The Prison House of Pattern Matching: What Philosophy Already Knew About LLMs

Everything you wanted to know about LLMs but were afraid of philosophy.

Introduction

The current excitement surrounding Large Language Models (LLMs) brings us to a philosophical juncture—one that recapitulates old debates about the nature of understanding and intelligence. As researchers and tech evangelists herald LLMs as steps toward artificial general intelligence (AGI), they inadvertently mirror philosophical critiques that have long exposed the limitations of symbolic manipulation.

This essay argues that the philosophical insights of the 20th century—especially those arising from structuralism, the problem of induction, and theories of meaning—anticipate the constraints of LLMs. By examining these systems through the lens of thinkers such as Saussure, Goodman, and Schaffer, we can better understand why current models fall short of genuine intelligence and how the integration of causal reasoning and grounded architectures might point the way forward.

The Saussurean Mirror: LLMs as Radical Structuralism Incarnate

In generating language, LLMs enact a radical form of Saussurean linguistics. Ferdinand de Saussure proposed that meaning arises from the relational structure of language, where signs are defined not by their intrinsic properties but by their differences from other signs. While Saussure acknowledged the social and historical dimensions of language, LLMs embody the most extreme interpretation of his structuralist vision: they operate within a self-contained system of statistical relationships between tokens.

Take, for instance, an LLM describing the taste of an apple. The model can generate plausible phrases—“sweet,” “crisp,” or “tart”—by statistically associating these descriptors with “apple.” Yet, as eloquent as its response may be, the model is akin to a scholar imprisoned in a library, endlessly reshuffling books without ever stepping into an orchard. This purely syntactic manipulation underscores a profound absence: the semantic tethering of language to sensory and experiential reality.

In this way, LLMs render explicit what structuralism left implicit: meaning as a purely internal affair, disconnected from external reality. This reduction reveals both the power and the limitations of LLMs—power in their ability to emulate linguistic forms, and limitation in their inability to ground those forms in anything but patterns.

Goodman’s Grue: The Problem of Projection in Pattern Learning

Nelson Goodman’s grue paradox raises a profound challenge for inductive reasoning: the question of projection. To project a pattern means to extend it from observed cases to unobserved ones, effectively deciding how past experience informs future expectations. Goodman’s insight was to show that any finite set of observations can support infinitely many patterns of projection. For example, an emerald that has always appeared green can be said to follow the predicate “green” (it will always be green) or the predicate “grue” (it is green before a certain time t but blue thereafter). Both projections fit the data equally well, yet only one aligns with our intuitions.

Goodman’s problem is not merely semantic; it strikes at the heart of inductive inference. Why do we choose some patterns over others? What principle legitimises the projection of one predicate (e.g., “green”) while rejecting others (e.g., “grue”)? Goodman argued that this choice is governed not by the data alone but by tacit conventions and shared expectations—factors that lie outside the observed patterns themselves. (Hence Goodman’s focus on aesthetic and world-making.)

When applied to LLMs, Goodman’s paradox reveals a critical limitation: these models excel at learning patterns from training data but lack an intrinsic basis for deciding which projections are valid. In other words, when an LLM generates text, why does it project this set of linguistic relationships rather than countless other statistically compatible alternatives? The paradox highlights that what LLMs call “learning” is, in fact, a constrained form of statistical generalisation, heavily shaped by the biases of their training environment.

By introducing Goodman’s grue paradox here, I want to underscore a key philosophical critique of LLMs: they cannot justify their projections beyond the statistical frequencies encoded in their training data. This inability to discriminate between natural and contrived patterns reveals why LLMs, despite their sophistication and undeniable utility, lack the deeper conceptual grounding necessary for genuine understanding. Their outputs are shaped by external design choices and human validation, not by an inherent ability to evaluate which patterns make sense in a broader, causal context.

I put Goodman’s paradox to work here in order to:

1. Expose the limits of pattern learning: It shows that LLMs cannot transcend the patterns encoded in their data to arrive at independent, justified inferences about the world.

2. Clarify the need for grounding: It underscores the importance of external frameworks—whether causal, hierarchical, or experiential—that can prioritise natural projections over contrived ones.

