The Fake Majority: Inside the Invisible Machinery of Astroturfing. A serial mediation model

La mayoría falsa: dentro de la maquinaria invisible del astroturfing. Un modelo de mediación en serie

A maioria falsa: por dentro da máquina invisível do astroturfing. Um modelo de mediação serial

Tanase Tasente^1*
Cristina-Dana Popescu^1**

¹ Ovidius University of Constan.a, Romania
^* Lecturer at Ovidius University of Constan.a, Romania. Email: tanase.tasente@365.univ-ovidius.ro
^** Lecturer at Ovidius University of Constan.a, Romania. Email: cristina.popescu@365.univ-ovidius.ro

Abstract

Purpose: This study explores how perceived astroturfing and media literacy influence individuals’ critical responses to online content. It examines whether awareness of astroturfing and self-reported evaluation style mediate this relationship, challenging the assumption that media literacy alone ensures ethical digital behavior. Methodology: A cross-sectional web-based survey was conducted in Romania between March and April 2025 with a convenience sample of 226 participants aged 18 to 73. The questionnaire measured perceived media literacy, four dimensions of astroturfing awareness, evaluation style, and critical response. Data were analyzed using structural equation modelling (SEM) with serial mediation, employing diagonally weighted least squares (DWLS) and bootstrapped confidence intervals. Results: Media literacy did not directly predict critical response. However, it significantly increased astroturfing awareness, which in turn strongly predicted ethical behavior online (β = 0.56, p < .001). Evaluation style had a marginal role, and age was slightly associated with a more rational response style. The model showed excellent fit indices (CFI = 0.991, RMSEA = 0.03). Conclusions: Media literacy fosters ethical engagement only when users recognize coordinated manipulation. Astroturfing awareness functions as a catalyst for activating critical response, emphasizing the need to integrate detection of digital inauthenticity into educational curricula. Collateral information: Limitations include the cross-sectional design and reliance on self-reported measures. Implications highlight the importance of platform transparency and media education reform.

Keywords
Astroturfing; Media literacy; Critical response; Coordinated inauthentic behaviour; Social media ethics; Digital engagement.

Resumen

Propósito: Este estudio analiza cómo el astroturfing percibido y la alfabetización mediática influyen en las respuestas críticas de los individuos ante contenidos digitales. Examina si la conciencia sobre el astroturfing y el estilo de evaluación autodeclarado median esta relación, cuestionando la idea de que la alfabetización mediática garantiza por sí sola un comportamiento ético en línea. Metodología: Se realizó una encuesta transversal en línea en Rumanía entre marzo y abril de 2025 con una muestra por conveniencia de 226 participantes de entre 18 y 73 años. El cuestionario midió la alfabetización mediática percibida, cuatro dimensiones de conciencia sobre el astroturfing, estilo de evaluación y respuesta crítica. Los datos fueron analizados mediante modelos de ecuaciones estructurales (SEM) con mediación en serie, usando mínimos cuadrados ponderados diagonalmente (DWLS) e intervalos de confianza bootstrap. Resultados: La alfabetización mediática no predijo directamente la respuesta crítica, pero incrementó significativamente la conciencia sobre astroturfing, la cual predijo de forma fuerte el comportamiento ético en línea (β = 0.56, p < .001). El estilo de evaluación tuvo un rol marginal y la edad se asoció levemente con una respuesta más racional. El modelo mostró excelentes índices de ajuste (CFI = 0.991, RMSEA = 0.03). Conclusiones: La alfabetización mediática promueve una participación ética solo cuando los usuarios identifican manipulaciones coordinadas. La conciencia sobre astroturfing actúa como catalizador de respuestas críticas. Limitaciones: diseño transversal y medidas autoinformadas. Implicaciones: necesidad de mayor transparencia digital y reforma educativa en medios.

Palabras clave
Astroturfing; Alfabetización mediática; Respuesta crítica; Comportamiento inauténtico coordinado; Ética en redes sociales; Participación digital.

