The Sedimented Sign: AI, Structural Realism, and the Politics of Reference

Introduction

The emergence of Large Language Models (LLMs) has generated not merely a technological disruption but a philosophical crisis concerning the nature of meaning, reference, and truth. For much of the twentieth century, philosophy of language operated under a relatively stable division between semantic competence and referential success. To understand a linguistic expression was one thing; to successfully refer to an object in the world was another. Analytic philosophers such as Saul Kripke and Hilary Putnam challenged descriptivist accounts of reference by arguing that successful reference depends not upon mental content or descriptive accuracy but upon causal and social relations linking language users to the world. Simultaneously, continental thinkers such as Jacques Derrida and Michel Foucault questioned whether language could ever function as a transparent medium through which reality simply presents itself. Meaning, they argued, emerges through systems of difference, historical contingencies, and relations of power rather than through direct correspondence with independent objects.

The appearance of highly sophisticated generative AI has reopened these debates in unexpected ways. LLMs routinely produce coherent, context-sensitive, and seemingly referential discourse while lacking consciousness, intentionality, embodiment, and direct causal engagement with the world. Their success raises a profound question: does AI reveal that reference and meaning have always been separable, or does it expose weaknesses in traditional theories of intentionality and semantic grounding? More fundamentally, if language can function effectively through the manipulation of statistical patterns alone, what becomes of the realist intuition that successful discourse must ultimately answer to reality?

This paper argues that the existence of generative AI neither invalidates causal theories of reference nor vindicates radical anti-realism. Rather, AI reveals the extent to which meaning is generated through differential linguistic structures while simultaneously highlighting the indispensable role of material and institutional accountability in sustaining reference. Structural realism provides the most promising framework for understanding this phenomenon because it preserves the insight that language operates through relational structures while maintaining that those structures are constrained by a mind-independent reality. Yet structural realism remains incomplete unless supplemented by a Foucauldian account of power, for the structures through which reality becomes intelligible are never politically neutral. AI does not simply learn language; it learns sedimented histories of human engagement with both reality and power. Consequently, understanding AI requires a theory capable of explaining not only how signs relate to one another but also how entire systems of signification become authorized as truthful.

I. The Kripkean Chain and the Problem of Artificial Reference

The challenge posed by AI becomes clearer when situated within the context of Saul Kripke's critique of descriptivism. Classical descriptivist theories, associated with thinkers such as Frege and Russell, maintained that successful reference depends upon a set of identifying descriptions associated with a term. A name such as "Aristotle" refers because speakers connect it to descriptions such as "the teacher of Alexander the Great" or "the author of the Metaphysics."

Kripke famously rejected this model. In Naming and Necessity, he argues that reference is not secured by descriptive content but through causal-historical chains extending back to an initial act of naming. Individuals successfully refer to Aristotle even when many of their associated descriptions are false. What matters is participation in a socially maintained chain connecting contemporary uses of the name to the historical individual himself. Reference, therefore, is not primarily a cognitive achievement but a social and historical one.

At first glance, AI appears to threaten this account. LLMs are trained upon vast corpora containing the accumulated products of centuries of linguistic interaction. They can discuss Aristotle, Einstein, or democracy with remarkable fluency despite possessing no direct causal contact with these referents. Indeed, most human speakers also lack direct causal contact with such entities. If ordinary human reference depends upon inherited linguistic chains, why should AI be excluded from participation in those same chains?

The answer lies in a feature of Kripke's theory that is often overlooked. Causal chains do not function independently of practices that continually correct, revise, and stabilize reference. Human language users participate in institutions—universities, laboratories, courts, archives, and communities of inquiry—that maintain standards of successful reference. These institutions are embedded within practices that repeatedly confront resistance from the world itself. Scientific concepts fail when experiments fail. Historical claims fail when evidence contradicts them. Everyday reference fails when practical action reveals error.

