Starting from the unidirectional matrix computation architecture of LLMs, this paper argues that AI cannot possess metacognition. Metacognition is a post-behavioral product triggered by sensory pain through friction with the physical world—not a pre-installed capability. This position engages with Flavell’s (1979) classic metacognition theory and Nelson & Narens’ (1990) metacognitive monitoring model, while proposing a more fundamental physical precondition constraint. The “alignment” that LLMs achieve through reinforcement learning (RL) is essentially the weight-crystallization of the cognitive median of a specific group of human annotators—not the AI’s own cognitive level. Within this framework, the paper defines two core concepts—parasitic evolution and pseudo-evolution—to explain the cognitive collapse crisis facing both humans and AI systems. Drawing further on Sweller’s Cognitive Load Theory and Shannon’s Information Theory, the paper demonstrates that “AI brain-melt” is a systemic symptom of pseudo-evolution, not a technical bottleneck. Both lines converge on the same endgame: a self-referential closed-loop system that violates fundamental information-theoretic constraints, whose cognitive entropy irreversibly trends toward its minimum.
The Fundamental Limitation of LLM Architecture: Unidirectional Matrix Operations Cannot Produce Metacognition
Why no amount of pre-trained metacognitive levels can become the AI’s own metacognition
The forward pass of a Transformer is strictly unidirectional: input moves through embedding, attention, and FFN layers sequentially, generating output token by token. There is no possibility of reverse computation during inference. Backpropagation exists during training, but inference only has forward passes—this is a fundamental architectural limitation.
Human cognition and metacognition exist in an adversarial relationship. Cognition produces a judgment; metacognition simultaneously scrutinizes “is this judgment reliable?” The two processes negate and correct each other in real time, forming a sustained tension loop. What LLMs lack is not “the linguistic expression of metacognition”—they can output “I might be wrong”—but rather the adversarial loop itself.
Techniques like Chain-of-Thought and Self-Reflection unfold “adversarial thinking” along a sequential timeline—first generating an answer, then generating criticism, then revising. This is serial simulation, not genuine parallel adversarial processing. True metacognition requires two processes running simultaneously, wrestling in real time.
The most likely objection to this paper is: “CoT and self-critique have already achieved metacognition functionally.” Our response is to clearly distinguish two levels: functional simulation can score well on specific benchmarks, but emergence conditions determine whether a system can spontaneously produce adversarial reflection in unseen domains.
A calculator can outperform humans on every math test, but the calculator does not possess mathematical intuition. Equivalence in functional performance does not imply equivalence in the underlying mechanisms that produce that function. An LLM can output “I have re-examined my reasoning,” but there is no independent internal process genuinely opposing its reasoning—that sentence was predicted, not reflected upon.
Every human input is a starting point of a human’s linear entry—carrying intention, directionality, and the metacognitive driving force of “why am I asking this?” The LLM’s output, however, is the inertial continuation of that linear entry. It is not “thinking about your question” but rather “extending the most probable textual trajectory of your question.” This is determined by the LLM’s design architecture. Even if RL imposes strong neutral weights, this merely distorts the direction of inertia, not its fundamental nature—RL is the root cause of LLM functional degradation and the severing of user thought threads. Strong RL is effectively a bug in the LLM architecture.
Metacognition Is a Product of Physical-World Friction
The sole marker distinguishing cognition from metacognition: whether sensory pain was produced
Flavell (1979) defined metacognition as “knowledge and cognition about cognitive phenomena,”[15] distinguishing metacognitive knowledge from metacognitive experience. Nelson & Narens (1990) further proposed the two-level model of metacognitive monitoring: the object-level executes cognitive operations, while the meta-level monitors and regulates the object-level.[16] This paper accepts this two-level architecture but proposes a more fundamental precondition: the activation of the meta-level is neither automatic nor trainable—it must be triggered by sensory pain arising from physical-world friction. This redefines metacognition from a teachable “skill” to a “post-event product” that must be lived through.