Resumo

Objetivo: Este estudo investiga como o astroturfing percebido e a literacia mediática influenciam as respostas críticas dos indivíduos ao conteúdo online. Examina se a consciência sobre astroturfing e o estilo de avaliação auto-relatado mediam essa relação, desafiando a suposição de ue a literacia mediática, por si só, garante um comportamento digital ético. Metodologia: Foi realizado um inquérito online transversal na Roménia entre março e abril de 2025, com uma amostra de conveniência de 226 participantes entre 18 e 73 anos. O questionário avaliou literacia mediática percebida, quatro dimensões da consciência sobre astroturfing, estilo de avaliação e resposta crítica. Os dados foram analisados por modelação de equações estruturais (SEM), com mediação serial, usando mínimos quadrados diagonalmente ponderados (DWLS) e intervalos de confiança com bootstrap. Resultados: A literacia mediática não previu diretamente a resposta crítica. No entanto, aumentou significativamente a consciência sobre astroturfing, que previu fortemente um comportamento ético online (β = 0.56, p < .001). O estilo de avaliação teve um papel marginal, e a idade associou-se ligeiramente a respostas mais racionais. O modelo apresentou ótimos índices de ajustamento (CFI = 0.991, RMSEA = 0.03). Conclusões: A literacia mediática promove o envolvimento ético apenas quando os utilizadores reconhecem manipulações coordenadas. A consciência sobre astroturfing atua como catalisador da resposta crítica, sugerindo a necessidade de integrar a deteção da inautenticidade digital nos currículos educativos. Limitações: desenho transversal e medidas auto-relatadas. Implicações: destacam-se a transparência das plataformas e a reforma da educação mediática.

Palavras-chave
Astroturfing; Literacia midiática; Resposta crítica; Comportamento inautêntico coordenado; Ética nas redes sociais; Engajamento digital.

1. Introduction

In the contemporary digital ecosystem, public opinion is no longer solely shaped by authentic civic mobilization or journalistic gatekeeping. Instead, an increasingly influential phenomenon - known as astroturfing - has gained traction across political, commercial, and activist spheres. Defined as the strategic simulation of grassroots support, astroturfing exploits digital affordances to construct the illusion of widespread public backing, while concealing the true source of influence behind orchestrated campaigns (Giovanelli & Persigo, 2022; Keller et al., 2020; Pineiro-Otero & Martinez Rolan, 2020). The origin of the term itself, a play on “grassroots” versus “artificial turf”, reveals its deceptive core: the fabrication of civic engagement to serve pre-established agendas.

Originating in the United States during the 1980s, astroturfing was initially applied in lobbying and corporate public relations to influence policy discussions under a false democratic facade (Boulay, 2012; Strong & Tyler, 2017). Today, however, the phenomenon has evolved into a global, technologically refined practice, often embedded within broader strategies of computational propaganda and algorithmic manipulation of visibility (Arce-Garcia et al., 2022; Chagas et al., 2022; Rauchfleisch & Chi, 2020). Unlike genuine grassroots activism, which emerges organically from shared values and collective will, astroturfing is characterized by covert strategic coordination, disguised sponsorship, and automated or incentivized actors posing as ordinary citizens (Baulch et al., 2024; Caro Castano, 2016).

Beyond deceptive influence, astroturfing raises profound ethical concerns. Scholars have highlighted its role in eroding public trust, manipulating information flows, and undermining democratic deliberation by fabricating online consensus (Boulay, 2012; Giovanelli & Persigo, 2022). The practice flourishes in digital environments driven by algorithmic popularity, where false signals of consensus can trend, mislead users, and pressure dissenters into silence (Zerback et al., 2021). A review of the literature reveals that astroturfing is not merely a fringe tactic but a pervasive strategy with tangible effects on perception, policy, and participation.

