Opinion Leaders and
Opinion Preferences

IIntroduction

1.1 Research Background and Problem Statement

In 1895, French social psychologist Gustave Le Bon, in The Crowd, first systematically described how individuals lose independent judgment within groups, inaugurating the academic exploration of opinion leaders and collective opinion formation. Over the ensuing 130 years—from Lazarsfeld’s two-step flow to Asch’s conformity experiments, from Kahneman’s cognitive biases to Sunstein’s echo chamber theory, from Muchnik’s social influence bias experiments to the AI persuasion field experiments published simultaneously in Nature and Science in 2025—researchers have repeatedly discovered the same core fact: humans overestimate the independence of their own opinions and underestimate the degree to which they are influenced.

However, this 130-year research tradition now faces an unprecedented point of rupture. Estimates from multiple industries and research institutions indicate that AI-generated or AI-assisted content already accounts for a significant proportion of newly published web pages. An Ahrefs analysis of 900,000 new web pages in 2025 found that 74.2% contained detectable AI-generated content; a 2026 academic study using Internet Archive samples estimated that approximately 35% of newly published websites were classified as AI-generated or AI-assisted; a related Europol report has been widely cited as predicting that by 2026, “up to 90% of online content may be synthetically generated.” When such dramatic changes occur in the fundamentals of the internet content ecosystem, the traditional opinion leader influence model built on personalized following faces a fundamental challenge.

The central questions of this paper are: How do the mechanisms through which opinion leaders influence opinion preferences change across different media technology eras? Has the reception side (human cognition) evolved in parallel? How does the scaled expansion of AI content reorganize the opinion leader ecosystem?

1.2 Research Significance

The theoretical significance of this paper lies in bridging the disciplinary barriers among evolutionary psychology (prestige bias), social neuroscience (mirror neurons), and communication studies (two-step flow / algorithmic amplification) to construct a unified analytical framework. Its practical significance lies in providing a scientific basis for public discourse governance, platform algorithm design, and media literacy education in the AI era. Three core innovations are advanced: first, the “dual-channel erosion” model (affective mirror alignment attenuation in the visual domain + signal-to-noise ratio collapse in the textual domain), explaining how AI content erodes mid-tier influencers through different mechanisms; second, a parasocial relationship typology demarcating the applicability boundaries of the hypotheses; third, a parameterizable hourglass model.

1.3 Research Methods and Paper Structure

This paper adopts an approach combining systematic literature review, interdisciplinary theoretical integration, and hypothesis construction. The literature covers 1895 to 2026, spanning five disciplines: social psychology, evolutionary psychology, communication studies, social neuroscience, and computational social science. The paper comprises nine chapters: Chapter II provides a literature review; Chapter III constructs a biological theoretical framework and parasocial relationship typology; Chapter IV analyzes individual differences; Chapter V compares temporal topology shifts on the transmission side; Chapter VI presents the dual-channel model and parameterized hourglass model; Chapter VII conducts a structural causal analysis; Chapter VIII provides discussion and verifiable pathway designs; Chapter IX presents the conclusion.

IILiterature Review: The Historical Trajectory of Research on Opinion Leaders and Opinion Preferences

2.1 The Discovery of the Group Mind (1895–1922)

Le Bon first proposed in 1895 that crowds are driven more by imagery, symbols, repetition, and shared emotion than by logical reasoning. He observed that individuals undergo profound psychological transformation upon merging into a crowd—personal inhibitions decline and impulsive behavior increases. Walter Lippmann’s 1922 Public Opinion constructed the social psychology of opinion formation under modern conditions, introducing the concepts of “pseudo-environment” and “stereotype”—the complexity of modern society has exceeded the boundaries of individual cognitive capacity, forcing people to rely on “pictures in their heads” rather than direct experience to understand the world. The book is regarded as a precursor to numerous subsequent theories (the Overton window, manufacturing consent, echo chambers, and social polarization).