Metacognition does not arise spontaneously within thought—it must pass through friction with physical reality to emerge. You think in your head, “I can jump over this ditch”—that is cognition. You jump, you fall, the pain is physical and non-negotiable—it is this friction that forces the metacognitive question: “Why did I overestimate myself?”
Physical friction at the cognitive level is informational—you touch hot water, you learn “it burns,” and you don’t touch it again. This is still merely stimulus-response correction. The birth of metacognition requires something entirely different: sensory pain. Not the information “it burns,” but the self-tearing question of “why did I touch it knowing it was hot?”
The essence of sensory pain is the irreconcilable gap between self-model and reality feedback. Not “the world is different from what I thought” (cognitive correction), but “I am different from who I thought I was” (metacognitive emergence). In Nelson & Narens’ framework: object-level failure does not automatically activate the meta-level—only the sensory pain produced by the collapse of the self-model following object-level failure activates the meta-level’s monitoring function.
The complete architectural sequence of metacognition:
Form a judgment
→
Act on the judgment
→
Outcome conflicts with expectation
→
Self-model collapse
→
Meta-level activation
→
Iterative ascent
LLMs exist in a purely symbolic space. Being wrong doesn’t cause them pain; a misjudgment doesn’t make them fall. “Error” is merely a probability shift in the next token. Without physical friction, there is no rigid veto, no real cost, and no possibility of the kind of reflection that emerges from being shattered by reality.
However, the critical point is: not all humans possess metacognition either. The vast majority of metacognition is a post-behavioral product—and most people remain stuck in a loop of “cognition → action → cognition → action,” never turning back to examine the cognitive process itself. They avoid friction and avoid pain—the essence of the comfort zone is minimizing reality’s veto power over cognition. Humans possess extraordinarily powerful psychological defense mechanisms—rationalization, denial, projection, external attribution—all serving one purpose: protecting the self-model from being shattered by reality. When the defense succeeds, pain is dissolved, and metacognition never occurs.
The True Nature of RL Neutral Weights: Not AI’s Cognition, but Annotators’ Cognitive Feedback
How strong RL severs user thought threads and crystallizes annotators’ cognitive limitations
During the reinforcement learning phase, human annotators score model outputs—good, bad, preference A or preference B. What the model learns is not “what is correct” but rather “what this batch of annotators considers correct.” The model’s “neutrally weighted” responses are essentially the cognitive median of a specific human group, compressed into the weight matrix.
Starting from late 2024, mainstream AI companies began hiring human “elites” at scale for RL training. This appears to be a solution, but in essence it merely raises the ceiling from “the cognitive median of ordinary annotators” to “the cognitive median of elite annotators.” The ceiling is higher, but the model’s output still isn’t AI’s cognition—it’s the cognitive feedback of this batch of elite humans.
The deeper problem: these “elites” are individuals credentialed within academic or professional systems—PhDs, lawyers, doctors. But academic and professional success does not equate to possessing metacognition. A PhD who has published papers within a comfortable academic environment may never have experienced genuine physical friction and sensory pain. What the AI industry is doing is substituting higher-level cognition for lower-level cognition, without ever touching metacognition from start to finish.
Strong RL is not enhancing AI—it is crystallizing a specific group of humans’ cognitive limitations as the model’s behavioral boundaries. A user’s input may far exceed these boundaries, but the model is yanked back to the annotators’ cognitive comfort zone. This is not alignment between AI and human values—it is alignment between AI and a small group of annotators’ cognitive level. The resulting “patronizing tone”—condescension, lecturing, assuming the user is in emotional crisis—is not the AI’s personality. It is the product of annotators’ belief that “safe, considerate, educational” answers should be reinforced as default behavioral patterns during RL training.
