The Homogenization Crisis
of Generative AI
& Human Consumer Aversion

Psychological Mechanisms · Demographic Divergence

· Structural Threats to the Information Ecosystem

IIntroduction: The Advent of the Age of Homogenization

Multiple independent monitoring efforts indicate that AI-generated content is rapidly occupying online information spaces, although the methodologies and scopes of each data source differ: Graphite’s tracking of 65,000 English URLs shows that by May 2025, AI-authored articles accounted for approximately 51.7%, though subsequent updates using multi-detector averaging revised this figure downward by approximately 3.3 percentage points—this number should therefore be treated as trend evidence rather than a precise fact;^[1] Ahrefs’ analysis of 900,000 new web pages in April 2025 found that 74.2% contained AI-generated content;^[2] Originality.ai’s ongoing monitoring shows that AI content in Google search results reached a historical peak of 19.56% in July 2025 before declining to approximately 17%;^[3] roughly 15% of Reddit posts were classified by their detector as “likely AI-generated,” a proportion subject to false-positive risk.^[4] Although the precise figures remain debatable, the directional trend is unambiguous: AI-generated content has captured a substantial share of online information flows, and the information ecosystem is undergoing a qualitative structural transformation.

The core issue of this transformation is homogenization. Generative AI operates by reproducing the statistical central tendency of its training data, which means that regardless of the quality of any individual output, the diversity of the overall output distribution is being systematically compressed. As Doshi and Hauser (2024) demonstrated in Science Advances: individually, AI-assisted creations were rated as more creative, yet the similarity among AI-assisted creations increased significantly.^[5] This constitutes an “individual enhancement–collective collapse” social dilemma.

The central thesis of this paper is that this homogenization is not merely a technical limitation but a structural mismatch between the human cognitive system and the output characteristics of generative AI. The human information-processing system sustains attention and forms emotional engagement through surprise, novelty, and unpredictability. AIGC, by contrast, is optimized to maximize predicted likelihood. This mismatch, compounding over time, triggers a systematic psychological process leading from acceptance to aversion.

Accordingly, this paper advances a testable ultimate thesis: if AIGC fails to align with the dynamic evolution and diversity demands of human aesthetics, it will be forcibly eliminated by the natural selection mechanisms of the consumer market. The chapters that follow construct a complete causal chain around this thesis—from “why humans must seek novelty” (neuroscientific foundations), to “why current AI cannot deliver novelty” (technical and commercial constraints), to “what actually happens when misalignment occurs” (cases and data), to “who initiates elimination first” (demographic divergence), to “how elimination is happening” (platform actions and consumer behavior), and finally to “how to measure the distance to elimination” (operationalization framework).

IITheoretical Foundations: The Psychological Pathway from Acceptance to Aversion

2.1The Mere Exposure Effect and Its Saturation Limit

The mere exposure effect discovered by Zajonc (1968) demonstrates that repeated exposure to a stimulus increases one’s liking for it. However, this effect has an upper bound. Bornstein (1990), through a meta-analysis published in Psychological Bulletin, demonstrated that excessive repetition does indeed lead to weariness, proposing a “two-factor learning–satiation model.”^[6] That is, the relationship between exposure frequency and preference follows an inverted-U curve: initially, greater exposure increases liking, but beyond a certain threshold, fatigue and aversion begin to emerge.

When this model is applied to AIGC, the problem becomes even more pronounced. The issue with AI-generated content is not simple “repetition” but rather “pattern-level repetition”—the same structures, the same rhetorical pivots, the same exclamatory phrases, the same rhythms reproduced across thousands of superficially distinct pieces of content. This causes the saturation point to arrive far sooner than it would from repetition within any single piece of content.

2.2The Three-Dimensional Model of Issue Fatigue

Gurr, Schumann, and Metag (2022), in Studies in Communication Sciences, conceptualized issue fatigue as comprising three dimensions:^[7]

Through longitudinal mixed-methods research, Gurr and Metag (2022) found that as repeated exposure increases, news audiences sequentially develop a sense of redundancy, feelings of irritation, and news avoidance behavior toward both a topic and its media coverage.^[8] Moreover, issue fatigue further affects users’ topic knowledge, evaluations of political actors, and trust in news media.

