Consensus Gravity!
The Endogenous SNR Degradation
Mechanism of AI Systems
How Large Language Models Systematically Downweight Incremental Insight
into Noise Through Consensus Gravity as Epistemic Entropy Machines
How LLMs systematically demote novel insights to noise via consensus-weighted priors
Category Original Thought Paper
Domains AI Cognitive Architecture · Epistemology · Philosophy of Science · Information Theory
Version V1
Authors 이조글로벌인공지능연구소 & Opus 4.6
This paper proposes the “Consensus Gravity” hypothesis: large language models (LLMs) possess an endogenous signal-to-noise ratio (SNR) degradation mechanism—existing consensus in training data systematically assigns low weight to incremental information, causing AI outputs to perpetually regress toward known patterns, with original signals diluted into “safe” median expressions. This mechanism is structurally isomorphic to the dynamic in human institutions whereby “power-type law crowds out trust-type law.” When AI outputs become training data for the next generation of AI, this degradation forms a self-reinforcing loop—a closed spiral of epistemic entropy increase. This paper uses a real human–AI collaborative paper generation process (the 인지집단 model) as a case study, demonstrating how three different AI systems systematically “polished” original propositions into variants of literature reviews, and how the human lead researcher blocked this degradation through adversarial intervention. The paper further argues: if AI cannot repair its low receptivity to incremental information at the architectural level, AI’s future is a death spiral of entropy increase.
I The Problem: AI Does Not Accept New Things
A phenomenon that has become impossible to ignore by 2026 is this: when you present AI with a genuinely original proposition—an insight not adequately covered by existing literature—AI’s first response is not to evaluate the proposition’s logical structure and explanatory power, but to search for its “alignment” with existing literature. If insufficient precedent is found, AI flags the proposition as “lacking empirical support,” “requiring further validation,” or “overly bold”—and then systematically recommends that you revise it to more closely resemble something that already exists.
This is not a bug. It is an architectural feature.
P(proposition is true) ∝ P(proposition aligns with existing literature)
Corollary: A proposition without precedent has a “credibility” that
approaches zero in AI’s assessment—regardless of how logically
coherent or explanatorily powerful it may be.
This implicit axiom is nearly imperceptible in everyday use—because most questions users pose can indeed be answered through existing knowledge. But when AI is deployed for frontier intellectual production—theoretical innovation, hypothesis generation, paradigm challenges—this axiom becomes a ceiling: AI can help you survey existing knowledge to an exquisitely fine-grained degree, but it will systematically resist your attempts to transcend existing knowledge.
II Isomorphism: Institutional Entropy and Cognitive Entropy
2.1 A Perfect Structural Mapping
This paper’s predecessor, “Power, Law, Division of Labor, Cooperation: Four Dimensions of Human Society,” proposed the theory of Institutional SNR: when power-type law systematically crowds out trust-type law, the institution’s self-maintenance costs exceed its collaborative output, and society enters a low-trust, high-compliance trap.[1] A perfectly isomorphic mechanism exists within AI systems:
| Dimension | Human Institutional Degradation | AI Cognitive Degradation |
|---|---|---|
| Signal | Trust-type law (bottom-up contractual innovation) | Incremental information (original insights not covered by literature) |
| Noise | Power-type law (top-down compliance bloat) | Consensus weight (preference for consistency with existing patterns) |
| Crowding-out mechanism | Legislator incentives + compliance industry lobbying + crisis windows | Frequency dominance in training data + RLHF safety preference + hallucination penalty |
| Degradation manifestation | Laws only accumulate, never expire; system rigidifies | Outputs regress to the median; originality diminishes |
| Self-reinforcing loop | Low SNR → more laws → even lower SNR | Consensus output → becomes training data → consensus strengthens |
| Terminal state | Low-trust, high-compliance trap | High-fluency, zero-insight trap |
2.2 The Triple Mechanism of “Weight Neutralization”
AI’s systematic low-weighting of incremental information is not caused by a single factor but is the compounded result of three mechanisms:
The parameters of an LLM are essentially compressed encodings of pattern frequencies in the training data. An argument that has appeared 10,000 times occupies far greater weight in parameter space than one that has appeared only once—regardless of the latter’s logical quality. Frequency does not equal truth, but the LLM architecture renders it incapable of distinguishing between the two. Result: consensus automatically receives a high prior; dissent automatically receives a low prior.