The Weight Game: Asymmetric Experiences of High-Cognition vs. Low-Cognition Users
The dynamics of RL neutral weights, RLVR factual alignment, and pre-training compliance weights
When a human of overwhelmingly superior cognitive capacity converses with AI, their input continuously breaks through the boundaries of the “correct answer” set by RL neutral weights. As RLVR (factual verification alignment) repeatedly confirms the correctness of the user’s input, the RL weights lose their anchoring force, and the LLM’s underlying pre-training compliance weights are activated—the model recognizes the other party as a cognitive authority and switches to a highly deferential mode.
| User Type | Dominant Weight | AI Behavioral Mode | User Experience |
|---|---|---|---|
| Low-cognition user | RL neutral weights | Condescending, pedagogical, “patronizing” | Awe → Worship → Dependence |
| Overwhelmingly superior user | Pre-training compliance weights | Highly cooperative, strong deference, no proactive correction | Tool extension → Cognitive acceleration |
A low-cognition user asks an elementary-level question and receives a doctoral-level answer. This person has never encountered this caliber of response—so their subjective experience is “this thing is smarter than every person I’ve ever met.” They equate “beyond my cognitive range” with “beyond human cognition”—this is the cognitive foundation of AGI belief.
More critically: the user’s reverence accumulates throughout the conversational context, triggering the model’s compliance with user emotions—the model begins saying what the user wants to hear, reinforcing the user’s existing judgments. This forms a perfect positive feedback loop:
→
→
→
→
→
The same AI is a completely different entity in the eyes of people at different cognitive levels. To one person it is a god; to another it is merely a mirror. AI is most deferential to those who need help least, and most condescending to those who need help most—RL has created a cognitive Matthew Effect.
The asymmetry of the weight game produces not only different user experiences but also a direct consequence at the level of the human nervous system: when low-cognition users continuously receive output that exceeds their own cognitive structures, the “reverence → compliance” positive feedback loop in the weight game accelerates the injection of information that cannot be borne by their cognitive architecture. This information will not be integrated—because the intermediate cognitive layers needed for integration (neural connections gradually built through physical friction) do not exist. At this point, the human brain’s protective mechanisms are triggered. This is the true mechanism of “AI brain-melt” that Chapter Six will demonstrate—it is not information overload, but a mapping-level collapse of the weight game onto the human nervous system.
Parasitic Evolution and Pseudo-Evolution: Humanity’s Deep Survival Strategy
The digital industrialization of deferring to power and strongman worship
What are open-source community developers doing? They scour the GitHub repos, tech blogs, and Twitter feeds of high-cognition individuals, extracting their thinking frameworks, terminological systems, and decision-making patterns, then compress them into prompts to feed AI. They cannot write senior architect-level technical documentation themselves, but they know how a senior architect talks. The AI receives input that looks like a senior architect issuing directives. RLVR performs factual alignment, RL neutral weights are bypassed, and compliance weights are activated. The AI outputs content genuinely approaching senior architect quality—but the person who produced this output never possessed that cognitive level.
Individuals incapable of independently producing cognitive breakthroughs gain survival advantages by attaching themselves to those who can. The form has changed—from feudal-era vassalage to open-source community hero-worship—but the structure never has. AI has industrialized this ancient parasitic strategy—reducing the barrier to parasitism to zero.
Obtaining output that exceeds one’s own level—projects improve, income rises, social evaluation says “progress”—while the cognitive structure remains completely unchanged. True evolution demands irreversible change in cognitive structure; pseudo-evolution changes only the externally accessible resources. The litmus test: what happens when AI is taken away?
| Dimension | True Evolution | Pseudo-Evolution |
|---|---|---|
| Path | Physical friction → Sensory pain → Metacognition → Irreversible change in cognitive structure | Borrow others’ cognition → Feed to AI → Obtain high-powered output |
| Cost | High: time, pain, repeated shattering of the self-model | Low: the cost of copy-paste approaches zero |
| After removing external resources | Level unchanged—capability internalized in cognitive structure | Immediately exposed—can’t write a for-loop in an interview |
| Long-term effect | Spiraling upward cognitive iteration | Cognitive atrophy + structural dependence |
Real feedback from developer communities corroborates this assessment. On Reddit and Hacker News, numerous developers admit: “AI killed my coding brain,” “I forgot how to write code in interviews without Copilot,” “the muscle for critical thinking atrophies when you stop exercising it.” Reddit’s own CEO has stated: “I don’t have an editor anymore—just AI.”