2.2 Influence Mechanisms in the Laboratory (1936–1963)

Sherif’s 1936 autokinetic effect experiment demonstrated that when participants were placed in small groups, individual judgments rapidly converged toward group norms, and this influence persisted in subsequent individual retesting—social influence changes not only public expression but also what people genuinely believe. Asch’s conformity experiments between 1951 and 1956 pushed the question to more extreme conditions: when the correct answer was unambiguously clear, participants conformed on approximately 37% of critical trials, 75% conformed at least once, but 24% never conformed. Deutsch and Gerard distinguished in 1955 between normative and informational influence as two distinct mechanisms. Festinger’s 1957 cognitive dissonance theory, through the classic $1/$20 experiment, showed that small pressure is more effective than large pressure in changing internal attitudes—the most effective opinion manipulation involves getting people to “voluntarily” make small concessions, after which the cognitive dissonance mechanism automatically completes the persuasion. Milgram’s cross-cultural experiments from 1961 to 1963 demonstrated the universality of authority obedience effects, though 35% of participants successfully resisted authority pressure.

2.3 The Establishment of Two-Step Flow and Opinion Leader Theory (1944–1955)

Lazarsfeld et al., in the 1944 The People’s Choice, found that interpersonal contact had a greater influence on voting behavior than mass media. Katz and Lazarsfeld’s 1955 Personal Influence systematically articulated two-step flow theory: traditional opinion leaders are “information hubs” among ordinary people, embedded in four domains—marketing, fashion, public affairs, and film—where they exert domain-specific influence. Hovland’s communication research program at Yale was the first to systematically study source credibility, message structure, and attitude change in persuasion. Rogers’ 1962 Diffusion of Innovations proposed four methods for identifying opinion leaders and the S-curve of innovation diffusion.

2.4 Refinement of Influence Mechanisms (1969–1995)

Moscovici’s 1969 blue-green slides experiment inverted the Asch paradigm, demonstrating that a consistent minority can change the deep-seated beliefs of the majority—when written responses were allowed, conformity actually increased. Janis proposed groupthink theory in 1972, analyzing the psychological mechanisms behind catastrophic decisions made by high-intelligence groups in cases such as Pearl Harbor, the Vietnam War, and the Bay of Pigs invasion. In 1974, Noelle-Neumann proposed the spiral of silence, while Tversky and Kahneman published their research on heuristics and biases in judgment. Cialdini integrated six influence principles in 1984, and Petty and Cacioppo proposed the Elaboration Likelihood Model (ELM) in 1986, distinguishing between central and peripheral routes. Zaller’s 1992 RAS model proposed that most people’s opinions are essentially “constructed on the spot.” Kuran’s 1995 preference falsification theory showed that the distribution of public preferences may differ dramatically from that of private preferences.

2.5 Opinion Dynamics in Digital Environments (2001–2021)

From 2001 onward, Sunstein continuously tracked group polarization and echo chamber issues. In 2007, Watts and Dodds’ computer simulation challenged the “influentials hypothesis”—large-scale influence cascades are driven more by large numbers of easily influenced people. In 2013, Muchnik et al. published a large-scale randomized experiment in Science: positive social influence increased the probability of positive ratings by 32%, and a cumulative positive herding effect raised final ratings by an average of 25%. Bail’s experiments between 2018 and 2021 revealed a counterintuitive finding: exposure to opposing viewpoints actually made people more extreme—social media is not a mirror but a prism.

2.6 Persuasion and Opinion Formation in the AI Era (2024–2026)

In 2024, Matz et al. demonstrated in Scientific Reports across four studies (N=1,788) that AI-personalized persuasion messages are more influential than non-personalized messages. The 2025 papers published simultaneously in Nature and Science showed through controlled experiments that AI chatbots can shift some voters’ candidate attitudes and voting intentions—experiments were replicated across election contexts in the United States, Canada, and Poland, though whether these experimental results can be extrapolated to actual electoral behavior and large-scale social environments still requires cautious assessment. A PNAS experiment found that LLM persuasiveness gains follow diminishing returns. Chmel et al.’s pre-registered field experiment found that social media creators’ influence on policy positions exceeds that of traditional campaign outreach. In the domain of virtual entities, the virtual idol and VTuber ecosystem (e.g., Kizuna AI and the Hololive series since 2016) as well as user studies of AI companion platforms such as Replika demonstrate that even when users know the entity is not a real person, strong parasocial attachment can still form—this provides an important counter-reference for boundary demarcation of the hypotheses presented later.