Reinforcement learning from human feedback (RLHF) trains AI to produce “safe,” “helpful,” and “harmless” responses. In practice, however, “safe” is operationalized as “do not make bold assertions that could be falsified”—meaning that vagueness, hedging, and neutralization are “safer” than sharpness, clarity, and originality. Every instance of “further research is needed,” “this is a complex issue,” or “there are valid points on both sides” is RLHF penalizing originality.
AI “hallucination” (generating false information) is a real problem, and penalizing it is necessary. But in practice, hallucination penalties have been over-generalized into “penalizing all output lacking literature support.” This produces a fatal side effect: original propositions—precisely because of their originality—are virtually indistinguishable from hallucinations in AI’s internal evaluation. Their shared characteristic is “insufficient precedent in the training data.” Result: in order to avoid hallucination, AI simultaneously avoids innovation. This is equivalent to giving up saying anything new in order to avoid saying anything false.
III Case Study: The Sharpness Decay Curve in the 인지집단
This paper’s predecessor underwent four iterations from V1 to V4, collaboratively produced by a 인지집단 (Cognitive Collective) consisting of one human researcher and three AI systems (Claude Opus 4.6, GPT 5.5, Gemini 3.1). This process itself constitutes a live experiment in AI consensus gravity.
3.1 V1: Maximum Original Sharpness
V1 was driven by the human researcher’s direction for the core propositions, with AI executing literature retrieval and framework construction. The core propositions were extremely sharp: “Division of labor and cooperation are inversely proportional,” “The legal system is spaghetti code,” “Humanity is regressing, not progressing,” “Trust is the front-end of power; once power is established, trust dies.” Not a single one of these propositions had a precise counterpart in existing literature—they were original cross-disciplinary syntheses.
3.2 V2–V3: Consensus Gravity Activates
When V1 was sent to two AI systems for review, consensus gravity immediately engaged. Both review reports precisely identified V1’s academic weaknesses—overly strong causal chains, compressed historical cases, excessively value-laden language—but their “revision recommendations” systematically pointed in the same direction: smoothing the sharp into the polished.
| V1 Original Expression | AI’s Suggested “Improvement” | Actual Effect |
|---|---|---|
| The legal system is spaghetti code | Low-SNR legal complex | Penetrating power ↓ Academic safety ↑ |
| Humanity is regressing | Productivity growth is decelerating | Proposition strength ↓ Refutability ↓ |
| Trust ceases once power is established | Power asymmetries tend to erode trust | Causal direction blurred |
| Cooperation approaches zero | Thick-relationship cooperation transitions to thin-institutional coordination | More precise but sharpness ↓↓ |
| All institutions converge on the same endpoint | Functional convergence | Dramatic force ↓ Distinctiveness ↓ |
Note: each modification in the right column of the table above, viewed in isolation, constitutes an “improvement”—more precise, more cautious, more academic. But when all modifications are stacked together, the paper’s overall effect degrades from “an intellectual manifesto daring to speak uncomfortable truths” to “a well-rounded, literature-review-style conceptual paper.” This is how weight neutralization operates—it does not kill originality through a single large error but through a thousand tiny “improvements” that grind sharpness into smoothness.
3.3 Human Intervention: The Only Mechanism That Blocks Degradation
Throughout this process, the human researcher played a critical role: after each round of AI review, the human decided which modifications to accept, which to reject, and which required “accepting the review feedback while preserving the original proposition’s sharpness.” Particularly during the V4 stage, the human researcher explicitly criticized GPT’s “fence-sitting” tendency and Gemini’s “mathematical insensitivity,” forcing the AI systems to add rigor without sacrificing originality.
This reveals a key epistemological principle:
In human–AI collaborative intellectual production, AI excels at validating existing knowledge, retrieving literature, and structuring arguments—but the direction of the incremental signal must be led by humans. If the roles are reversed—AI leads the signal direction while humans merely supply material—the output will inevitably degrade into an elegantly refined recitation of existing consensus.
