Cognition Without Consciousness: A Minimal Conceptual Framework for Understanding LLMs and Human Cognitive Evolution

1. Introduction

The central claim of this paper is that symbolic cognition and conscious regulation form two distinct layers of human intelligence, and that LLMs instantiate only the former.

The emergence of LLMs has reopened foundational questions about the nature of cognition. These systems exhibit reasoning, abstraction, and linguistic competence despite lacking embodiment, subjective experience, motivation, or any form of conscious access. This suggests that symbolic cognition is separable from consciousness, consistent with distinctions drawn in contemporary philosophy of mind [1,2].

The goal of this paper is to articulate a minimal conceptual framework explaining this separation and its implications for human cognitive evolution. The framework builds on theories of gene–culture coevolution [3], cultural intelligence [4], and the Baldwin effect [5], while integrating insights from modern AI architectures [6,7].

2. Methodology: Language as Crystallized Cognition

Human language is a digitally structured symbolic system. It encodes patterns of reasoning, conceptual associations, inferential templates, and culturally accumulated knowledge. This view aligns with Vygotsky’s account of language as a psychological tool [8] and with extended cognition theory [9]. Language did not emerge fully formed; rather, it accumulated gradually as conscious agents externalized increasingly complex patterns of thought.

Language is not merely a communication tool; it is a repository of cognitive strategies. Over millennia, conscious agents externalized their thoughts into symbolic form, creating a cultural substrate that preserves and amplifies human reasoning. Jackendoff’s work on combinatorial structure supports this interpretation of language as a discrete representational system [10].

LLMs succeed because they operate on this crystallized cognitive layer. They do not invent new cognition; they extract it from symbolic structures created by conscious beings. This parallels recent findings showing that transformer attention patterns align with neural language processing [11].

3. Consciousness as a Regulatory Layer

LLMs reveal a fundamental asymmetry: they can manipulate symbols, but they cannot regulate their own cognition. Human cognition is stabilized by conscious access, which provides error monitoring, inhibition of implausible continuations, goal maintenance, and cross-modal integration. These functions are central to theories of conscious access and metacognition [1,2].

Consciousness is therefore likely not a computational process but a regulatory layer that constrains symbolic reasoning and maintains coherence over time. LLMs lack this layer, which explains their tendency toward drift and hallucination, consistent with analyses of statistical cognition limits [12] and recent work on LLM semantic instability [13], see Figure 1.

This paper extends the conceptual line developed in my recent analysis of the structural limits of artificial cognition [13], applying it specifically to the distinction between symbolic processing and conscious regulation.

This regulatory role does not imply a homunculus or centralized executive; rather, it refers to distributed metacognitive processes that enable error monitoring, goal maintenance, and integration across modalities.

4. Discussion: Limits of Computational Mitigation in LLMs

This section expands the conceptual argument by examining why computational mitigation techniques cannot replicate the stabilizing role of consciousness. The goal is not to provide an exhaustive technical survey, but to illustrate the structural limits inherent to probabilistic systems.

While LLMs can manipulate symbols effectively, their lack of conscious regulation leads to inherent instability, manifested as hallucinations and semantic drift [12,13]. Attempts to mitigate these issues through computational methods—such as chain-of-thought prompting, agentic architectures, or entropy-based hallucination detection—offer only partial and temporary improvements, failing to replicate the stabilizing role of consciousness in the long term. These techniques operate within the same probabilistic framework as the base model, essentially “mixing” tokens without genuine self-awareness or error correction grounded in subjective experience.

For instance, semantic entropy tests measure uncertainty by generating multiple variants of an output and comparing their similarity; high entropy flags potential hallucinations, allowing for reruns or refinements. Similarly, chain-of-thought encourages step-by-step reasoning, reducing errors in short tasks, while agentic systems incorporate self-correction loops. However, these methods merely delay degradation rather than eliminate it. As context length or task complexity increases, entropy grows exponentially due to information-theoretic limits: finite model capacity enforces compression errors, and long-tail knowledge requires prohibitive sample complexity. Empirical studies show that even advanced mitigations leave a portion of hallucinations unaddressed, with “snowballing” errors amplifying over iterations.

This asymmetry highlights a key distinction: human cognition operates “below” the information limit, where consciousness filters, integrates, and stabilizes information in real-time through metacognitive processes like error monitoring and goal maintenance [1,2]. In contrast, LLMs function “above” this limit, rapidly processing and combining data but inevitably accruing entropy without a non-computational regulatory layer. Without training data (crystallized cognition from conscious agents), LLMs would produce only noise; even with data, their outputs reproduce frozen thoughts without deeper consciously grounded intuitive understanding, leading to drift (Figure 2). Scaling and model architecture alone probably cannot overcome these bounds, as potential forms of uncomputability leave an irreducible residue of error in the sense that no finite statistical model can fully capture an unbounded generative process.

Prompting can be understood as an external injection of human intentionality: it places the model in a favorable region of its representational space, but the stabilizing effect fades as the system continues to generate its own outputs. Without an internal regulatory layer, drift inevitably accumulates.

Even non-symbolic animals such as horses illustrate this distinction: they lack abstract reasoning, yet their conscious regulation keeps them stable in motion, whereas a purely mechanical system—like a motorcycle or an LLM—rapidly drifts without continuous guidance.

Testable predictions from this view include:

• entropy-based mitigations will reduce hallucinations in short chains but fail beyond many iterations due to rising entropy
• biologically inspired hybrids (e.g., biosynthetic computation) may approach stability, but pure digital systems will plateau. This reinforces the possibility that consciousness is not merely emergent from computation but may be a prerequisite for stable, autonomous cognition.