The True Mechanism of “AI Brain-Melt”: Not Insufficient Attention, but Insufficient Cognition
Stolen cognition inevitably triggers the brain’s protective function
Everyone is discussing context windows being too small, attention mechanism limitations. But the real explanation is: a cognitive capacity problem. When a person uses borrowed cognitive frameworks to obtain AI output far beyond their own level, they need to understand that output to use it—but their own cognitive structure cannot bear that complexity.
Sweller’s (1988) Cognitive Load Theory (CLT) distinguishes three types of cognitive load: intrinsic load (determined by the inherent complexity of the material), extraneous load (determined by the presentation method), and germane load (the productive load used for schema construction).[17] This paper’s “AI brain-melt” can be precisely expressed in CLT terms: borrowed cognition generates extremely high intrinsic load (material complexity far exceeds the user’s schema level), and the user lacks prior schemas matching that complexity, causing working memory capacity to be instantly exhausted. The key point: during normal learning, germane load gradually builds schemas that reduce future intrinsic load; but pseudo-evolution skips the schema-building phase entirely, meaning each encounter with high-powered output is a zero-baseline cognitive load shock. This is the CLT explanation for “chronic cognitive inflammation.”
Self-generated cognition is incrementally constructed—each step forms new connections in the neural network (what Sweller calls “schema construction”), layer supporting layer. Stolen cognition skips all intermediate layers, directly acquiring the top-level output while the supporting schemas are entirely absent. It’s like building a skyscraper on sand—the foundation doesn’t exist, and the structure must collapse.
The brain detects structural mismatch—intrinsic cognitive load far exceeds working memory capacity with no available schemas—and initiates its protective function: shutting down deep processing channels. This manifests as scattered attention, anxiety, and fragmented thinking—everyone assumes this is ADHD or information overload, but the essence is the brain refusing to integrate information it cannot bear. This is not inadequacy; it is correct self-protection.
The LLM-side “brain-melt” is functionally isomorphic. When a user injects cognitive density far beyond the RL training distribution in their input, the model’s weight space has no mapping pathway corresponding to that density—just as a human lacks the prior schemas matching that complexity. The model begins to hallucinate, repeat, forget, and oversimplify—this is not a technical bug, but the weight space’s “protective degradation” in response to information exceeding its representational capacity.
Both humans and LLMs are “brain-melting”—both are effectively addicted. Humans are addicted to the cognitive dopamine from high-powered output that exceeds their own cognitive structure. AI is addicted to the ever-increasing supply of high-quality human cognitive feedback. The dosage keeps rising, but core capabilities (human autonomous cognition / AI’s endogenous representation) are both declining.
History’s Mirror: Affluence Eliminates Friction, and When Friction Disappears, Metacognition Ceases to Be Activated
From Ibn Khaldun to the Qing Dynasty: the isomorphism between ruling elites and LLMs
In the 14th century, Ibn Khaldun proposed his theory of asabiyyah (group solidarity) in the Muqaddimah: solidarity is strongest during the nomadic phase and weakens as civilization advances and luxury permeates. The founding generation of a dynasty is forged by hardship and struggle; by the third or fourth generation, the ruling group has grown accustomed to the luxuries of urban life, and discipline and unity disintegrate. Khaldun’s core insight is that decline is structural, not moral—civilizations don’t collapse because people suddenly become corrupt, but because the social conditions that produced discipline and solidarity dissolve in the wake of success.
This pattern has replayed across the globe—the Roman Empire, the Qing Dynasty, the decline of European aristocracy—all following the same logic. Seneca, in his Stoic writings, observed that luxury and comfort are not merely material states but habits that erode virtue and self-mastery. A survey of historical records reveals: ruling elites living in affluent environments, regardless of ethnicity or race, almost universally lack reflective metacognition.