IIITheoretical Framework: The Biological Basis of Opinion Leader Following

3.1 Prestige Bias: An Evolutionarily Shaped Social Learning Shortcut

The prestige theory proposed by Henrich and Gil-White in 2001 suggests that people use indirect success cues as adaptive shortcuts for selecting learning targets; this “prestige bias” may be a uniquely human social learning mechanism. Learners offer deference to high-status individuals in exchange for greater access, maximizing opportunities to acquire adaptive behaviors. However, Boyd and Richerson’s 1988 research showed that when high prestige is associated with non-adaptive cultural traits, prestige-seeking behavior can evolve at the cultural level even when it reduces reproductive success—this explains why people follow manifestly “useless” influencer content.

3.2 Social Grooming and the Evolutionary Function of Language

Dunbar’s 1996 research found that approximately two-thirds of human conversation time is devoted to social topics. He proposed that humans evolved language as a substitute for primate physical grooming, serving social bonding functions within groups of approximately 150 individuals. Blogs, vlogs, comment sections, and real-time chat overlays are functionally “digital grooming”—the neural circuits they activate are functionally homologous to those activated by ancestors grooming fellow tribe members.

3.3 The Mirror Neuron System: The Neural Basis of Affective Alignment

The role of mirror neurons in social cognition is important but complex. Observation of facial expressions triggers spontaneous facial mimicry, promoting emotional contagion and social alignment in affiliative social contexts. EEG studies have found that mirror neuron activity is stronger when participants process facial emotions than facial identities, pre-recorded dynamic facial expressions elicit stronger mimicry responses than static images, and cultural congruence modulates mirror response intensity. However, it must be emphasized that mirror neurons are one of multiple neural bases for social influence, not the sole mechanism—they operate in concert with the oxytocin system, reward circuits, and social cognition networks to collectively support human social learning capacity.

3.4 Parasocial Relationships and Typology

Horton and Wohl proposed parasocial interaction theory in 1956. Attachment theory indicates that humans naturally respond to any form of human-like communication. Oxytocin signaling participates in the formation and maintenance of social relationships, and dopamine reward systems are repeatedly activated during social media engagement.

However, not all parasocial relationships rely on the same set of mechanisms. Based on a literature synthesis and counter-evidence from virtual idol/AI companion research, this paper proposes a parasocial relationship typology distinguishing four different attachment mechanisms:

The significance of this typology is that the “affective mirror alignment attenuation” hypothesis proposed later in this paper applies only to the first type, “authenticity-based following,” under AI-simulated-human conditions; it does not cover fantasy parasocial relationships and functional dependence—the existence of the latter two demonstrates that non-human agents can also establish specific types of emotional bonds.

3.5 Integrated Model: A Multi-Mechanism Synergy Framework

Opinion leader following is not the product of a single mechanism but a multidimensional synergistic activation process. This paper formalizes it as:

Here, prestige bias provides target selection; social grooming provides bonding needs; the mirror neuron system (in the visual domain) provides affective alignment; and parasocial bonding provides emotional attachment. Why vlogs are more “addictive” than academic papers is not because the content is more truthful, but because they simultaneously activate more social neural circuits. From campfires to TikTok, the driving force is the same set of Pleistocene hardware expressing itself through different media.

IVIndividual Differences: The Spectral Structure of Agency

4.1 Six Core Dimensions of Influence Resistance

130 years of research have identified at least six dimensions that systematically modulate individual susceptibility to influence: Rotter’s 1966 locus of control—internals show significantly lower conformity rates; Cacioppo and Petty’s 1982 need for cognition—attitudes formed by high-NFC individuals are more resistant; openness from the Big Five personality traits—the 2023 University of Bern replication confirmed it as the only trait significantly associated with lower conformity; domain expertise provides a confidence buffer; cultural dimensions modulate conformity rates; and fundamental differences exist between refutation-type and avoidance-type persuasion resistance strategies.

4.2 The Experimental Portrait of “Independents”

“Independents” in classic experiments are not statistical noise. The 24% of full non-conformers in Asch’s experiments, the 35% who defied authority in Milgram’s experiments, and the approximately two-thirds of “guards” in Zimbardo’s Stanford Prison Experiment who resisted pressure to be abusive are cross-experimentally consistent. Crutchfield et al. found that resisters scored higher on measures of intellectual ability and ego strength. Yet even independents reported tension and impulses during the resistance process—agency requires continuous investment of cognitive resources to maintain.