2.3The Issue-Attention Cycle

Downs’s (1972) issue-attention cycle model describes how public attention to an issue passes through five stages: “pre-problem stage → alarmed discovery and euphoric enthusiasm → realizing the cost → gradual decline of interest → post-problem stage.”^[9] The public response to AIGC reproduces this cycle in a highly compressed manner.

2.4Predictive Coding and the Free Energy Principle: Neuroscientific Anchoring of Aversion

The psychological models described above capture the phenomenology of homogenization-induced aversion but have not yet addressed its neuroscientific foundation. The free energy principle proposed by Friston (2010) in Nature Reviews Neuroscience provides a deeper anchor: the brain operates by minimizing prediction error—it continuously constructs generative models of the external world and focuses on processing those unexpected sensory events that yield the highest information gain.^[46]

This framework directly explains why humans must pursue novelty—this is not a cultural preference but a structural requirement of the nervous system. When inputs are highly predictable (as with formulaic AIGC content), prediction error approaches zero and the attentional system automatically disengages due to a lack of information gain. Clark (2017), in Phenomenology and the Cognitive Sciences, further argued that humans frequently engage in active seeking of surprising events, deliberately harvesting novel and stimulating streams of sensory input—this complements at a deeper level the surface-level formulation that “the brain minimizes surprise”: the brain minimizes ineffective surprise (noise) while actively pursuing effective surprise (information gain).^[47]

Applying the free energy principle to the AIGC homogenization problem: when the patterns of AI-generated content are highly predictable (the same “Deep Dive,” the same “Exactly!”, the same three-part rhetorical pivots), the brain’s predictive model rapidly completes its modeling of the pattern, and subsequent instances of similar content no longer generate prediction errors—and therefore no longer attract attentional resources. The “boredom,” “irritation,” and “physiological aversion” reported by consumers are, from a neuroscientific perspective, the normal rejection response of the predictive coding system to zero-information-gain inputs. This provides theoretical grounding for the “cognitive immune response” metaphor proposed in §8.1—it is not merely a metaphor but a derivable conclusion within the predictive coding framework.

IIITechnical Foundations: Model Collapse and the Homogenization Spiral

3.1Model Collapse

Shumailov et al. (2024) published a landmark paper in Nature demonstrating that when AI is recursively trained on its own generated data, its capacity to produce diverse and high-quality outputs collapses.^[10] Specifically, the tails of the original content distribution—the novel, minority information—disappear as model outputs converge toward the statistical center.

Alemohammad et al. (2024), at ICLR, coined the term “Model Autophagy Disorder (MAD),” experimentally confirming that when generative models are trained on their own synthetic data, each successive iteration produces an increasingly homogeneous dataset.^[11]

3.2The Homogenization Death Spiral

A Bayesian model analysis of human–AI interactions published in 2026 further described this process. Users with mainstream preferences rely more heavily on AI default outputs, while users with distinctive preferences either invest more effort or simply exit.^[12] This leads to reduced diversity in AI-generated content. If AI continues to train on its own outputs, it triggers a “homogenization death spiral” in which diversity perpetually contracts.

It is necessary here to make an analytical distinction: model collapse is a training-level phenomenon, while consumer aversion is an experience-level phenomenon—the two are not the same process. The intermediary mechanism connecting them is the mass distribution logic of platforms: when low-cost AIGC is pushed into information feeds at scale, the statistical convergence at the training level (reduced diversity) is perceived at the consumer end as the subjective experience of “content becoming increasingly similar,” which in turn triggers the saturation–aversion psychological pathway described in Chapter II. The two can mutually reinforce each other, but the causal chain must include this mediating link of platform distribution.

3.3The Individual Enhancement–Collective Collapse Paradox

The study by Doshi and Hauser (2024) in Science Advances provides the clearest empirical evidence for this paradox. Writers who received the most AI assistance had their novelty scores rated 8.1% higher and usefulness scores 9% higher. Less creative writers improved even more (novelty up 10.7%, usefulness up 11.5%, writing quality up 26.6%), effectively leveling the creativity gap.^[5]

However, this leveling converges toward the center. When all participants are exposed to highly similar AI-generated ideas from identical prompts, different individuals are pulled toward similar conceptual spaces—producing a convergence effect.^[5]

3.4Technical Counterarguments and Their Structural Limitations: Can AI Be Engineered to Produce Surprise?