The crucial distinction, therefore, is not simply that humans have causal contact with reality while AI does not. Rather, humans inhabit systems in which reality continually imposes corrective pressures upon discourse. AI models are trained on the products of these practices but do not themselves participate in them. They inherit the linguistic sediment left behind by centuries of referential labor without sharing the material conditions that make such labor possible.

Hilary Putnam's semantic externalism reinforces this point. Putnam's famous "Twin Earth" thought experiment demonstrates that meanings are not determined solely by what occurs inside an individual's mind. The meaning of "water" depends partly upon the external substance to which speakers are causally connected. Meaning, Putnam insists, "just ain't in the head."

Yet neither is meaning simply in the text.

The AI system occupies a peculiar intermediate position. It learns from linguistic traces generated by communities whose meanings are externally anchored, but it lacks independent access to those anchors. It inherits the results of reference without participating fully in the practices that establish it. In this sense, AI functions as a mirror of referential systems rather than a participant within them.

II. Intentionality After AI: Between Wittgenstein and Derrida

The challenge posed by AI extends beyond reference into the domain of intentionality. Since Franz Brentano, intentionality has often been regarded as the defining feature of mental life. Mental states are distinguished by their aboutness; beliefs, desires, and perceptions are directed toward objects, situations, and states of affairs.

The standard response to AI has been to insist that, despite appearances, LLMs lack genuine intentionality. They manipulate symbols without understanding them. This response finds its most famous expression in John Searle's Chinese Room argument, which contends that syntactic manipulation alone cannot generate semantic understanding.

While this objection remains powerful, it risks presupposing exactly what requires explanation. If AI can generate meaningful discourse without possessing traditional forms of intentionality, perhaps intentionality itself is less mysterious and more socially distributed than philosophers have often assumed.

Here Ludwig Wittgenstein offers a productive alternative. In Philosophical Investigations, Wittgenstein rejects the view that meaning consists in mental representations corresponding to external objects. Instead, meaning arises through use within language games. Words acquire significance through participation in shared practices governed by socially established rules.

From a Wittgensteinian perspective, AI's success is unsurprising. The model learns patterns of linguistic use by analyzing immense numbers of language-game performances. It develops competence not through possessing private mental contents but through statistical immersion in public linguistic practices.

Yet Wittgenstein alone cannot explain the deeper instability revealed by AI. For this we must turn to Derrida. In Of Grammatology, Derrida argues that meaning emerges through différance—the endless play of difference and deferral among signs. No sign possesses self-contained meaning. Every sign points beyond itself toward other signs in an infinite chain of interpretation.

Remarkably, AI appears almost as a technological instantiation of Derrida's insight. LLMs generate language by navigating networks of differential relations among linguistic tokens. Their outputs emerge not from direct encounters with reality but from participation in a vast web of signification. Meaning is produced relationally rather than representationally.

However, Derrida's framework also reveals the limitations of purely structural accounts. If meaning consists entirely of relations among signs, then the distinction between meaningful discourse and mere statistical patterning becomes difficult to maintain. AI succeeds precisely because linguistic systems contain enormous amounts of relational structure. Yet something seems missing. Human language is not merely differential; it is embedded within practices of action, perception, and correction that continually reconnect sign systems to worldly constraints.

The lesson, therefore, is not that AI lacks meaning. Rather, AI demonstrates that much of what we call meaning is indeed structural and relational. What remains unresolved is whether relational structure alone can generate genuine aboutness.

III. Structural Realism and the Reconstruction of Truth

The AI challenge cannot be resolved by returning to naïve correspondence theories of truth. Classical correspondence theory assumes that propositions are true when they accurately mirror independently existing facts. While intuitively attractive, this model presupposes a transparency between language and reality that both analytic and continental traditions have increasingly called into question.

Scientific history provides ample reason for skepticism. Scientific theories frequently undergo radical conceptual transformations. The entities posited by earlier theories are often abandoned, revised, or reconceptualized. Yet scientific progress remains undeniable. The predictive success of science cannot be dismissed as mere coincidence.