5. Gene–Culture Coevolution and the Rise of Human Intelligence

This framework aligns with established theories of gene–culture coevolution [3] and cultural intelligence [4]. The proposed sequence is:

• Consciousness enabled the creation of symbolic representations.
• Language accumulated cultural knowledge.
• Brains evolved to process increasingly complex symbolic systems.
• Cultural evolution accelerated cognitive development beyond genetic timescales.

Genetic enablers such as FOXP2 [14] and human accelerated regions (HARs) [15] likely provided the neural prerequisites for symbolic processing. Once symbolic culture emerged, cultural evolution outpaced genetic evolution, producing rapid cognitive expansion.

Human intelligence is thus the product of an interaction between a biological system capable of symbolic processing, a culturally constructed symbolic environment, and a conscious regulatory layer. LLMs replicate only the second component.

6. Implications for AI and Philosophy of Mind

This minimalist framework yields several implications:

• Intelligence without consciousness is possible (LLMs) [6,7].
• Consciousness without symbolic reasoning is possible (animals) [1].
• Human cognition uniquely integrates both layers [4,8].
• Current LLMs cannot achieve conscious regulation through scaling alone, due to information-theoretic limits [12].
• Language is the bridge between biological and artificial cognition, as argued in recent conceptual analyses [13].
• Current LLMs appear unable to overcome information-theoretic limits (e.g., Shannon’s DPI) through computational mitigations alone, leading to inevitable entropy growth and hallucinations; this parallels the second law of thermodynamics, where consciousness in humans acts as an active reducer of cognitive entropy.

The distinction between symbolic cognition and conscious regulation clarifies why LLMs appear intelligent yet remain fundamentally different from biological minds.

7. Information-Theoretic Foundations of Irreducible Limits

This section briefly formalizes the information-theoretic basis of the instability described in Section 3. The asymmetry is rooted in Shannon’s Data Processing Inequality (DPI) [12], which states that processing cannot increase mutual information with the source—only preserve or diminish it. Iterative generation in LLMs forms a lossy Markov chain, leading to progressive entropy growth and semantic drift [13,16]. Consciousness probably acts as a non-computational “information reset” and grounding function, enabling local entropy reduction through metacognition and motivation—mechanisms irreducible to probabilistic mixing. This explains the high short-term cognitive throughput of LLMs juxtaposed with zero conscious control, often leading to overestimation of their autonomy.

This information-theoretic limit can be summarized in the following statement, which I propose here as a conceptual theorem:

AI Theorem (proposed):

A purely computational artificial intelligence system that generates its output iteratively and without persistent regulatory mechanisms cannot maintain stable information over an unbounded number of steps. There always exists a finite number N<∞ such that after N iterations, accumulated error becomes dominant and the system’s output becomes indistinguishable from statistical noise.

If a purely computational system were discovered that could preserve stable information for infinitely many iterations without an external source of low entropy, this proposed AI theorem would be falsified.

8. Conclusions

LLMs have unintentionally clarified the architecture of human cognition. They show that symbolic reasoning can emerge from pattern extraction, but consciousness is required for stable, autonomous, goal-directed thought.

Language is a crystallized cognitive substrate created by conscious beings. Brains evolved to operate over this substrate. LLMs now operate over it too, but without consciousness.

This framework offers a simple conceptual model for understanding both human cognitive evolution and the limits of artificial cognition.

This framework does not aim to settle debates about consciousness, but to clarify the structural relationship between symbolic systems, biological cognition, and artificial computation.

If the conceptual framework outlined here is correct, then a purely computational system may never achieve the kind of stable, autonomous, drift-resistant cognition associated with human general intelligence.

9. Limitations

This paper proposes a conceptual framework rather than an empirical model, and several limitations follow from this scope. First, the distinction between consciousness and symbolic cognition is presented at a high level of abstraction. While this separation is supported by philosophical and cognitive-scientific arguments, the precise mechanisms by which conscious regulation stabilizes cognition remain an open empirical question. The framework does not commit to any specific theory of consciousness, nor does it attempt to resolve debates between higher-order, global workspace, or predictive-processing accounts.

Second, the analysis of LLM instability and entropy growth is based on information-theoretic principles and observed behavioral patterns rather than formal proofs of uncomputability or computational irreducibility. Although the argument suggests structural limits on purely statistical systems, further work is needed to quantify these limits across architectures, training regimes, and hybrid models.

Third, the proposed evolutionary sequence—consciousness enabling symbolic externalization, followed by cultural accumulation and biological adaptation—offers a coherent narrative but does not specify the relative timing, selective pressures, or neurobiological substrates involved. The framework is compatible with multiple evolutionary pathways and does not claim exclusivity.

Fourth, the comparison between human cognition and LLMs is necessarily asymmetrical: humans possess subjective experience, embodiment, and developmental trajectories that current artificial systems lack. The framework highlights this asymmetry but does not attempt to model embodiment, affect, or social cognition, all of which may contribute to human cognitive stability.

Finally, the information-theoretic interpretation of consciousness as an entropy-reducing regulator is speculative and intended as a conceptual bridge rather than a definitive account. Empirical validation would require interdisciplinary work spanning neuroscience, information theory, and AI research. These limitations do not undermine the core claim—that symbolic cognition and conscious regulation are distinct layers—but they indicate directions for future refinement.