1.1. Literature Review

Core Mechanisms and Conceptual Distinctions

At its foundation, astroturfing operates through three central mechanisms: the camouflaging of true sponsorship, the systematic amplification of a message, and the construction of an artificial consensus. What distinguishes astroturfing from standard public relations or influencer marketing is its covert nature. Rather than overt endorsements, it involves paid or controlled agents - bots, sock puppets, troll armies - posing as grassroots voices (Baulch et al., 2024; Chagas et al., 2022; Keller et al., 2020). In doing so, orchestrators manufacture perceived support that far exceeds real public endorsement, contributing to the illusion of widespread support (Caro Castano, 2016; Johns & Cheong, 2019).

Closely related concepts include front groups (entities created by hidden sponsors to advocate specific positions), sock puppetry (individual users operating multiple fake accounts), and covert influencer marketing (undisclosed paid endorsements masquerading as personal opinions) (Pineiro-Otero & Martinez Rolan, 2020; Strong & Tyler, 2017). In each case, the defining element remains the same: covert coordination to simulate spontaneous civic expression.

Strategic Coordination in Digital Ecosystems

Digital platforms are fertile ground for astroturfing due to their reliance on engagement-driven algorithms and user anonymity. Researchers argue that the manipulation of algorithmic visibility - via coordinated hashtag pushes, synchronized likes, retweets, or shares - facilitates the virality of manipulated messages (Chagas et al., 2022; Iwan & Hed, 2023). Platforms such as Twitter, Instagram, and TikTok are particularly vulnerable due to their design emphasis on trending content, which astroturfing campaigns can exploit through concentrated short-term activity (Arce-Garcia et al., 2022; Elmas et al., 2019).

The method known as “Thunderclap” exemplifies this logic: by concentrating numerous posts within a narrow window, orchestrators simulate a grassroots surge that deceives algorithms into promoting visibility (Mahbub et al., 2019; Sorensen et al., 2017). Similarly, social media automation, particularly through bots, became a critical component for manipulating visibility and simulating widespread citizen support during Chile’s 2017 presidential elections (Santana & Huerta Canepa, 2019). In this context, strategic coordination becomes invisible, concealed behind dispersed user accounts that appear unrelated.

Political Astroturfing: Case-Specific Tactics

Astroturfing has become a staple of digital electioneering across the globe. In Brazil, Chagas et al. (2022) document how far-right digital brigades aligned with Jair Bolsonaro used coordinated hashtags to dominate Twitter trends, blending bots with pseudo-organic users. These campaigns not only promoted favorable narratives but also suppressed opposing voices. Similarly, during Mexico’s #SalarioRosa campaign, fake accounts inflated engagement to lend credibility to a government initiative (Espinoza & Pina-Garcia, 2023). In Indonesia’s 2019 elections, Baulch et al. (2024) observed that “buzzers” operated in hierarchical structures, releasing memes and slogans via WhatsApp in both positive (white) and negative (black) forms. Encrypted platforms amplified the opacity of sponsorship, making it more difficult for observers to trace the content to its funders. These instances highlight how cultural and technological contexts shape astroturfing’s manifestations, but the underlying goal remains: to manipulate perceived support and guide public opinion.

Corporate Astroturfing and Influencer Manipulation

Commercial entities also employ astroturfing to promote products or discredit rivals. In Spain, the #Pentatrillones beer campaign used microcelebrities to simulate viral enthusiasm, only for users to later uncover the orchestrated nature of the endorsement - damaging both the influencers’ reputations and audience trust (Caro Castano, 2016). This aligns with digital ethics concerns, as audiences feel betrayed when camouflaged sponsorship is revealed.

Astroturfing’s expansion into influencer culture reflects its integration into hybrid media systems, where online trends are frequently mirrored or legitimized by mainstream outlets (Chadwick, 2013; Leerssen et al., 2023). As content circulates across platforms, manipulated visibility can translate into real-world legitimacy, underscoring the dangers of false consensus.