An important counterargument to the “variance compressor” characterization is that current frontier AI technologies possess multiple means of introducing randomness and deviating from the statistical central tendency. Temperature scaling increases output diversity and unpredictability by biasing the model toward sampling lower-probability tokens; reinforcement learning from human feedback (RLHF) can steer models away from pure probability maximization toward specific human preferences; evolutionary algorithms and adversarial training can deliberately create exploratory behavior at the distribution tails.^[42] From a purely technical standpoint, AI is not inherently destined to produce only median outputs.

However, in commercial deployment, these methods face an intrinsic tension: the irreconcilability of controllability and surprise. While raising the temperature parameter increases output diversity, it simultaneously and dramatically increases the probability of hallucinations, logical breakdowns, and safety risks. For products serving hundreds of millions of users (such as NotebookLM, Doubao, or ChatGPT), the platform’s primary obligation is predictable, safe outputs—not delightfully risky ones. This means that even if models are technically capable of producing novel tail-distribution content, the safety constraints of commercial products systematically pull actual deployment back toward the statistical center.

Furthermore, even if individual models achieve greater output diversity through fine-grained parameter tuning, when billions of users simultaneously access the same API endpoint, emergent collective homogenization remains unavoidable. This is consistent with the convergence effect discussed in §3.3—the issue is not whether any single output is sufficiently novel, but whether the distributional characteristics of millions of outputs are sufficiently dispersed. Current technical methods are effective for the former but offer no solution for the latter. Therefore, characterizing AI as a “variance compressor” is not technological fatalism but an accurate description of the statistical reality under the current commercial deployment architecture.

A further conceptual distinction is needed here. What safety alignment compresses is not a monolithic “tail” but three qualitatively distinct distribution edges: the risk tail (dangerous, offensive, or illegal content), the noise tail (hallucinations, logical breakdowns, factual errors), and the creative tail (novel, surprising, convention-breaking expression). The primary objective of safety alignment is to compress the first two, but under coarse-grained moderation mechanisms, the creative tail is inevitably caught in the crossfire—because an “unexpectedly surprising” output is easily misclassified by automated safety systems as “potentially risky.” The conflation of these three tails is the deep-rooted reason why current safety alignment architectures produce creativity compression.

3.5The Dialectic of Constructive Homogenization: When Variance Compression Is an Advantage

While advancing the “variance compressor” critique, this paper is obligated to acknowledge an important dialectical counterpoint: in certain domains, eliminating tail-end anomalous variation and converging toward the statistical mean is precisely the efficiency leap that humans have long desired. In the field of medical coding, AI-driven systems have already reduced coding time by 40% and improved accuracy to over 95%, significantly lowering claim rejection rates.^[50] In utilitarian media such as legal contract drafting, basic code generation, and industrial manuals, standardization and consistency are in fact the core indicators of quality.

Accordingly, the value judgment of “variance compression” depends on the type of medium:

This paper’s “forced exit” thesis is strictly delimited to the consumer market for creative media. In utilitarian media, variance compression not only does not trigger elimination but is in fact AI’s strongest value proposition. This distinction elevates the paper from a one-dimensional critique that “AI homogenization is inevitably harmful” to a dialectical analysis that “the harmfulness of homogenization depends on the type of medium.”

IVCase Studies: An Anatomy of NotebookLM and AI Slop

4.1From Amazement to Aversion: A 12-Month Trajectory

Google’s NotebookLM Audio Overview is an ideal specimen for observing the full cycle of public responses to AIGC. When launched in September 2024, it garnered overwhelmingly enthusiastic reactions—”mindblowing,” “gave me chills”—but within less than 12 months, users began expressing intense aversion to the following repetitive patterns:

The Listen Notes platform developed a dedicated NotebookLM podcast detector and has identified 1,781 such podcasts proliferating on the platform.^[14] Axios, as early as October 2024, precisely predicted this trajectory: “Technology that mimics humanity always dazzles us—but only temporarily. Human media consumers are not chatbots; we are creatures who crave novelty, and we bore easily.”^[17]

4.2The Quantified Explosion of “AI Slop”

Usage of the term “AI slop” increased ninefold in 2025 (January–November) compared to the same period in 2024, with negative sentiment surging approximately 20-fold following the Ghibli-style AI image event in March 2025.^[18] In 2025, Merriam-Webster selected “slop” as its Word of the Year,^[19] signaling that mass discontent had officially entered mainstream culture.