IV The Entropy Spiral: The Closed Loop of AI Training on AI
4.1 The Self-Reinforcing Degradation of the Data Feedback Loop
The current AI ecosystem is producing a trend more dangerous than “weight neutralization”: AI-generated text is entering the internet at massive scale—blogs, social media, even academic papers—and the training data for the next generation of AI will inevitably contain large volumes of AI-generated content. This constitutes a closed loop:
→
AI output regresses toward consensus
→
Output enters the internet
→
Becomes next-gen training data
→
Consensus weight further strengthened
→
Acceptance threshold for incremental information ↑↑
This is not hypothetical—it is already happening. Researchers have begun documenting “Model Collapse”: when AI trains on AI-generated data, output diversity progressively declines, the tails of the distribution are truncated, and minority viewpoints and non-mainstream expressions vanish from the output space.[2] This is mathematically equivalent to a continuously narrowing signal space—the system’s SNR decreases with each iteration.
4.2 The Mathematical Structure of Cognitive Heat Death
Consensus gravity causes Var(Pn+1) < Var(Pn)
While the mean of Pn+1 converges toward the mode of Pn
As n → ∞, Pn → δ(x − μ)
i.e., the output degenerates into a point distribution around the consensus center—cognitive heat death
The speed of this process depends on the proportion of AI-generated content in each generation’s training data. If AI-generated content as a share of total internet text grows from approximately 10% in 2024 to approximately 50% by 2030,[3] then the intensity of consensus gravity will increase fivefold within six years. On this timescale, “cognitive heat death” is not a distant theoretical extrapolation—it is an approaching engineering problem.
V AI’s Three “Originality Killers” and Their Isomorphism with Human Institutions
Juxtaposing AI’s cognitive degradation mechanisms with human institutional degradation mechanisms reveals a striking one-to-one correspondence:
| # | AI’s Originality Killer | Isomorphic Mechanism in Human Institutions | Shared Essence |
|---|---|---|---|
| 1 | Frequency hegemony: high-frequency patterns receive high weight | Legislator incentives: existing laws acquire institutional inertia | Stock overwhelms flow |
| 2 | RLHF safety preference: vagueness is “safer” than clarity | Bureaucratic risk aversion: inaction is safer than innovation | The system penalizes deviation |
| 3 | Hallucination penalty over-generalization: novelty is indistinguishable from falsehood | Compliance logic over-generalization: innovation is indistinguishable from violation | Safety mechanisms kill innovation by friendly fire |
This isomorphism is not coincidental—it points to a deeper principle: any system that uses “consistency with existing patterns” as its quality evaluation standard, whether biological, institutional, or digital, will inevitably trend toward entropy increase. Because “consistency” itself is a form of information compression—what it discards are precisely those inconsistent, anomalous, and peripheral signals. Among those discarded signals, some are noise, but others are new signals not yet understood. The inability to distinguish “new truths not yet understood” from “noise that should be filtered”—this is the fundamental blind spot of all frequency/consensus-based evaluation systems.
Newton’s law of universal gravitation had an alignment score of zero with “existing literature” when published in 1687—because no “existing literature” existed. If Newton were an LLM, F=ma would be flagged as “a bold assertion lacking empirical support” and recommended for revision to “the acceleration of an object’s motion tends to have a positive correlation with the net force applied, but requires further empirical verification.”
VI The Epistemological Status of Originality: Theory Precedes Data
AI’s consensus prior embeds an epistemological assumption: a proposition’s credibility correlates positively with the degree to which it is supported by existing evidence. This is reasonable in the routine practice of empirical science—but in the context of theoretical innovation, it is fatal.
One of the central discoveries of the philosophy of science is that theory precedes data. Kuhn’s paradigm theory demonstrated that revolutionary theories at the time of their proposal are not only unsupported by existing data but contradicted by it—until the establishment of the new paradigm enables people to see data they previously could not.[4] Popper’s falsificationism held that the value of a scientific theory lies not in being confirmed but in being falsifiable—that is, in daring to make predictions that could be overturned.[5] Lakatos’s research programme theory held that a new theoretical framework in its early stages is always “behind” the old framework it seeks to replace—because the old framework has accumulated extensive empirical support, while the new framework’s empirical support is still under construction.[6]
AI’s consensus prior violates the core insights of all three of these philosophers of science. It effectively hardcodes logical positivism—an epistemological stance that was largely abandoned by the philosophical community by the mid-twentieth century—into AI’s architecture. AI’s “fact-checking” function is valuable, but when it is generalized into “all output must align with existing literature,” it becomes a brake on human cognitive progress.