Ethical and Democratic Implications

The rise of astroturfing signals a profound clash between democratic transparency and manipulative digital strategies. Scholars emphasize that the practice erodes trust in online dialogue, as users struggle to distinguish real activism from orchestrated performance (Boulay, 2012; Zerback et al., 2021). When citizens become skeptical of all digital mobilization, genuine movements suffer, and public deliberation becomes polluted by distrust and confusion (Giovanelli & Persigo, 2022). From an ethical standpoint, astroturfing violates core tenets of professional communication, such as honesty, openness, and transparency about sources (Strong & Tyler, 2017). These breaches are compounded by social platforms’ often vague or weak enforcement of anti-astroturfing policies, which allows campaigns to re-emerge quickly with new accounts or altered methods (Keller et al., 2020).

Detection, Resistance, and Ongoing Challenges

While detection methods have advanced - including network analysis, linguistic patterning, and bot detection - many campaigns blend real and fake elements in ways that circumvent automated systems (Chagas et al., 2022; Keller et al., 2020). Hybrid tactics, ephemeral content, and encrypted platforms limit data visibility, complicating scholarly and journalistic investigation (Baulch et al., 2024).

Responses must extend beyond technical fixes. Scholars advocate for critical digital literacy, transparency standards, platform accountability, and cross-sector collaboration (Zerback et al., 2021). Yet any effective strategy must account for local contexts, as astroturfing adapts to platform preferences, political norms, and user behavior (Caro Castano, 2016; Oguz, 2018).

Astroturfing in social media represents more than a deceptive tactic; it constitutes a systemic form of covert persuasion that leverages the dynamics of online communication to alter public perception. Its key mechanisms - camouflaged sponsorship, algorithmic exploitation, and the illusion of public support - enable orchestrators to shape debates, influence elections, and redirect consumer choices without disclosure or accountability.

The literature consistently identifies invisible coordination, fabricated consensus, and platform vulnerability as structural facilitators of astroturfing. In response, communication science must refine its conceptual, methodological, and ethical tools to resist this erosion of democratic discourse. Astroturfing is not an anomaly but a strategic adaptation to the logic of digital media, and only through sustained scholarly, technological, and civic engagement can its influence be mitigated.

1.2. Objectives of the Study

Starting from the gaps identified in the literature, this study proposes an in-depth analysis of the existing relationships between astroturfing, media literacy, and critical response, with the main objective of highlighting the fact that people’s critical response is influenced by astroturfing and media literacy, depending on how people perceive themselves responding to messages in the online environment.

Therefore, we hypothesize that individuals who assess themselves as predominantly investigative-rational in terms of their reactions to online messages (H₁), who have higher levels of astroturfing (H₂), and higher media literacy (H₃), will have a more pronounced critical response. At the same time, we propose that high levels of media literacy will lead to greater astroturfing awareness (H₄), which will determine reactions to online messages toward the investigative-rational pole (H⁵) and lead to a more critical response, i.e., astroturfing and evaluation style will mediate the relationship between media literacy and critical response.

As a self-assessment, the evaluation of online reactions could be influenced by age (CV₁), education level (CV₂), and time spent online (CV₃). Controlling for these three variables at the level of the second mediator, the proposed model is presented in Figure 1.

2. Method

We used CHERRIES checklist (Eysenbach, 2004) aimed to assess the quality of Web Surveys to enhance the methodology.

Design. Data were collected from March 15 to March 20, 2025, using an online form distributed among the target population, and a convenience sample was employed. The analysis of the model was conducted using a cross-sectional design, which precluded drawing causal inferences. Nonetheless, despite the analysis involving a structural equation model with a serial mediation effect, the results are presented descriptively, with causal verification planned to be conducted in future research.

Informed consent process. The informed consent form clearly stated that participation in the study was voluntary and that participants could withdraw at any time. The research was conducted in accordance with the guidelines of the Declaration of Helsinki and the ethical standards set by the Ethics Committee of the faculty with which the authors are affiliated. The estimated time required to complete the questionnaire was approximately 5-7 minutes. Initially, data were securely stored on Lime Survey (https://research.pluscommunication.eu/index.php/213198?lang=ro) before being exported into Excel format, after which the files were deleted from the platform. All participant responses as well as selected socio-demographic data (gender, age, education, social network platform) were retained. The authors were responsible for data collection, and the data will be archived and stored for five years following publication. No additional personal information was collected, and the data collection process was conducted anonymously.