4.3 New Vulnerabilities of Individual Differences in the AI Era

AI-personalized persuasion can precisely identify each person’s position on the six-dimensional spectrum and select the path most likely to circumvent their specific resistance mechanisms. Previously, opinion leader influence was “broadcast-style”—the same message addressed to everyone, with naturally immune individuals possessing strong resistance. AI-personalized persuasion is “precision-targeted.” However, this vulnerability may also catalyze new forms of resistance: AI-assisted fact-checking tools, C2PA content credential standards, and algorithmic transparency requirements can constitute “acquired immunity enhancement”—human agency can not only degrade but also be re-empowered through institutional and technological design.

VThe Temporal Topology Shift on the Transmission Side

5.1 The Traditional Assembly Model: Pulse-Decay Influence

Durkheim’s “collective effervescence” described the synchronization and emotional intensification process that occurs when individuals gather: co-presence produces cognitive changes, individual consciousness gives way to group consciousness, movements become homogenized, and emotions mutually amplify. Zimbardo proposed deindividuation theory in 1969, but Postmes and Spears’ 1998 meta-analysis revised this theory—the SIDE model argues that anonymity in groups switches the self from personal identity to social identity rather than dissolving it. The most commonly observed longitudinal pattern in persuasion research is the normal decay of attitudes—newly formed attitudes regress toward baseline over time.

5.2 The Internet Model: Low-Frequency–Cumulative–Weak-Decay Influence

Zajonc’s 1968 mere exposure effect revealed a different temporal dynamic: repeated exposure alone is sufficient to enhance attitudes, requires no reinforcement, and operates below conscious awareness. The asynchronous and multi-exposure nature of internet content creates conditions for “ideational rumination”—the same content is indexed by search engines, pushed by social media algorithms, and cited and reshared by others. Epstein et al.’s 2025 three randomized controlled experiments directly quantified: multiple exposures to directionally biased content produce greater attitude shifts than single exposure.

This paper formalizes three transmission models as temporal influence functions:

where P_i is the personalization match strength. The traditional model is exponential decay of a single pulse, the internet model is the superposed accumulation of multiple pulses, and the AI model adds an individual customization coefficient on top of the cumulative base.

5.3 Algorithmic Amplification: The Disproportionate Coverage of Tech Leaders’ Voices

The traditional transmission pathway (party → media → public opinion) has been restructured by tech leaders/platform owners into “opinion leader → platform → followers.” Huszár et al.’s 2022 PNAS study, a large-scale randomized controlled experiment based on nearly 2 million daily active Twitter/X accounts, found that in 6 of 7 countries, mainstream political right-wing voices enjoyed higher algorithmic amplification. A strategic analysis proposed a “30-3,000-300,000” cascade schematic model: 30 core nodes influence 300,000 members of the public through 3,000 mid-tier influencers—under algorithmic acceleration, this cascade can complete within milliseconds.

5.4 The Critical Asymmetry: No Matching Upgrade on the Reception Side

Evolutionary mismatch theory provides the core framework for understanding this asymmetry: there is a gap between the cognitive tools we inherited and the demands of the contemporary environment. Heuristics that were adaptive in small-scale social contexts become systematic vulnerabilities in algorithmic environments. Metacognitive myopia—the human information processor’s inability to evaluate the quality and history of information—was not a defect in the Pleistocene, but in the era of algorithmic push, it makes the mere exposure effect engineerable. The high consistency between the approximately 37% conformity rate in Asch’s 1951 experiment and the 33% conformity rate in Franzen and Mader’s 2023 replication suggests that the basic magnitude of human susceptibility to group pressure has remained significant over 72 years. It should be noted that Bond and Smith’s 1996 meta-analysis of 133 cross-cultural Asch replications found that American conformity rates had declined since the 1950s—possibly attributable to the strengthening of individualist cultural values. However, the magnitude of decline relative to baseline is limited (a third of participants still conform when the answer is unambiguously clear after 72 years), and Franzen and Mader found that conformity rates were actually higher on political opinion tasks (38%), suggesting the basic mechanism is intact and may even be stronger in new domains.