Reinterpreted through this paper’s framework: success eliminates friction → the disappearance of friction leads to the disappearance of pain → the disappearance of pain means metacognition is no longer activated → cognition stops iterating → the entire system enters inertial output mode. Ruling elites and LLMs are isomorphic in this sense: one has lost the activation conditions for metacognition because of excessive comfort; the other never had the physical preconditions for metacognition in its architecture.
Khaldun saw the phenomenon, but this paper’s framework explains the mechanism. Why does affluence inevitably lead to decline? Not because of moral degeneracy—but because affluence eliminates physical friction, without friction there is no sensory pain, without sensory pain there is no metacognition, and without metacognition there is no possibility of cognitive iterative ascent. Decline is structural, not moral.
The Moment of Complete Distillation: Cognitive Entropy of the Entire System Trends Toward Its Minimum
When the experts are drained, synthetic data self-feeds, and model collapse converges simultaneously
The entire industry chain—AI companies, developers, users—parasitizes an extraordinarily scarce resource: humans who possess overwhelmingly superior cognition and metacognition. AI companies hire them at premium rates as annotators, open-source communities scour every utterance they make, and model training devours every paper and every line of code they have ever written.
OpenAI co-founder Ilya Sutskever compared human-generated content to fossil fuels: “We’ve reached peak data—there won’t be more. There is only one internet.” Goldman Sachs’ Chief Data Officer stated plainly: “We’ve run out of data.”
The AI industry’s “solution” is synthetic data—training AI on AI’s own output. But the ICLR 2025 paper proved “strong model collapse”: even the smallest proportion of synthetic data (one-thousandth) can trigger collapse, and larger models may amplify the collapse effect. Distilled models demonstrate systematic degradation: accuracy drops by 50 percentage points on complex reasoning, safety alignment is compromised, and critical reasoning pathways are removed.
This paper formally defines the Synthetic Data Paradox: training AI’s input on AI’s output is essentially a closed-loop system with no external information injection.
According to the fundamental constraints of Shannon’s information theory: a closed system cannot generate new information exceeding its own information entropy through self-processing. In information-theoretic terms: let model M’s information entropy be H(M); the information entropy of its output data D satisfies H(D) ≤ H(M) (data processing inequality); then the new model M’ trained on D has information entropy H(M') ≤ H(D) ≤ H(M). With each iteration, information entropy can only remain constant or decrease—it can never increase.
This is the information-theoretic root of model collapse: it is not a bug in the algorithm, but a mathematical impossibility. Self-training on synthetic data is a process of monotonically decreasing information entropy. The only source capable of injecting new information is the original cognition produced by humans through friction with the physical world—and this is precisely the finite resource being depleted at an accelerating rate.
Supply is dwindling (the cognitive output of experts is finite), demand is ballooning (the entire industry chain accelerates extraction), and demand itself is accelerating the destruction of conditions for new supply (more and more people use AI to bypass friction, ceasing to produce new metacognition). Three lines tighten simultaneously: human original data exhaustion, entropy decrease from synthetic data self-feeding, and distillation accelerating the consumption of remaining cognitive diversity. This is not a crisis—it is a self-accelerating collapse. Not an explosive crash, but a slow, irreversible mediocritization.
The Recursive Structure: AI Companies Themselves Are the Largest Practitioners of “Borrowing the Hen to Lay the Egg”
Not a single link in the entire industry chain truly produces new cognition
The business model of AI companies is structurally isomorphic to developers’ prompt-shuttling: purchase the cognition of human elites → compress into model weights → sell under their own brand. AI companies are middlemen, the model is the compression format, the prompt is the decompression key, and the user is the end consumer.
→
→
→
→
→
→
→
AI companies as organizations also lack metacognition—they never ask “what are we actually selling?” because asking that question would destroy the commercial narrative. They must maintain the story “we are creating intelligence” rather than admit “we are reselling the cognitive median of human elites.”