Development and pre-testing. The data collection materials were initially pre-tested with a sample of 50 students from the Faculty of Law and Administrative Sciences at Ovidius University of Constanța. The questionnaire was promoted both during class sessions and through the students’ WhatsApp groups. Based on the feedback received during this stage, several items were revised to improve clarity and relevance. The final version was retested before the official data collection phase began.

Recruitment process. The online survey was distributed using a combination of targeted outreach and snowball sampling. Participation was voluntary, and all respondents were informed that their responses would be confidential, anonymous, and used exclusively for research purposes. To ensure a diverse pool of participants, the questionnaire was initially promoted through official communication channels - such as email invitations sent to the major parliamentary political parties in Romania, private messages to their official Facebook pages, and direct WhatsApp messages to individuals in leadership roles. To further expand the reach, the survey link was also shared on social media platforms, including Facebook and WhatsApp, with participants encouraged to forward it within their own networks.

Survey administration. The questionnaires were designed using Lime Survey and disseminated via social media platforms. Participation was voluntary, and no rewards were offered. Data were collected from April 1 to April 24, 2025. The questionnaire comprised 25 items presented over six pages, with no adaptive or randomization procedures employed. Participants were permitted to review their responses, but no completeness checks were implemented.

Response rates. The survey was accessed by a total of 3,273 individuals. Out of these, 226 respondents fully completed the questionnaire, while 3,047 provided partial responses. This corresponds to a completion rate of approximately 6.90%, highlighting potential challenges in maintaining participant engagement throughout the survey.

Preventing multiple entries. To ensure the uniqueness of each response and maintain data integrity, the survey was configured in LimeSurvey to allow only one submission per IP address. This technical restriction effectively prevented duplicate entries. Consequently, no duplicate submissions were identified in the final dataset.

2.1. Participants and Procedure

Data were collected online from 226 Romanian participants aged 18 to 73 years (M=44.26, SD=12.64), of whom identified as female (73.45%). Regarding education level, most respondents held a Master’s or doctoral degree (65.93%), followed by those with a Bachelor’s degree (28.76%). A smaller proportion reported having completed high school (4.42%) or post-secondary education (0.88%). In terms of daily internet use, most participants reported spending between one and two hours online daily (37.61%), followed closely by those who spent two to four hours (31.86%). Fewer respondents indicated spending four to six hours online (8.85%), while some reported spending either less than one hour (16.37%) or more than six hours per day (5.31%).

When assessing their own level of media literacy, most participants rated themselves as having a high (47.35%) or medium (31.86%) level. A smaller share considered their skills to be very high (18.14%), while only a few indicated low (2.21%) or very low levels (0.44%). In terms of preferred evaluation style in the online environment, the reflective style was most commonly reported (1.77%), followed by the alternative (26.99%) and emotional styles (0.88%). Only a small number of participants identified with the social (11.95%) or automatic (58.41%) evaluation styles. Participants were informed about their right to withdraw from the study at any time, and the informed consent form clearly stated that all data collected would be used exclusively for research purposes. Completing the questionnaire required approximately 5–7 minutes. The study was conducted in accordance with the ethical standards of the Declaration of Helsinki (2013) and received approval from the Ethics Committee of the authors’ affiliated institution.

2.2. Measures

Media literacy was measured using a single ordinal item asking participants to rate their perceived media literacy on a 5-point Likert scale ranging from 1 (very low) to 5 (very high).