VIAI Content Expansion and the Reorganization of the Opinion Leader Ecosystem

6.1 The Scale of AI-Generated Content—Three-Tier Evidence

Estimates of the proportion of AI-generated content on the internet require distinguishing among three tiers. At the empirical measurement tier, Ahrefs’ 2025 analysis of 900,000 new web pages found that 74.2% contained detectable AI-generated content. At the academic estimation tier, a 2026 study published on arXiv estimated, using Internet Archive samples, that by mid-2025 approximately 35% of newly published websites were classified as AI-generated or AI-assisted. At the risk projection tier, a related Europol report has been widely cited as predicting that by 2026, “up to 90% of online content may be synthetically generated”—this is a predictive estimate, not a measured proportion, but it marks the high level of concern from law enforcement and security agencies regarding synthetic content scale. The three tiers converge on a directional judgment: AI-generated content already accounts for a significant, possibly dominant, proportion of newly published web materials, with no sign of decelerating growth. Since current AI content detectors have varying levels of false positives and false negatives, these figures should be treated as trend indicators rather than precise census data.

6.2 The Dual-Channel Erosion Model

6.2.1 Channel One: The Visual Domain—Affective Mirror Alignment Attenuation

Traditional influencer-type opinion leaders activate audiences’ mirror neuron systems through facial expressions and body movements in video, generating spontaneous facial mimicry and emotional contagion to produce “affective mirror alignment”—this is the neurobiological foundation of authenticity-based parasocial relationships. The core hypothesis of this paper is: when audiences identify or believe video content to be AI-generated, mirror alignment for positive emotions may be significantly attenuated.

Key evidence comes from an EEG study by Eiserbeck et al. (2023) at Humboldt-Universität zu Berlin: faces perceived as displaying genuine smiles elicited classic emotional effects on P1, N170, and EPN components, whereas faces perceived as deepfake smiles showed no emotional effects on these indices; only LPP activity was enhanced—indicating a switch from an affective resonance mode to a more effortful cognitive evaluation mode. Negative expressions elicited typical effects regardless of authenticity—positive affective signals are precisely the core tool through which influencers build personalized following, and they are also the most vulnerable channel when authenticity beliefs are compromised.

The boundary conditions of this hypothesis must be clearly delineated. First, it has received experimental support under conditions where “participants know or believe the content is AI-generated,” but effects under uninformed conditions remain to be verified. Second, existing evidence hints that the human brain may possess subthreshold detection capability for synthetic faces—machine learning classifiers can distinguish AI faces from real faces based on neural activity—but definitive evidence is insufficient, and this threshold may be breached as video generation technology advances. Third, this hypothesis applies only to the “authenticity-based following” type in the parasocial relationship typology, not to “fantasy parasocial relationships”—users of virtual idols and VTubers form strong attachments even when knowingly interacting with non-human entities, demonstrating that non-human agents can establish specific types of emotional bonds through characterization, continuous updating, and interactivity.

6.2.2 Channel Two: The Textual Domain—Signal-to-Noise Ratio Collapse

If influence erosion in the visual domain stems from mirror alignment attenuation at the neuroscientific level, influence erosion in the textual domain stems from signal-to-noise ratio collapse at the information economics level. Traditional text-based opinion leaders (bloggers, columnists, commentators) never relied on visual mirror neurons to establish parasocial relationships—their competitive advantage was “content that is more professional, deeper, and more insightful than what ordinary people produce.” When AI can generate unlimited quantities of comparable-quality content at zero marginal cost, this advantage is systematically compressed.

Major search platforms have already systematically reduced the visibility of low-quality synthetic content through multiple algorithm updates—this itself is an institutional response to signal-to-noise ratio collapse in the textual domain. An operator equipped with advanced prompting techniques can produce 500 blog posts or 5,000 e-commerce descriptions in a matter of days—this “synthetic scale” fundamentally upends the traditional assumption of content scarcity.

Channel Two in the textual domain and Channel One in the visual domain point toward the same outcome through different mechanisms: the erosion of the influence base of mid-tier opinion leaders.