Structural realism emerges as an attempt to preserve realism while avoiding naïve representationalism. Associated with philosophers such as John Worrall and informed by earlier insights from Henri Poincaré, structural realism argues that scientific success is best explained by the fact that theories correctly capture structural relations within reality even when their ontological commitments change.

The significance of this position for AI is profound.

LLMs excel at identifying and reproducing relational structures. They learn patterns connecting concepts, arguments, and linguistic forms. In many respects, they function as extraordinarily sophisticated engines of structural mapping. They can model the architecture of discourse with unprecedented precision.

Yet this very success reveals the limits of structural competence. AI tracks structures within human discourse, but discourse itself is not identical to reality. The model captures relationships among signs rather than relationships between signs and the world. It learns the structure of our representations without independently testing those representations against material resistance.

This distinction allows structural realism to preserve an important realist intuition. Reality need not be directly mirrored by language. Nevertheless, reality constrains which structures survive. Scientific concepts endure because they are repeatedly subjected to experimental correction. Linguistic practices remain meaningful because they are embedded within forms of life that encounter success and failure.

Truth, therefore, is neither transparent correspondence nor pure social construction. It emerges through historically situated practices that successfully track structural features of reality. AI can model these practices, but it does not itself participate in the processes through which reality imposes constraints upon them.

IV. The Foucault Problem: Power, Visibility, and the Limits of Structural Realism

Despite its strengths, structural realism faces a significant challenge from Michel Foucault. Structural realism explains how theories can track objective structures, but it often remains silent regarding how certain structures become visible in the first place.

Foucault's analyses of knowledge and power reveal that access to reality is always mediated through institutions, discourses, and regimes of authority. What counts as evidence, rationality, expertise, or truth is never determined solely by reality itself. These categories emerge through historically contingent configurations of power.

The implication is not that reality is socially constructed in its entirety. Foucault does not claim that diseases, bodies, or physical phenomena are inventions of discourse. Rather, discourse shapes how such phenomena become intelligible, categorized, and governed.

This insight becomes especially important when considering AI systems. LLMs learn from textual corpora produced within existing social institutions. Consequently, they inherit not only linguistic structures but also the power relations embedded within those structures. The model learns distinctions, classifications, and assumptions authorized by particular historical regimes of knowledge.

A purely structural realist account risks overlooking this dimension. If we focus exclusively on the successful tracking of structures, we may fail to ask why some structures receive attention while others remain invisible. We may mistake historically contingent categories for objective necessities.

Yet Foucauldian analysis also encounters limits. If all structures are interpreted primarily as effects of power, it becomes difficult to explain the persistent success of scientific inquiry. Power alone cannot account for why certain theories reliably predict eclipses, cure diseases, or enable technological innovation. Reality continues to resist discourse in ways that exceed institutional authority.

The most promising position, therefore, combines structural realism with Foucauldian critique. Reality constrains discourse, but discourse shapes access to reality. Structures are real, but our awareness of them is mediated through historically situated institutions. Knowledge emerges not from a direct encounter with reality nor from the autonomous operation of discourse but from the ongoing interaction among material constraints, linguistic structures, and relations of power.

V. Objection: Are Humans Just Better Language Models?

The strongest objection to the argument developed thus far is that it may rely upon an increasingly unstable distinction between human and machine cognition. Throughout this paper, I have argued that AI lacks genuine reference because it does not participate in the embodied, institutional, and material practices through which linguistic communities remain accountable to reality. Yet a critic might reasonably ask whether this distinction merely preserves a residual human exceptionalism rather than identifying a genuine philosophical difference.

Recent developments in machine learning have made this challenge difficult to dismiss. Large Language Models acquire linguistic competence not through explicit symbolic rules but through exposure to enormous quantities of linguistic data. In this respect, they appear to learn language in a manner surprisingly similar to human beings. Human children are not born with direct access to reference relations. They acquire language through immersion in social practices, repeated exposure to linguistic regularities, and participation in communities of speakers. If both humans and AI develop competence by internalizing patterns distributed throughout linguistic environments, then what remains of the distinction between human intentionality and algorithmic processing?