3. Point toward architectural gaps: By situating LLMs within Goodman’s critique, we see the necessity of augmenting pattern-matching systems with mechanisms for causal reasoning and world interaction.

Thus, Goodman’s grue paradox is useful not only for critiquing the current state of LLMs but also deepens our understanding of what’s missing: a principle for distinguishing between arbitrary statistical fits and meaningful generalisations grounded in reality.

The Grounding Hierarchy: Schaffer Meets LeCun

Jonathan Schaffer’s metaphysical hierarchy offers a provocative lens for understanding LLMs’ limitations. In his wonderful essay, "On What Grounds What”, Schaffer posits that reality is structured hierarchically, with some entities more fundamental than others. Meaning, in this view, depends on grounding: higher-level abstractions must be anchored in lower-level realities.

Yann LeCun’s critique of current AI architectures resonates with Schaffer’s framework. LeCun argues that true intelligence requires hierarchical world models, starting with basic physical principles and advancing to abstract reasoning. By contrast, LLMs operate in reverse: they begin at the highest level of abstraction—language—without grounding it in more fundamental levels such as perception, object permanence, or causation.

This disconnect explains why LLMs struggle with tasks requiring embodied understanding. Without grounding language in a reality that includes physical interaction and causal relationships, their outputs remain unmoored, a facsimile of meaning that lacks the ontological depth to bridge words with the world.

The Circular Logic of Validation

The problem of grounding extends to how we evaluate LLM “intelligence.” The validation process is deeply circular:

1. LLMs are trained on human-generated text.

2. They internalise the statistical patterns within this text.

3. Their outputs are deemed “intelligent” when they align with these patterns.

4. We validate their intelligence by checking against the training set’s patterns.

This self-referential logic mirrors the pitfalls of post-structuralism, where the claim “all is language” undermines its own grounding. Similarly, the claim “all is pattern-matching” raises the question: what validates the patterns themselves? Without an external criterion—such as causal engagement with the world—the validation collapses into a tautology.

Breaking Out: Pearl’s Causal Revolution

To transcend the prison house of pattern matching, we must heed Judea Pearl’s insights on causality. Pearl argues that intelligence requires navigating a “ladder of causation” that progresses from:

1. Association: Detecting correlations.

2. Intervention: Understanding how actions alter outcomes.

3. Counterfactual reasoning: Imagining alternative scenarios.

LLMs operate at the first rung, correlating tokens without causal understanding. Pearl’s framework suggests that genuine intelligence demands systems capable of intervention and counterfactual reasoning—abilities that require interaction with the world, not just its linguistic representation.

The Way Forward: Hybrid Architectures and Multiple Ways of Knowing

Philosophy’s critique of LLMs does not merely expose their limitations; it points toward better paths forward. Future AI systems must integrate diverse forms of reasoning, grounded in hierarchical models of reality and novel neural network architectures. This requires:

1. Hierarchical World Models

• Ground language in perception and physics.

• Respect metaphysical dependencies between levels of abstraction.

• Build understanding incrementally, from sensory input to symbolic reasoning.

2. Causal Reasoning Systems

• Move beyond correlation to causal inference.

• Enable interaction with and intervention in the world.

• Support counterfactual thinking to explore alternative possibilities.

3. Multiple Types of Understanding

• Combine statistical pattern recognition with causal reasoning.

• Integrate symbolic manipulation with embodied knowledge.

• Ground abstract thought in the concrete realities of physical experience.

Conclusion

The enthusiasm for LLMs, and make no mistake, I’m an enthusiast, often mistakes linguistic mimicry and simulation for genuine understanding. This mirrors philosophical debates long explored by structuralists, metaphysicians, and logicians. While LLMs represent remarkable achievements in language processing, treating them as steps toward AGI risks repeating the same error: conflating sophisticated symbol manipulation with intelligence. It needs a more robust and grounded architecture.

The path forward lies not in perfecting pattern-matching systems but in integrating them into architectures that respect the hierarchical and causal nature of reality. By grounding symbolic manipulation in physical understanding and embedding multiple forms of reasoning, we might move closer to systems that genuinely understand.

Philosophy, far from being an idle critique, offers a guide to this transformation. By taking seriously its insights into meaning, causation, and grounding, we can build AI systems that escape the prison of pattern matching and ascend toward genuine intelligence.