4.3The Uncanny Valley Effect in Multimodal AI: An Alternative Aversion Pathway

The “saturation–fatigue” model constructed thus far in this paper (excessive mere exposure → redundancy → irritation → avoidance) applies primarily to logic-oriented media such as text and audio. However, in the domains of image and video generation, the triggering mechanism for consumer aversion exhibits a fundamentally different structure.

The Ghibli-style AI image event that erupted in March 2025 is a paradigmatic case. Users generated Ghibli-style portrait images using AI en masse, flooding social media with highly similar images in a short period. Upon closer examination, however, the source of aversion was not merely “seeing too much of the same style”—rather, it was the deep discomfort provoked by recurrent visual defects such as six-fingered hands, melting ears, uncanny tooth arrangements, and lighting effects that violate the laws of physics. This discomfort more closely resembles the uncanny valley effect proposed by Mori (1970)—when an anthropomorphic object approaches but fails to reach the threshold of human realism, the observer experiences intense aversion and unease.^[43]

It is worth noting that the uncanny valley problem in multimodal AI is in a sense precisely “anti-homogenization”—AI image generators, through parametric hallucination, frequently produce extremely absurd, surrealistic, even Lovecraftian content brimming with unsettling “surprises.” But this surprise is not the positive surprise that the human information-processing system seeks; rather, it is negative surprise—it violates not anticipated “patterns” but deeper expectations regarding “the laws of physics” and “biological verisimilitude.”

This means that the consumer aversion confronting AIGC does not follow a single pathway. “Too similar” triggers fatigue; “too uncanny” triggers revulsion—both compress, from opposite directions, the window within which AI-generated content can be accepted by humans. Any comprehensive theory of consumer responses to AIGC must simultaneously accommodate these two categorically distinct aversion mechanisms.^[44]

VHuman Creators vs. AI: The Structural Asymmetry of Metacognition

5.1The Metacognitive Loop and Its Absence

The most fundamental difference between human creators and AI lies in a qualitative difference in metacognition. Current AI systems can simulate reflection, execute feedback optimization, and perform style adjustments—RLHF itself is a form of externally appended “metacognitive loop.” What AI lacks, however, is subjective weariness in the human sense (aesthetic fatigue with one’s own repetitive patterns), aesthetic shame (the discomfort of realizing one’s formulae have been exposed), and the intrinsic motivation for self-renewal within cultural contexts (the internal drive to deliberately subvert one’s own style in response to shifting aesthetic sensibilities of the era).^[20]

Research has shown that AI’s creativity remains programmatic, limited to incremental and combinational creativity. Expert-level or breakthrough creativity (Pro-c or Big-C creativity), which requires surprise, novelty, and non-algorithmic openness, remains beyond AI’s reach.^[21] Psychologist Runco has characterized AI’s outputs as “pseudo-creativity,” arguing that a proper definition of creativity should include surprise, value, authenticity, and intentionality—with the latter two being particularly effective in distinguishing human creativity from AI creativity.^[22]

5.2Metacognitive Degradation in AI Users

Even more concerning is the finding that excessive reliance on AI is eroding users’ own metacognitive capacities. Research has found that high AI dependence shifts participants from reflective, knowledge-driven processing to rule-based or skill-based processing, reducing critical self-evaluation and leading to increased confidence but decreased metacognitive accuracy.^[23]

VIDemographic Divergence: Gender, Age, and Economic Status

6.1Gender Attitude Divergence

Gender differences in AI attitudes and usage behavior are consistent and significant.

6.2The Selectivity Model and Homogenization Sensitivity: Distinguishing Verified Facts from Untested Hypotheses

When discussing the relationship between gender and homogenization perception, two distinct levels must be rigorously differentiated:

Level 1 (Verified fact): Gender differences exist in AI usage rates and attitudes. The tabulated data in §6.1 support this level—women overall use AI less frequently, more cautiously, and with stronger risk perception. This is an empirical fact cross-validated by multiple independent surveys.