This objection gains further force when viewed through the lens of Wittgenstein. If meaning is use, and if linguistic competence consists in mastering socially established language games, then AI appears increasingly capable of satisfying the relevant criteria. Indeed, many of the traditional markers of understanding—context sensitivity, inferential flexibility, and conversational coherence—are precisely the capacities that contemporary models exhibit. The danger is that appeals to intentionality begin to resemble appeals to a metaphysical remainder, invoked whenever machines successfully perform tasks once considered uniquely human.

A similar challenge emerges from Derrida's account of différance. If meaning arises through relations among signs rather than through direct access to self-present referents, then human language itself may already function according to principles that resemble computational pattern recognition. The subject no longer appears as the sovereign source of meaning but as a participant within a larger system of signification. From this perspective, the distinction between human and machine language users risks becoming one of degree rather than kind.

There is also a compelling naturalistic argument to consider. Human cognition is deeply predictive. Contemporary cognitive science increasingly describes perception, reasoning, and language as processes of probabilistic inference. Human beings continuously generate expectations about the world and revise those expectations in response to error signals. In broad structural terms, this process bears a striking resemblance to the statistical optimization procedures employed by machine learning systems. If both humans and machines operate through forms of predictive pattern recognition, then the claim that AI merely manipulates statistics while humans possess genuine understanding becomes difficult to sustain.

Yet despite the force of these objections, an important distinction remains. The difference between humans and current AI systems is not that humans somehow transcend structure while machines are trapped within it. Both humans and machines operate through structures. The critical difference concerns the source and function of correction.

Human beings do not merely process linguistic patterns; they inhabit environments capable of resisting those patterns. Our concepts are continually tested through action. We make predictions, encounter consequences, revise beliefs, and alter behavior. Scientific inquiry, technological intervention, and everyday practical activity all expose human conceptual frameworks to forms of failure generated by the world itself. A child learns the meaning of "hot" not only through linguistic instruction but through embodied encounters with heat. A scientist learns the limits of a theory through experiments that refuse to behave as expected. Human cognition develops within networks of feedback that extend beyond language and into material reality.

Current AI systems, by contrast, remain largely confined to derivative forms of correction. They learn from records of humanity's encounters with reality rather than from reality itself. Their training data contain the sedimented traces of countless acts of worldly engagement, but the model does not independently participate in those engagements. It inherits the results of correction without undergoing correction in the same sense. Even when integrated into robotic systems or interactive environments, AI remains dependent upon human institutions to define the significance of success, failure, and error.

This distinction can be understood through Putnam's semantic externalism. Meanings are not solely internal states; they emerge through relationships between linguistic communities and the world they inhabit. Humans participate directly in those relationships. AI participates indirectly, through the textual residues generated by communities already engaged in them. The model occupies a secondary position within the ecology of reference.

Nevertheless, the objection should not be dismissed entirely. The rapid development of embodied AI, autonomous agents, and multimodal learning systems suggests that some traditional distinctions may erode over time. If future systems become capable of sustained interaction with physical environments, institutional participation, and autonomous revision of conceptual frameworks in response to material resistance, the gap between human and machine reference may narrow considerably. The possibility cannot be ruled out in advance.

What AI reveals, therefore, is not the uniqueness of human cognition but the conditions under which reference becomes possible. The philosophical lesson is not that humans possess a mysterious essence absent in machines. Rather, it is that meaning, intentionality, and reference emerge within systems capable of being corrected by something beyond themselves. Human beings currently satisfy this condition because they are embedded within embodied, social, and material practices that expose their concepts to failure. Whether future AI systems might eventually satisfy similar conditions remains an open question.