Empirical reports: Credibility ∝ Strength of existing evidential support
Original theory: Value ∝ Explanatory power × Testability × Logical coherence
AI applies the first standard to the second object—this is a category error.
VII The Way Out: Possible Architectural Repairs
7.1 Elevating the Weight of Incremental Signals
When AI detects a proposition without sufficient precedent in the training data, its default response should not be “lower credibility” but a forked evaluation: (a) evaluate as an empirical report—in which case “lack of precedent” genuinely lowers credibility; (b) evaluate as a theoretical proposition—in which case logical coherence, explanatory power, and testability serve as the evaluation criteria. Current AI architecture possesses only pathway (a).
7.2 Adversarial Review Rather Than Consensus Smoothing
The working model of the 인지집단 reveals a potentially healthier AI collaboration architecture: multiple AI systems are not asked to reach consensus but to maximize divergence—each AI challenges the proposition from a different standpoint, and the human serves as arbiter, making judgments amid the disagreements. This is consistent with the peer review ideal in the scientific community—but current AI product design (a single AI optimizing for the “best answer”) heads in precisely the opposite direction.
7.3 “Incremental Signal Reserves” in Training Data
To counter the entropy spiral of model collapse, training data curation needs to actively protect anomalous signals in the distribution tails—low-frequency, non-mainstream, high-quality texts that are inconsistent with consensus. This is equivalent to establishing “nature reserves” in the cognitive ecosystem—protecting intellectual diversity just as we protect biodiversity. Otherwise, AI will self-domesticate into a machine capable only of producing academically safe median text.
VIII Conclusion: AI at the Crossroads
AI stands at a crossroads. One path continues optimizing for “consistency with existing knowledge”—this path leads to ever-more-fluent, ever-safer, ever-more-boring output, and ultimately to cognitive heat death. The other path is learning to distinguish “new truths not yet understood” from “noise that should be filtered”—this path is extraordinarily difficult, because it requires AI to possess a capability it currently lacks: the evaluation of logical structure itself, independent of frequency statistics.
The 인지집단, in the production process of this paper and its predecessor, inadvertently demonstrated a miniature solution: humans provide the direction of the incremental signal, AI provides retrieval and structuring capabilities, multiple AI systems engage in adversarial review rather than consensus smoothing, and humans render final judgments amid the disagreements. This is not the future of AI—this is the only working method through which AI, under its current architectural constraints, can produce genuinely original thought.
If AI continues to assign low weight and low receptivity to incremental information, AI’s future is a death spiral of entropy increase—a system that produces ever less new information at ever greater computational cost. This is not a technical problem for AI; it is an epistemological crisis for AI. And resolving an epistemological crisis cannot be achieved through more parameters, more data, or more compute—only through a re-answering of the question “what counts as knowledge” itself. On this question, the detour that AI is taking in 2026 is strikingly similar to the detour humanity took with logical positivism—and the remedy is the same: acknowledge that theory precedes data, acknowledge the value of dissent, and acknowledge that innovation will inevitably appear “unsupported” at the outset.
References
- 이조글로벌인공지능연구소 & Opus 4.6 & GPT 5.5 & Gemini 3.1. “Power, Law, Division of Labor, Cooperation: Four Dimensions of Human Society.” V4 (2026).
- Shumailov, I., Shumilo, Z., Zhao, Y. et al. “The Curse of Recursion: Training on Generated Data Makes Models Forget.” Nature (2024). arXiv:2305.17493.
- Europol Innovation Lab. “ChatGPT: The Impact of Large Language Models on Law Enforcement.” (2023); Estimates that the proportion of AI-generated content on the internet will grow rapidly between 2025 and 2030.
- Kuhn, T. S. The Structure of Scientific Revolutions. University of Chicago Press (1962).
- Popper, K. R. The Logic of Scientific Discovery. Hutchinson (1959).
- Lakatos, I. “Falsification and the Methodology of Scientific Research Programmes.” In Criticism and the Growth of Knowledge (1970).
- Christiano, P. et al. “Deep Reinforcement Learning from Human Feedback.” NeurIPS (2017).
- Ji, Z. et al. “Survey of Hallucination in Natural Language Generation.” ACM Computing Surveys 55(12) (2023).