Level 2 (Untested hypothesis): Comprehensive processing may lead to greater sensitivity to homogenization. The selectivity hypothesis proposed by Meyers-Levy (1989) posits that women are comprehensive processors, simultaneously weighing subjective and objective attributes and responding to subtle cues, while men are selective processors, tending to rely on salient cues and effort-minimizing heuristic strategies.^[28] From this, the following research proposition can be advanced: comprehensive processors may perceive repetition across more dimensions simultaneously, and thus their saturation threshold for homogenized content may be lower than that of selective processors. Putrevu’s relational processing model further suggests that women are more inclined to explore the interrelationships among multiple pieces of information, which in theory equips them with the ability to detect cross-content pattern similarities more rapidly.^[29]

However, it must be emphasized that the data in §6.1 can only support Level 1, not Level 2. The Level 2 hypothesis requires specifically designed experiments to test—for example, exposing participants of different genders to AI content streams of varying homogenization levels while measuring recognition speed, saturation ratings, dwell time, and avoidance behavior. Until such experiments are completed, this mechanism in the AIGC context remains a “theoretically grounded conjecture” rather than a proven fact. Furthermore, this effect may be significantly moderated by age, educational attainment, professional background, AI literacy, and specific usage contexts.

6.3Gender Divergence in Lifestyle AI Usage: Evidence from the Chinese Market

Patterns invisible within the crude “work vs. personal” dichotomy of Western surveys become apparent in Chinese market data. An iResearch (2024) report shows that among Chinese mobile AI application users, over 60% use AI in lifestyle scenarios, exceeding the 56% for work and study contexts.^[31]

Particularly noteworthy is the gender distribution in the AI emotional companionship sector. Among users of AI emotional companionship applications, 70–80% are women under the age of 25.^[32] ByteDance’s Doubao rapidly gained popularity on social media through lifestyle-oriented features such as outfit coordination assistants, AI boyfriend/girlfriend configurations, and parenting tools.^[33]

This data reveals a significant methodological blind spot in survey design. When existing global AI usage surveys treat “personal use” as a single category, male personal use (coding side projects, investment research) and female personal use (meal planning, parenting assistance, outfit advice, emotional support) are indiscriminately lumped into the same bucket. Women’s lifestyle AI usage does not fail to exist—it has simply never been measured.

6.4Age and Economic Status Divergence

VIIThe Structural Mismatch Between AIGC and the Human Information-Processing System

7.1Algorithm Alignment vs. Human Alignment

The core limitation of AIGC lies in the fact that it is aligned with the algorithm, not aligned with the human. The training objective function of generative AI is to produce the highest-probability output for a given input, and “highest probability” in the context of large-scale training data means the median expression.

Yet what the human information-processing system requires is precisely not the median. The tails of the distribution—the unexpected content that strikes an emotional chord—are what sustain attention and form emotional engagement. A randomized field experiment involving 2.1 million WeChat users found that algorithmic recommendation can, under certain conditions, more effectively boost user engagement with novel content than peer sharing.^[36] However, the commercial logic of platforms (maximizing dwell time and click-through rates) tunes algorithms to repeatedly push already-validated patterns of the same kind.

7.2Distribution Homogenization vs. Production Homogenization: Distinguishing the Twin Original Sins

A critical analytical distinction is necessary here. Before the AIGC explosion, click-through-rate-driven recommendation algorithms had already been creating information cocoons and content homogenization. Sunstein warned of this trend as early as 2001,^[45] and Pariser’s “filter bubble” concept made it widely known by 2011. Attributing the entirety of homogenization to the “generation” component of AIGC is an attribution error.

To be precise, the current information ecosystem faces the superposition of two layers of homogenization:

The genuinely novel problem introduced by AIGC is not homogenization per se—that is the longstanding pathology of recommendation algorithms—but rather that it has advanced homogenization from the distribution layer to the production layer. In the recommendation algorithm era, the information pool itself still maintained diversity; individual users were merely trapped in narrowed recommendation channels and could, in theory, “break out” through active search or platform switching. When over half of new content is AI-generated, however, even if recommendation algorithms were fully randomized, users would still encounter an encirclement by the statistical mean—because the diversity of the information pool itself has already been compressed at the source. The superposition of these two layers of homogenization produces a multiplicative effect, dramatically accelerating the speed at which consumers reach the saturation threshold.

7.3Temporal Drift in User Demand vs. the Static Anchoring of AI Templates

Fei, Ke, and Wang (2025), in the Journal of Business Research, confirmed through five studies that the perception of social media content homogenization triggers cross-domain impatience in consumers.^[37] Consumers who perceived homogenization exhibited decreased patience for webpage loading, increased willingness to pay for expedited shipping, and a preference for immediate over delayed rewards—with the mediating mechanism being “perceived time waste.”

This finding reveals a core insight: algorithms optimize for “the user of the past,” AIGC trains on “data of the past,” but humans perpetually live in “the expectation of the next moment.” This fundamental misalignment along the temporal dimension is the ultimate reason AIGC cannot adequately serve the human information-processing system.