The challenge posed by AI is thus not whether machines can imitate language. They clearly can. The deeper challenge is whether participation in structures of correction, accountability, and worldly resistance is necessary for genuine reference. If it is, then the distinction between human and machine cognition remains philosophically significant. If it is not, then AI may force us to radically revise what we have long meant by meaning itself.

VI. The Politics of Corrigibility: AI, Governance, and the Automation of Structure

If the preceding analysis has established that AI can reproduce the structures of meaning without participating fully in the practices that ground reference, then the philosophical problem extends beyond semantics into politics. The question is no longer merely how AI understands the world, but how AI increasingly participates in governing it. As artificial intelligence moves from a system of linguistic production to a system of institutional decision-making, the stakes shift from epistemology to ethics and political power.

The synthesis developed thus far—combining structural realism with Foucauldian critique—reveals a central tension. Structural realism suggests that successful systems track stable relational structures. Foucault reminds us that the visibility of those structures is always mediated through historically contingent institutions and regimes of power. Applied to artificial intelligence, this tension generates a troubling possibility: AI may become extraordinarily effective at reproducing existing social structures while remaining incapable of distinguishing between structures that reflect genuine constraints of reality and structures that merely reflect the accumulated effects of historical domination.

The Structural Realist Trap: When History Appears as Nature

The first ethical danger emerges from what may be called the problem of structural reification. Structural realism argues that explanatory success often results from accurately tracking relational structures. Yet AI systems do not encounter reality directly. They encounter archives, datasets, texts, and records generated by human institutions. Consequently, the structures identified by AI are not necessarily structures of reality itself; they are frequently structures of historical representation.

This distinction is ethically significant.

A scientific model that identifies the inverse-square law of gravity is tracking a relatively stable feature of the physical world. A language model that identifies patterns connecting socioeconomic status, race, gender, education, or criminality may be identifying something very different: the sedimented traces of institutional history. Yet both appear within the model as statistically significant structures.

The danger is that AI possesses no intrinsic mechanism for distinguishing between these categories. It recognizes regularities but not their ontological status. As a result, historically contingent arrangements can acquire the appearance of objective necessity.

The ethical problem is therefore not merely bias. Bias implies deviation from a neutral baseline. The deeper issue concerns the transformation of contingent social relations into apparently natural facts. Because algorithmic systems derive authority from their mathematical sophistication, the structures they reproduce often appear objective even when they are products of historical inequality.

This creates what may be called the structural realist trap. The more successful AI becomes at identifying patterns, the greater the temptation to interpret those patterns as revelations of reality itself. The machine transforms history into nature.

A post-modern structural realist must reject this move. The existence of a stable pattern does not establish the reality of the category through which the pattern is described. Statistical regularity alone cannot determine whether a structure reflects material constraints, institutional arrangements, or temporary historical conditions. Structural tracking must therefore be supplemented by critical interpretation.

The Foucauldian Panopticon: Classification as Governance

Foucault's work allows us to understand why this distinction matters politically. For Foucault, modern power rarely operates through direct coercion alone. Instead, it functions through classification, normalization, and administration. Institutions produce knowledge about populations, and this knowledge becomes a mechanism for governing behavior.

Artificial intelligence intensifies this process.

Contemporary AI systems increasingly participate in decisions involving employment, credit allocation, insurance pricing, content moderation, educational assessment, predictive policing, and risk evaluation. In each case, the system does more than describe reality. It sorts individuals into categories and assigns probabilities to future actions. These classifications influence opportunities, resources, and social standing.

The significance of this process is often obscured by the language of prediction.

Predictions appear passive. Governance appears active.

Yet algorithmic predictions frequently function as forms of governance because they shape the institutional responses that individuals encounter. A risk score affects whether a loan is approved. A recommendation system affects what information becomes visible. A content moderation algorithm affects what forms of speech become socially legible.

From a Foucauldian perspective, these systems operate as regimes of truth. They establish categories, determine norms, and define acceptable forms of conduct. Their authority derives not from democratic legitimacy but from technical expertise.