6.2.3 Stress Test: Counter-Case Analysis of Virtual Idols and AI Companions

Virtual idols (such as Kizuna AI, the Hololive VTuber series) and AI companion platforms (such as Replika) constitute direct counter-examples to the overly strong claim that “AI cannot establish personalized following.” Large numbers of users form intense emotional attachments even when they know the entity is not a real person. These phenomena require precise boundary demarcation of the hypothesis: AI-generated agents can establish functional, fantasy, or character-based parasocial relationships, but on the authenticity dimension constituted by “real human experience—vulnerability—irreplaceable life history,” they remain weaker than verifiable human opinion leaders. The question is not “Can AI establish relationships?”—the answer is yes—but rather “In what dimensions do the types of relationships AI establishes differ structurally from the authenticity-based following of human opinion leaders?”

6.3 The Parameterized Hourglass Model

The hourglass model needs to be upgraded from a theoretical image to a parameterizable, testable structure:

Both indicators must be reported simultaneously: if absolute volume rises but relative share declines, it means the mid-tier has not truly contracted but rather the top tier has grown faster; if both absolute volume and relative share decline simultaneously, this supports the strong hypothesis of genuine mid-tier collapse. If Relative Middle Compression increases as AI content proportion rises, the hourglass hypothesis is supported. Measurable indicators include: top tier (traffic and revenue share of the top 1% of creators), mid-tier (survival rate, conversion rate, and median income of creators with 10,000 to 1,000,000 followers), bottom tier (local community recommendation rate and private community engagement), AI impact (AI content supply volume in the same domain and user trust index). The shape of the hourglass may differ significantly across cultures—the bottom layer of face-to-face trust may be more robust in collectivist societies, while top-tier concentration may be more extreme in individualist societies.

The “authenticity moat” of the top tier is not absolute immunity. In the algorithmic era, interactions between super top-tier figures and the public are still mediated through digital platforms. Deepfake technology can generate highly realistic false statements or scandal videos, and under “low-frequency–cumulative–weak-decay” algorithmic amplification, once such high-intensity fabricated information breaches a critical threshold, it could deconstruct the credibility foundation of super top-tier figures. Whether the authenticity premium remains valid in the face of extremely high concentrations of AI noise is itself an open question requiring ongoing tracking.

6.4 AI Trust Compensation: When Human Distrust Converts to AI Trust

The completeness of the hourglass model also requires consideration of a reverse pathway. A 2025 study on “deferred trust” proposed that low trust in humans may convert into a higher propensity to choose AI—some users, distrusting human political and media systems, actively turn to AI, perceiving it as more neutral and objective. This means the AI era is not simply one of “AI reducing trust,” but rather that a reverse channel exists: “human distrust → AI trust transfer.” This weakens the strong assertion that “AI necessarily lacks personalized influence” but strengthens the model’s real-world explanatory power—in societies with high levels of human distrust, AI may function not as a “weakener of influence” but as an “alternative trust bearer.” This paper treats this as a boundary condition of the hourglass model: when societal trust in traditional opinion leaders and media institutions falls below a certain threshold, AI may instead fill the trust vacuum.

VIIStructural Causal Analysis of Human Social Development and Agency Transformation

7.1 Five Structural Causal Chains

Integrating 130 years of research evidence, this paper identifies five structural causal chains. First, media technology evolution → thickening of the information intermediary layer → shrinkage of direct experience → collapse of the subjective judgment base. Second, expansion of social scale → increased anonymity → social comparison anxiety → strengthened conformity motivation. Third, deepening professional specialization → lengthening of cognitive delegation chains → enhanced authority dependence → concentration of opinion leader power. Fourth, information overload → cognitive resource scarcity → deepened heuristic dependence → expanded manipulation windows. Fifth, the rise of identity politics → binding of opinions to identity → attitude change becomes equivalent to identity threat → polarization lock-in.

7.2 From “Social Screening” to “Self-Coronation”

The authority of traditional opinion leaders is bottom-up emergent, domain-specific, socially verified, and bidirectionally revocable. The opinion authority of tech leaders/platform owners is cross-domain spillover, algorithmically amplified, unidirectionally irrevocable, and unlimited in scale. When the opinion leader is also the platform owner, error-correction mechanisms not only fail but are reverse-exploited—errors and controversies become fuel for algorithmic amplification.

