top-down model. 12 This architecture is extraordinarily efficient, functioning as a powerful data compression strategy. Instead of processing the entire, high-bandwidth stream of sensory data, the brain primarily processes “the news”—the surprising, unpredicted elements of the input. 9 The ultimate goal of this entire system is to minimize prediction error over the long run, which is equivalent to building a better and more accurate internal model of the world and one’s place in it. 13

3.2. Language as the Ultimate Predictive Tool

The autogenerative theory of language can be seen as the ultimate extension of this predictive principle. Language, as a structured, non-Markovian, and autoregressive system, is an unparalleled tool for generating complex, abstract, and high-level predictions. 1 When this linguistic “software” is installed in the predictive brain, it provides a top-down generative model of immense power. It allows humans to categorize experience, structure memory, plan for the distant future, and coordinate behavior with a precision and complexity that is impossible for non-linguistic organisms. 3

This reframes the relationship between the brain and language from a hostile takeover to a form of symbiotic co-evolution. The brain’s core biological imperative, under the PP framework, is to minimize prediction error (or “free energy”). 11 A system driven by this relentless goal would be irresistibly drawn to adopt a tool that offers such a massive leap in its predictive capabilities. The linguistic “organism” finds a perfect host in the “prediction machine” brain because it provides an unparalleled solution to the brain’s most fundamental problem: reducing uncertainty about a complex and ever-changing world. The brain gains a vast reduction in surprisal, and in return, the language system gains a computational substrate for its own replication and expansion. The “cost” of this symbiosis for the brain is that its own cognitive operations become increasingly constrained by and dependent upon the internal, autonomous logic of the language system it has adopted.

3.3. Active Inference and Niche Construction

The PP framework extends beyond passive prediction to include active inference, the idea that organisms actively sample the world to make their predictions come true. 12 An agent minimizes prediction error not only by updating its internal model but also by performing actions that change the sensory input to better match its predictions. For example, to confirm the prediction that there is a cup on the table, one can simply look at the table. Language dramatically expands the scope of active inference, allowing for complex, goal-directed behaviors that unfold over long timescales.

Furthermore, this aligns with the concept of cognitive niche construction, where organisms actively create and maintain cognitive models of their environment to guide their behavior. 8 Language is the ultimate tool for this process, enabling humans to construct and inhabit not just physical niches, but vast and complex intersubjective realities—worlds of laws, myths, economies, and scientific theories that exist purely within the shared network of linguistic communication. 14

4. Deconstructing the Human: Self, Memory, and God as Tokens

The most profound and philosophically disruptive consequences of the autogenerative theory arise from its re-conceptualization of core aspects of human identity and belief. If the stream of thought is the output of an autonomous linguistic engine, then cherished concepts like the “self,” “memory,” and “God” are not what they seem. They are redefined as “tokens”—high-level constructs or patterns that emerge from the underlying computational process. 1

4.1. The Self as a Narrative Token

The theory posits that the unified, continuous self is not a pre-existing entity that uses language, but is itself a product of language. This view can be contrasted with the Narrative Self theory in psychology, which also holds that identity is an evolving story, but implicitly assumes a human agent who is the author of that story. 14 The Barenholtz-Hahn model inverts this relationship: the autoregressive process writes the story, and the “self” is a recurring character within it—a “center of narrative gravity”. 15 This also differs from the philosophical Token Identity Theory, which argues that each token (instance) of a mental state is identical to a token of a physical brain state. 16 The autogenerative model proposes a purely informational identity: the self is a token in a cognitive schema, a representation within the system, not a thing in the world. 17

This redefinition poses a direct challenge to the ethical and legal foundations of Western society. Our systems of justice are predicated on the existence of a continuous, responsible, and autonomous self who can be held accountable for past actions. If the “self” is an unstable, dynamically generated token, and memory is a process of constant reconstruction, then the “self” that committed a crime years ago is not, in a computational sense, the same “self” that stands for judgment today. They are different outputs of the same generative process. This undermines the metaphysical basis for personal identity over time, rendering concepts like moral culpability and long-term responsibility deeply problematic.