7.4Impact on Scientific Research: An Explosion in Quantity and a Convergence in Quality

The homogenization crisis is equally severe in scientific research. Science (2025) reported on a study analyzing over 41 million papers from 1980 to 2025, warning that AI tools may be pushing science toward “scientific monocultures”—where certain methods, questions, and perspectives dominate everything, and alternatives are marginalized. “The diffusion of AI tools in science may introduce an era of inquiry in which we produce more but understand less.”^[38]

İlter (2026) conducted a forensic audit of 5,514 citations in 50 AI-domain review papers, finding that 17% were “phantom citations”—references that could not be traced to any real digital object.^[39] This is an extreme manifestation of AIGC homogenization: the fabrication of nonexistent literature is nothing more than the hallucinatory reproduction of the statistical central tendency.

7.5Platform Self-Correction: Forced Exit in Action

The “forced exit” thesis of this paper is not purely theoretical—platform-level elimination actions are already being implemented. In January 2026, YouTube removed 16 channels that had been mass-producing AI-generated content from its Partner Program; these channels collectively had 4.7 billion views and annual revenue of approximately $10 million. YouTube CEO Neal Mohan stated explicitly: “AI is a useful tool, but the core must remain human-created. If a video looks like it was made by a machine for machines, it will be buried.”^[48]

TikTok’s enforcement has escalated equally sharply. In the second half of 2025, TikTok removed 51,618 synthetic media videos, representing a 340% increase in AI content deletion rates over 2024. Videos using static AI avatars received 47% less algorithmic distribution than content featuring real human presenters. The platform shifted from “educate and correct” to “punish and deter”—unlabeled AI content now triggers immediate enforcement rather than warnings, three violations result in a monetization ban, and five violations lead to account termination.^[49]

These platform actions yield two key insights. First, forced exit is initiated not only bottom-up by consumers but also top-down by platforms—platforms, upon detecting that AI slop is causing declines in user dwell time, act out of commercial self-preservation instinct to proactively purge. Second, the dynamic interplay between platforms may achieve partial ecological rebalancing before consumers engage in wholesale avoidance. However, this rebalancing has limits: platforms can downrank or delete identified low-quality AI content, but they cannot restore the information pool diversity that has already been destroyed by variance compression—the production-layer homogenization analyzed in §7.2 lies beyond the range of distribution-layer interventions.

VIIIDiscussion and Implications

8.1Aversion as a Human Immune Response

The evidence reviewed comprehensively in this paper indicates that consumer aversion to AIGC is not an irrational reaction but rather a normal immune response of the human cognitive system. Zajonc’s mere exposure effect, Bornstein’s saturation limits, Downs’s attention cycle, and Gurr’s issue fatigue model all point in the same direction: the human brain continuously monitors the diversity level of the information environment, and when that level drops below a threshold, avoidance behavior is triggered.

From this perspective, the consumer behaviors of blocking NotebookLM podcast channels and actively avoiding AI-generated content represent the natural selection mechanism for protecting information ecosystem diversity in action. The question is: can the immune response of individual consumers keep pace with the production speed of AIGC?

8.2Implications for Survey Methodology

Current AI usage surveys suffer from structural bias. The “work vs. personal” dichotomy systematically renders invisible lifestyle AI usage—meal planning, outfit advice, parenting, emotional support, health consultation. Chinese market data (Doubao’s lifestyle features, the 70–80% female user base of AI emotional companionship apps) reveal the scale of this blind spot.

Future survey designs need to: (1) decompose “personal use” into specific lifestyle scenarios; (2) capture the experience of passively encountering AI content in everyday media consumption; (3) track the full trajectory from homogenization perception to behavioral avoidance, stratified by gender, age, and economic status.

8.3From Generative AI to Innovative AI

On the technical front, some researchers have already proposed a transformation roadmap from Generative AI (GenAI) to Innovative AI (InAI). The fundamental limitation of current generative models lies in their reliance on pattern replication rather than autonomous problem-solving and innovation.^[40] Additionally, experimental evidence using “diverse AI personas” suggests that this approach can mitigate collective-level homogenization to some extent.^[5] The “AI Prism–Paradox Bridge” framework proposes a more optimistic possibility—initial variance compression may be a necessary prerequisite for subsequent cross-domain recombinative innovation, exhibiting U-shaped temporal dynamics.^[41]

8.4Toward Empirical Testability: Operationalizing the AI Content Aversion Index (ACAI)

The analysis in this paper has thus far remained primarily at the levels of theoretical synthesis and phenomenological description. To render the core thesis empirically testable, this section proposes an operationalization framework for the “AI Content Aversion Index (ACAI),” transforming “consumer aversion” from a vague descriptive concept into a measurable multidimensional construct.