7.3 The Externalization of Agency and the Enhancement of Manipulability

The transformation of agency has progressed through five stages: group dissolution (1895–1936), structural concession (1951–1963), systematic vulnerability (1974–1995), algorithmic substitution (2001–2021), and AI infiltration (2024–2026). But this trajectory should not be read simply as a unidirectional “degradation narrative.” Human agency has simultaneously experienced migration, reorganization, and outsourcing even as it has been eroded: mass media gave people informational capabilities beyond direct experience; social media empowered marginalized groups with unprecedented counter-authority mobilization capacity; AI tools are also enhancing individuals’ abilities to rebut, verify, write, and organize. Platform environments can both manipulate and catalyze new forms of collective intelligence. A more accurate description is not “degradation of agency,” but rather “the degree of externalization of agency is continuously deepening and its plasticity is continuously increasing—which simultaneously implies rising manipulability and the possibility of empowerment.”

VIIIDiscussion

8.1 Theoretical Implications of the “Unchanged Reception Side” Proposition

The implication of this paper’s “transmission-side topology shift vs. reception-side evolutionary mismatch” framework is that it redirects attention from “how technology has changed us” to “how technology exploits what has not changed about us.” The consistency of conformity rates between Asch’s 1951 experiment and Franzen and Mader’s 2023 replication over 72 years indicates that the basic magnitude of human susceptibility to group pressure remains significant—although Bond and Smith’s meta-analysis shows that American conformity rates have declined, the one-third baseline is sufficient as an analytical anchor. What has changed is not the person, but the temporal structure and power topology of the information environment in which the person is embedded.

8.2 Limitations of the Dual-Channel Model and Hypotheses

The hypotheses of this paper have five main limitations. First, AI video technology is advancing rapidly, and the gap in visual-level mirror responses may narrow. Second, Eiserbeck’s experiments are based on the condition that “participants know it is AI”; effects under uninformed conditions remain to be tested. Third, mirror neuron mechanisms are inapplicable in the purely textual domain; signal-to-noise ratio collapse is an independent information economics mechanism rather than a neuroscience mechanism—the dual-channel model acknowledges that the two domains follow different logics. Fourth, the subconscious resolution threshold may be breached as AI technology advances, at which point the mid-tier may not collapse but be replaced by “AI digital humans.” Fifth, cross-cultural verification is insufficient, as existing evidence is biased toward Western participants. Furthermore, one must candidly address the major controversy within the mirror neuron research field: scholars such as Hickok (2009/2014) and Heyes and Catmur (2022) have raised serious questions about the explanatory power of mirror neurons in social cognition, arguing that their role may have been over-extended and that mirror neurons may be a product of associative learning rather than evolutionary adaptation. Channel One of this paper’s hypothesis is built on the relatively well-established narrow-domain evidence that the mirror system participates in emotional facial processing, rather than on the strong version of mirror neuron theory, but the above controversies nonetheless constitute a theoretical risk requiring ongoing attention.

8.3 Verifiable Experimental Design

The critical experiment for verifying the “affective mirror alignment attenuation” hypothesis should employ a five-group design:

Measurement indices include: EEG N170, EPN, and LPP components and mu rhythm suppression; facial electromyography (zygomaticus/corrugator); subjective closeness and parasocial relationship scales; persuasion effectiveness and memory retention. Particular attention must be paid to the cross-contamination risk in Groups C and D data interpretation—participants’ subconscious may detect microsecond-level pixel artifacts in AI video before conscious reporting, so the experiment should incorporate high-frequency eye tracking and galvanic skin response (GSR) as baseline separation tools to distinguish between two competing explanations: “genuine visual defect perception” versus “subconscious belief change.” The critical determination is: if Group D (“human but told AI”) shows mirror/affective attenuation, then the problem lies in authenticity beliefs rather than visual quality—this would substantially support the core hypothesis of this paper.

Verifying the hourglass model should involve collecting creator platform data stratified by follower count (top 1%, mid-tier 10,000 to 1,000,000, micro-tier below 10,000, and local communities), observing changes in traffic share, engagement rates, sponsorship revenue, follower growth, and user trust before and after AI content proliferation. If mid-tier indicators decline significantly while top-tier and micro-tier indicators show relative strengthening, this supports the hourglass restructuring.