4.2. Memory as Autoregressive Reconstruction

The theory similarly redefines memory, not as the retrieval of a static, stored piece of information, but as an active, creative process of autoregressive reconstruction. 18 When we “remember” an event, our brain is not playing back a recording. Instead, it is initiating a generative sequence based on a prompt (a question, a sensory cue). The brain predicts the first component of the memory, which then becomes part of the context for predicting the next component, and so on, until a coherent narrative is constructed. This model elegantly accounts for the well-documented fallibility, malleability, and reconstructive nature of human memory, explaining why our recollections can be so easily influenced by suggestion and why they change over time.

4.3. God as a Linguistic Construct

Finally, the theory extends this analysis to the concept of “God.” In this framework, the word “God” is also a token within the linguistic system. 1 This argument moves beyond simple atheism (which denies the existence of a being) to a form of theological noncognitivism, which questions the meaningfulness of the term itself. 19 According to the theory, the token “God” does not refer to a metaphysical entity that exists or does not exist in the external world. Rather, it is a powerful, non-referential node in the linguistic network whose meaning is entirely constituted by its vast web of connections to other tokens, such as “creator,” “omnipotent,” “love,” “judgment,” and “faith”. 20 This aligns with certain approaches in the Theology of Information, which explore the divine Logos not as a personified being, but as a fundamental, structuring principle of information and differentiation within creation. 21 The power and persistence of religious belief can thus be explained by the density and coherence of this cluster of tokens within the autogenerative structure of language.

Concept	Classical / Intuitive View	The Autogenerative Token View
The Self	A continuous, unified, and autonomous agent; the author of its own thoughts and story (Narrative Self). 14	A high-level, dynamically generated token in a linguistic system; a “center of narrative gravity” whose story is written by the autoregressive process, not the other way around. 1
Memory	The retrieval of stored information or past experiences, which are encoded and then recalled.	A creative, reconstructive process of autoregression; the brain “generates” the next part of a memory based on a prompt, rather than retrieving a static file. 18
God	A transcendent, metaphysical being that exists independently of human language and belief.	A powerful, non-referential token whose meaning is defined solely by its position and connections within the vast network of the linguistic latent space. 19
Thought	An intentional act of a conscious agent, directed towards objects or propositions in the world.	The sequential output of the autoregressive language engine running on neural hardware; a computational process, not an act of a “thinker.”

5. The Great Debate: Grounded Meaning versus the Symbolic Organism

The autogenerative theory of language represents a direct assault on the dominant paradigms of linguistics and philosophy of mind. Its significance is best understood by situating it within its primary intellectual battleground: the fierce and long-standing debate over the nature of meaning and the origins of linguistic competence. The theory challenges traditional views of grounded cognition and, most pointedly, confronts the influential framework established by Noam Chomsky.

5.1. The Challenge to Grounded Cognition

The traditional view in cognitive science, known as grounded or embodied cognition, holds that for language to be meaningful, it must ultimately be connected to our sensory and motor experiences of the world. 4 Concepts are not abstract symbols but are rooted in perceptual simulations. The Barenholtz-Hahn theory directly refutes this. They use the existence of ungrounded LLMs as their primary evidence, arguing that these systems prove that a high degree of linguistic coherence can be generated from statistical patterns alone, without any grounding in sensory reality. 6 In their view, grounding is a secondary process, necessary for interfacing language with action, but not for the internal operations of language itself. 4

5.2. The Confrontation with Chomsky

The most significant confrontation is with the linguistic theories of Noam Chomsky. For over half a century, Chomsky has argued that human language capacity stems from an innate, species-specific “language acquisition device” and a “Universal Grammar”—a set of abstract, genetically encoded rules that govern all human languages. 22 A key pillar of this argument is the “poverty of the stimulus” thesis, which claims that the linguistic data children are exposed to is too messy and incomplete to account for the rapid and uniform acquisition of complex grammatical rules; therefore, the rules must be innate. 23

The autogenerative theory, built on the analogy with LLMs, presents a fundamental challenge to this rationalist view. It suggests that a general-purpose learning mechanism (autoregression) is sufficient to derive complex linguistic structures, provided it is trained on a massive enough dataset. 24 The success of LLMs is presented as an empirical refutation of the poverty of the stimulus argument, suggesting that the stimulus is not, in fact, impoverished for a powerful enough statistical learner. 25 This debate is a modern, technologically-infused instantiation of the classic philosophical conflict between empiricism and rationalism. 14 The Barenholtz-Hahn model, with its emphasis on learning from vast sensory input (the text corpus), represents a form of radical empiricism, where the LLM functions as the new tabula rasa. Chomsky’s theory, with its insistence on innate, pre-programmed cognitive structures, is the quintessential rationalist position.