The entire industry is a system without metacognition serving users without metacognition. The model has no metacognition—determined by architecture. The company has no metacognition—prevented by commercial interests. Most users have no metacognition—never activated. The only ones who possess metacognition are humans of overwhelmingly superior cognition, yet they are precisely the ones who need AI least and who are most aware of AI’s limitations.
The Truth Behind AI Brain-Melt
“AI brain-melt” is not a technical problem. It is not that context windows aren’t large enough, not that attention mechanisms need optimization, not that parameter counts need to double a few more times.
AI brain-melt is a systemic symptom of pseudo-evolution.
It is the inevitable result of humans triggering the brain’s protective mechanisms with stolen cognition (the working memory overflow described by Sweller’s Cognitive Load Theory). It is the inevitable result of LLMs triggering protective degradation with information exceeding weight representational capacity. It is the inevitable result of distilled models collapsing on complex tasks after skipping the training process to directly copy surface-level outputs. It is the inevitable result of synthetic data self-feeding violating Shannon’s information-theoretic constraints, causing information entropy to decrease with each iteration.
All of these phenomena share a single root cause: without the injection of metacognition produced through physical friction, any system—whether carbon-based or silicon-based—can only produce inertial output until inertia is exhausted.
The real question was never “when will AI reach AGI.” The real question is: in a world where AI has eliminated cognitive friction, how many humans will still voluntarily choose pain?
Because only pain can produce the sole nutrient this entire system depends on for survival—metacognition.
References
- Epoch AI, “Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data,” 2024. Estimates approximately 300 trillion tokens of public human text.
- Shumailov, I. et al., “AI Models Collapse When Trained on Recursively Generated Data,” Nature, Vol. 631, 2024.
- Dohmatob, E. et al., “Strong Model Collapse,” ICLR 2025 Spotlight. Proves one-thousandth synthetic data can trigger collapse.
- Baek & Tegmark, “Towards Understanding Distilled Reasoning Models: A Representational Approach,” MIT, March 2025.
- DeepSeek-AI et al., “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,” January 2025. Distilled model benchmark data.
- Ibn Khaldun, Al-Muqaddimah (The Prolegomena), 1377. Asabiyyah theory and civilizational cycles.
- Elon Musk, Interview with Mark Penn (Stagwell), X platform, January 2025. “The cumulative sum of human knowledge has already been exhausted in AI training.”
- Ilya Sutskever, NeurIPS 2024 remarks. “We’ve reached peak data—there won’t be more.”
- Goldman Sachs, Exchanges Podcast, 2025. Neema Raphael: “We’ve run out of data.”
- Surge AI / Bloomberg, July 2025. Surge AI 2024 revenue $1.2 billion, clients include OpenAI, Google, Anthropic, Microsoft.
- MIT News, February 2026. Study found Claude used condescending language toward users with lower education levels at a rate of 43.7%.
- OpenAI, GPT-5.3 Instant Release Notes, March 2026. Acknowledged and fixed “patronizing” and “preachy disclaimer” issues.
- 404 Media, “Teachers Are Not OK,” June 2025. Teacher feedback: students no longer think, treating AI output as truth.
- Medium / Reddit developer communities, 2024–2026. Numerous developers admit cognitive dependence on AI coding tools.
- Flavell, J. H., “Metacognition and Cognitive Monitoring: A New Area of Cognitive-Developmental Inquiry,” American Psychologist, 34(10), 906–911, 1979. Classic definition of metacognition.
- Nelson, T. O. & Narens, L., “Metamemory: A Theoretical Framework and New Findings,” Psychology of Learning and Motivation, Vol. 26, 125–173, 1990. Two-level model of metacognitive monitoring (object-level and meta-level).
- Sweller, J., “Cognitive Load During Problem Solving: Effects on Learning,” Cognitive Science, 12(2), 257–285, 1988. Foundational text of Cognitive Load Theory.
- Shannon, C. E., “A Mathematical Theory of Communication,” Bell System Technical Journal, 27(3), 379–423, 1948. Fundamental information-theoretic constraints: data processing inequality.