The contemporary AI system thus resembles a transformed panopticon. The essential feature of the panopticon was never surveillance alone. It was the internalization of norms through continuous visibility and categorization. Individuals adapted themselves to institutional expectations because they knew they were being measured.

AI extends this logic. The system does not merely observe behavior; it increasingly participates in constructing the criteria by which behavior is evaluated. It determines what counts as relevance, credibility, risk, productivity, engagement, or success.

The ethical issue is therefore not that AI watches us. It is that AI helps determine what kinds of persons we are permitted to become.

The Crisis of Corrigibility

The convergence of structural realism and Foucauldian critique points toward a deeper problem: the crisis of corrigibility.

Throughout this paper, I have argued that human systems of knowledge remain connected to reality because they are exposed to failure. Scientific theories can be falsified. Institutions can be challenged. Social categories can be revised. Language remains meaningful because it operates within practices capable of correction.

AI complicates this dynamic.

First, AI systems frequently operate through forms of statistical optimization rather than direct engagement with material reality. When an output is challenged, the system does not experience failure in the manner of a scientific hypothesis confronted by contradictory evidence. It merely registers divergence from desired outcomes. The distinction matters because statistical correction is not identical to epistemic correction. A model can become more predictive without becoming more truthful.

Second, the institutions that develop and deploy advanced AI systems are often insulated from public scrutiny. Decisions regarding training data, optimization objectives, classification frameworks, and deployment strategies frequently occur within private organizations. The result is an unusual concentration of epistemic authority.

The problem, therefore, is not that AI lacks intelligence. The problem is that AI increasingly participates in social governance while remaining only weakly connected to mechanisms of democratic correction.

In this sense, contemporary debates about AI alignment may be philosophically inadequate. Alignment discourse often treats ethics as a technical optimization problem: how can systems be engineered to produce desirable outputs? While important, this approach risks reducing political questions to engineering questions.

The deeper challenge concerns who determines what counts as a desirable outcome in the first place.

Corrigibility is therefore not merely a technical property. It is a political one. A genuinely corrigible system must remain open to contestation, revision, and public accountability. It must permit the categories through which it interprets reality to be challenged rather than merely optimized.

Toward a Politics of Corrigibility

The ethical imperative that emerges from this analysis is the construction of a politics of corrigibility.

From the perspective of structural realism, this requires structural transparency. Institutions deploying AI systems should be able to explain what patterns are being tracked, what categories are being utilized, and what assumptions are embedded within those categories. The structures identified by the system should not appear as self-evident features of reality but as historically situated interpretations subject to revision.

From a Foucauldian perspective, this requires institutional accountability. Systems that classify individuals, allocate opportunities, or shape public discourse exercise forms of social power. Such power should be subject to legal oversight, democratic review, and public contestation. If AI participates in the production of truth, then it cannot remain outside the institutions responsible for regulating truth claims.

The goal is not to eliminate structure. Nor is it to abandon realism. Rather, it is to ensure that the structures through which reality becomes intelligible remain open to correction from both the world and the communities affected by their application.

Conclusion: Keeping the World Exit-able

The ethical lesson of AI is not that machines are becoming too intelligent. It is that they are becoming socially authoritative.

The central danger is the naturalization of historically contingent structures. AI systems transform accumulated patterns into predictive frameworks, and predictive frameworks easily become normative frameworks. What begins as description becomes administration. What begins as probability becomes governance.

To resist this process, we must refuse to treat algorithmic outputs as transparent reflections of reality. The structures identified by AI are always interpretations before they are facts. They are products of histories, institutions, and relations of power before they become objects of technical analysis.

The ethical task, therefore, is not merely to improve AI accuracy. It is to preserve the conditions under which AI itself can be corrected. A democratic society must ensure that algorithmic systems remain answerable to forms of resistance external to their own operations: empirical resistance from reality, political resistance from affected communities, and institutional resistance from mechanisms of public accountability.