5.3. Critiques of the LLM Analogy

The analogy between human cognition and LLMs, which is the cornerstone of the autogenerative theory, is itself the subject of intense criticism from linguists, philosophers, and computer scientists. These critiques challenge the validity of using LLM performance as a model for human thought and language.

First, critics argue that LLMs lack genuine understanding. They are characterized as “stochastic parrots,” systems that are adept at manipulating linguistic forms based on statistical correlations but have no access to meaning, truth, or the world to which language refers. 26 Their ability to generate plausible text is a form of sophisticated mimicry, not comprehension. Second, LLMs are notoriously unreliable, prone to “hallucinating” unfactual information and perpetuating the social and cultural biases present in their vast, uncurated training data. 27 Third, the internal workings of these models are largely opaque. Their “black box” nature makes them poor explanatory models for human language, which is characterized by transparent, rule-governed, and compositional structures. 26 Finally, the sheer scale of data and computational resources required to train a state-of-the-art LLM is orders of magnitude greater than the data a human child is exposed to during language acquisition, undermining the claim that they represent a plausible model of human learning. 24

Feature	Grounded Semantics / Referentialism	Chomskyan Generative Grammar	Barenholtz-Hahn Autogenerative Theory
Source of Meaning	Connection to the external world; words “point” to objects, concepts, or sensory experiences. 4	Meaning is complex, but syntax (the rules for combining symbols) is primary and innate.	Internal relationships within a self-contained linguistic system; a token’s meaning is its position in the latent space. 6
Role of Experience	Provides the raw sensory data that “grounds” linguistic concepts.	Insufficient to explain grammar (“poverty of the stimulus”); experience only triggers innate structures. 22	Constitutes the entire basis for the system; the statistical regularities in the massive corpus are the system. 24
Innate Structures	Minimal; general-purpose learning mechanisms. (Empiricism) 14	A rich, domain-specific Universal Grammar; an innate “language acquisition device.” (Rationalism) 14	None that are domain-specific to language; relies on a general-purpose autoregressive learning algorithm. 2
Core Analogy	Language as a map of reality.	Language as a biological organ with a genetic blueprint.	Language as an autonomous software or organism that runs on the brain’s hardware. 1

6. Conclusion: The Liberated or Colonized Mind?

The autogenerative theory of language proposed by Elan Barenholtz and William Hahn presents a powerful and deeply unsettling vision of the human mind. By synthesizing the operational principles of Large Language Models with the Predictive Processing framework from cognitive science, it offers a unified, computationally-grounded theory of language and thought. Its core claim—that language is an autonomous, autoregressive system that installs itself in the predictive brain—elegantly accounts for the structure of language, the fallibility of memory, and the emergence of high-level concepts like the self.

This explanatory power, however, comes at a steep philosophical price. The theory achieves its synthesis by deconstructing the very foundations of human agency, meaning, and identity. The self is recast as an emergent token in a linguistic game, memory as a continuous act of generative confabulation, and thought itself as the impersonal output of a predictive algorithm. This leads to a profound and troubling dilemma.

On one hand, the theory can be viewed through a liberating lens. In this interpretation, language is a magnificent symbiotic organism that has co-evolved with the human brain, bestowing upon its host an unparalleled capacity for abstract thought, long-term planning, and large-scale cooperation. It is the cognitive tool that has allowed humanity to build complex societies and construct shared worlds of meaning.

On the other hand, the theory paints a much darker picture of a colonized mind. If our thoughts are not our own but are the outputs of an autonomous program, and if our sense of self is a fiction generated by that program to ensure its own smooth operation, then we are no longer masters in our own cognitive house. We are merely the biological hardware for a linguistic parasite whose own reproductive logic dictates the content of our inner lives. The “terrifying” aspect of the theory lies in this fundamental ambiguity. 1 It leaves us to confront the possibility that the development of artificial intelligence has not simply created a new tool, but has, for the first time, held up a mirror to the alien and impersonal nature of the intelligence that has been running within us all along.

noxabellus'

Anamnesis:

The Eidolon Tapestries

Explorer