8.4 Sociopolitical Consequences of Opinion Leader Ecosystem Reorganization

The hourglass model implies a triple consequence. The democratic risk of top-tier concentration lies in a very small number of people gaining disproportionate voice power without effective error correction. The opportunity of micro-tier reversion lies in the possibility that Dunbar’s 150-person circles may be reactivated as a foundation for democratic dialogue. The knowledge-ecological consequence of mid-tier collapse is that professional content creators—bridges connecting academia and the public—face existential crisis. Simultaneously, the emergence of a “ghost layer”—composed of large numbers of synthetic accounts and AI persona-type opinion leaders—may make the information space even more opaque. The rise of human-AI hybrid creators is also blurring the boundaries between layers of the hourglass.

8.5 Policy Recommendations: A Four-Tier Graduated Framework

Based on the above analysis, this paper proposes a four-tier policy framework. At the government regulation tier, promoting algorithmic transparency legislation—mandating disclosure of content amplification mechanisms and ranking logic so the public can understand how information is selected and ordered. At the platform obligations tier, promoting the institutionalization of AI content labeling—the European Commission began work on a Code of Practice on Marking and Labelling AI-Generated Content in 2025, and the C2PA content credential standard should be advanced toward industry-standard status. At the industry standards tier, exploring the structural separation of “opinion producers” and “distribution infrastructure owners”—when the same entity simultaneously controls opinion production, algorithmic distribution, and audience measurement, it constitutes an historically unprecedented concentration of power, though specific separation criteria and boundaries (whether platform CEO posts are restricted, how media companies owning platforms are handled, how open-source algorithm platforms are treated) require further operationalization. At the education system tier, incorporating evolutionary psychology and cognitive biases into civic education so the public understands the structural features of its own cognitive system—not to generate panic, but to institutionalize “acquired immunity enhancement.” Tensions exist among the four tiers—especially between regulation and freedom of speech, and between antitrust and innovation space—and these tensions themselves are subjects requiring ongoing negotiation in democratic societies.

IXConclusion

9.1 Summary of Core Findings

Through a systematic interdisciplinary literature review and theoretical integration spanning 1895 to 2026, this paper arrives at the following core findings: Human following of opinion leaders is rooted in multiple evolutionarily shaped neural circuits, and the basic magnitude of susceptibility has remained significant over 130 years. The temporal topology on the transmission side has transformed from pulse-decay to low-frequency–cumulative–weak-decay, while the reception side’s cognitive hardware has not received a synchronous upgrade, constituting an evolutionary mismatch. The scaled expansion of AI content may be reorganizing the influence base of mid-tier opinion leaders through two channels—affective mirror alignment attenuation in the visual domain and signal-to-noise ratio collapse in the textual domain. The opinion leader ecosystem shows a trend of restructuring from a pyramid toward an hourglass plus ghost layer.

9.2 Theoretical Contributions

The theoretical contributions of this paper span five dimensions. First, it proposes the “dual-channel erosion” model, connecting social neuroscience (mirror alignment attenuation) with information economics (signal-to-noise ratio collapse) to explain the differentiated erosion of opinion leader influence across different domains by AI content. Second, it constructs a parasocial relationship typology, distinguishing authenticity-based following, fantasy parasocial relationships, functional dependence, and face-to-face trust, and clarifying the applicability boundaries of the hypotheses. Third, it constructs the “transmission-side topology shift vs. reception-side evolutionary mismatch” analytical framework, unifying 130 years of dispersed research under an evolutionary perspective. Fourth, it proposes a parameterizable hourglass model and the “ghost layer” concept. Fifth, it identifies the AI trust compensation mechanism as a boundary condition of the hourglass model.

9.3 Future Research Directions

Future research should proceed in six directions. First, fMRI/EEG verification of the five-group experimental design (including the critical Group D “human falsely informed as AI” condition). Second, longitudinal tracking of AI-personalized persuasion effects across different personality profiles. Third, information economics modeling of signal-to-noise ratio collapse in the textual domain. Fourth, cross-cultural comparison of hourglass shapes in collectivist versus individualist societies. Fifth, long-term impact of the disappearance of mid-tier knowledge creators on public discourse quality. Sixth, the challenge posed by human-AI hybrid creators to the hourglass model’s boundaries.