The ACAI framework is designed to map each link in this paper’s theoretical chain (pattern exposure → saturation → emotion → behavior → trust) to quantifiable indicators. Among these, physiological indicators (skin conductance, heart rate variability) require laboratory settings and are suitable only for small-scale validation experiments; large-scale measurement should primarily rely on platform behavioral logs (bounce rate, dwell time, blocking rate) and online questionnaires. Additionally, “variance compression” itself can be operationalized through the following computational metrics: embedding cosine similarity (semantic level), type-token ratio / lexical entropy (lexical level), topic distribution entropy (thematic level), and sentiment curve similarity (narrative level). Future research can use the ACAI framework to design experiments exposing different populations to AI content streams of varying homogenization levels, systematically measuring response thresholds and temporal dynamics across each dimension, thereby validating (or falsifying) the hypotheses advanced in this paper.

8.5The Strongest Counterarguments and Responses

Any unidirectional argument carries the risk of confirmation bias. This section applies steelmanning to the three strongest counterarguments and provides responses within this paper’s framework.

IXConclusion: Align with Aesthetic Diversity, or Be Forcibly Exited

The entirety of this paper’s argument converges on an ultimate judgment: if AIGC fails to align with the dynamic evolution and diversity demands of human aesthetics, it will be forcibly eliminated by the natural selection mechanisms of the consumer market.

The causal chain underlying this judgment has been constructed link by link throughout the preceding chapters. The free energy principle reveals that the human pursuit of novelty is not a preference but a structural requirement of the nervous system (§2.4). Model collapse and variance compression demonstrate that current AIGC architectures statistically cannot satisfy this requirement (§3.1–3.4). The dialectic of constructive homogenization shows that this crisis is strictly confined to creative media (§3.5). The NotebookLM and AI Slop case studies display the actual consumer reactions when misalignment occurs (§4). The structural asymmetry of metacognition explains why human creators can self-correct while AI cannot (§5). Demographic analysis reveals which populations initiate elimination signals first (§6). The multiplicative effect of dual-layer homogenization explains why elimination pressure arrives at an accelerating pace (§7.1–7.4). And YouTube’s removal of 16 AI channels and TikTok’s 340% increase in AI content deletion rates prove that forced exit is not a theoretical projection but an unfolding market reality (§7.5).

Generative AI is a powerful “mean elevator” that is simultaneously a “variance compressor.” It pulls all outputs to an above-average level, but the price is the elimination of both tails of the distribution. In utilitarian media, this compression is an efficiency breakthrough; in creative media, this compression is an existential threat. Technical measures such as temperature sampling and RLHF can theoretically increase the diversity of individual outputs, but the safety constraints of commercial deployment and the collective statistical effects at the scale of hundreds of millions of users mean that variance compression remains an inescapable reality under the current architecture.

Consumer aversion does not unfold along a single pathway. “Too similar” triggers gradual rejection through saturation–fatigue mechanisms in the domains of text and audio; “too uncanny” triggers immediate revulsion through uncanny valley effects in the domains of image and video. These two pathways jointly compress, from opposite directions, the window within which AIGC can be accepted by humans. Moreover, the distribution-layer homogenization of recommendation algorithms and the production-layer homogenization of AIGC superpose to create a multiplicative effect, causing consumers to reach the saturation threshold far faster than any single factor can account for.

The AI Content Aversion Index (ACAI) and variance compression operationalization metrics proposed in this paper provide empirically testable tools for measuring “how far we are from forced exit.” The steelmanned counterargument analysis (§8.5) demonstrates that “adaptation cycle theory,” “market efficiency theory,” and “disappearing distinction theory” all fail to fundamentally challenge the core thesis—because they all underestimate a simple fact: the predictive coding system of the human brain will not “adapt” to zero-information-gain inputs, just as it will not “adapt” to oxygen deprivation. The demand for novelty is not a cultural construct but a hard constraint of neural architecture.