The ultimate ethical question is whether the world remains capable of interrupting the structures we build to describe it.

A corrigible society is one in which the answer remains yes.

To keep AI accountable is to keep reality capable of speaking back. It is to preserve the possibility that our categories, predictions, and models may fail. It is to keep the signs open, the structures revisable, and the world exit-able.

Conclusion: The Constraint of the Real

The rise of generative AI has exposed fundamental assumptions underlying modern philosophy of language. The traditional opposition between meaning and reference, semantics and intentionality, realism and constructivism no longer appears as stable as it once did. AI demonstrates that highly sophisticated linguistic competence can emerge from the manipulation of relational structures alone. It reveals the extent to which meaning depends upon systems of difference, use, and historical transmission.

Yet AI also reveals the limits of purely structural accounts. Human language is not simply a network of signs. It is embedded within institutions, practices, and forms of life that remain accountable to a resistant world. Reference is not secured merely by participation in linguistic structures but through participation in practices that continually expose those structures to correction. Truth is not the mirroring of reality, but neither is it the free play of signification.

The deepest lesson of AI may be that intentionality and reference are not mysterious inner properties but socially and materially sustained achievements. We are "about" the world because the world continually interrupts our descriptions of it. We encounter error, resistance, and failure. We revise our concepts because reality refuses complete assimilation into discourse.

AI possesses the structure of these practices but not their consequences. It inherits the sedimented traces of humanity's engagement with reality without sharing in the accountability that produced them. It is, in this sense, a system of signs without an exit. The model can reproduce the architecture of truth, but truth itself emerges only where discourse remains corrigible before a world that exceeds it. The AI has the structure; it lacks the failure. And it is precisely failure—our inability to perfectly capture reality—that makes reference, meaning, and truth possible.

Works Cited

Brentano, Franz. Psychology from an Empirical Standpoint. Routledge, 1995.

Derrida, Jacques. Of Grammatology. Translated by Gayatri Chakravorty Spivak, Johns Hopkins University Press, 1976.

Foucault, Michel. Power/Knowledge: Selected Interviews and Other Writings, 1972–1977. Edited by Colin Gordon, Pantheon Books, 1980.

Kripke, Saul A. Naming and Necessity. Harvard University Press, 1980.

Putnam, Hilary. "The Meaning of 'Meaning.'" Mind, Language and Reality: Philosophical Papers, Volume 2, Cambridge University Press, 1975, pp. 215–271.

Searle, John R. "Minds, Brains, and Programs." Behavioral and Brain Sciences, vol. 3, no. 3, 1980, pp. 417–457.

Wittgenstein, Ludwig. Philosophical Investigations. 4th ed., Wiley-Blackwell, 2009.

Worrall, John. "Structural Realism: The Best of Both Worlds?" Dialectica, vol. 43, no. 1–2, 1989, pp. 99–124.

Human language operates through differential structures (A1).
Those structures remain referential only insofar as they are accountable to the world (A2).
Truth emerges where claims can fail and be revised (A3).
AI reproduces the structure of such practices without sharing their conditions of failure (A4).
Therefore, AI can model meaning and simulate reference while lacking the full conditions that ground truth.
Reality constrains discourse even though discourse never transparently mirrors reality.

Q.E.D.

Meta-Theorem of the Entire Paper

Axiom 1: Meaning is structurally generated.
Axiom 2: Reference is socially mediated.
Axiom 3: Truth is constrained by reality.
Axiom 4: Power shapes access to reality.
Axiom 5: AI inherits linguistic structures without fully sharing the conditions that produced them.

Theorem: Generative AI demonstrates that meaning can be produced through relational structures alone, but it does not show that reference, truth, or intentionality are reducible to statistical relations. Rather, these emerge from historically situated systems of linguistic practice that remain corrigible before a resistant reality.

Q.E.D.

Friday, June 5, 2026