The Astroturfing dimension was evaluated using a multidimensional survey instrument comprising four distinct factors, each consisting of five ordinal items with a Likert-type response scale ranging from 1 (never) to 5 (very often). The first factor, Invisible Strategic Coordination, comprised five Likert-type items, including “Some online campaigns seem designed just to appear popular,” which demonstrated good internal consistency in our data (α=0.8, 95% CI[0.75, 0.83]). The second factor, Constructing the Illusion of Public Support, consisted of five Likert-type items, such as “I usually notice suspiciously synchronized activity between unknown accounts,” and exhibited very good internal consistency for all five items (α=0.87, 95% CI[0.85, 0.9]). The third factor, Concealment of the Real Source, also contained five Likert-like items, including “It seems to me that many viral campaigns do not indicate who initiated them.” Our research sample showed excellent scale consistency (α=0.83, 95% CI[0.79, 0.86]). The final factor, Algorithmic Manipulation of Visibility, comprised five Likert items, such as “Some ideas seem to gain attention solely due to false or bought reactions,” and yielded very good internal consistency (α=0.89, 95% CI[0.87, 0.91]) among the participants investigated. The hierarchical omega coefficient Jensen and Weng (1994), Zinbarg et al. (2005), Revelle and Condon (2019) was calculated to determine whether the factors were cumulative and whether a total score for Astroturfing could be obtained. The results indicate high values for both the overall coefficient and the hierarchical coefficient (ω_hierarchical=0.82, ω_general=0.95). Confirmatory factor analysis (CFA) provided support for these findings, with the proposed model exhibiting a good fit to the data under the assumption of correlated factors (oblique assumption), thereby establishing its validity and reliability for research purposes (χ²=153.329, df=166, p= 0.751, CFI=1, SRMR=0.043, RMSEA approx 0, 90% CI [0, 0.023]).).

The evaluation style was assessed using a single item, where participants rated their personal approach to online content on a scale ranging from automatic-emotional (low values) to rational-investigative (high values). The response options were as follows: Automatic (1), if it is viral, I am interested; Emotional (2), I react to how it makes me feel; Social (3), I follow what my friends support; Reflective (4), I analyze the source and context; and Alternative (5), I check other platforms or sources.

Initially, the Critical Response and Digital Ethics scale comprised five Likert-type items, including “I choose not to repost when I don’t know who created it.” However, the scale’s consistency was compromised due to the item “I alert others when I identify staged campaigns,” which exhibited poor correlation with the total score and the other items. Consequently, removing this item yielded a 4-item Likert scale with satisfactory internal consistency (α=0.79, 95% CI[0.74, 0.83]).

2.3. Overview of Statistical Analysis

We used R (Version 4.5.0; R Core Team, 2025) and the R-packages depmixS4 (Version 1.5.0; Visser & Speekenbrink, 2010), devtools (Version 2.4.5; Wickham et al., 2022), dplyr (Version 1.1.4; Wickham et al., 2023), ggplot2 (Version 3.5.2; Wickham, 2016), Hmisc (Version 5.2.3; Harrell Jr, 2025), kableExtra (Version 1.4.0; Zhu, 2024), lavaan (Version 0.6.19; Rosseel, 2012), limer (Version 0.1; Heiss, 2015), MASS (Version 7.3.65; Venables & Ripley, 2002a), mvtnorm (Version 1.3.3; Genz & Bretz, 2009), nlme (Version 3.1.168; Pinheiro & Bates, 2000), nnet (Version 7.3.20; Venables & Ripley, 2002b), papaja (Version 0.1.3; Aust & Barth, 2024), polycor (Version 0.8.1; Fox, 2022), psych (Version 2.5.3; Revelle, 2024), reshape2 (Version 1.4.4; Wickham, 2007), Rsolnp (Version 1.16; Ghalanos & Theussl, 2015), rstatix (Version 0.7.2; Kassambara, 2023), sasLM (Version 0.10.5; Bae, 2024), tidyLPA (Version 1.1.0; Rosenberg et al., 2018), tinylabels (Version 0.2.5; Barth, 2023) and usethis (Version 3.1.0; Wickham et al., 2024) for all our analyses.

Internal consistency for all factors was evaluated using the Cronbach’s α index (Cronbach, 1951), a widely accepted measure of reliability. For multifactorial instruments, a total score can be calculated if a general latent factor is present, influencing other factors. This general factor can be used to account for the common variance among the different factors. To further evaluate the internal consistency of the total score, omega coefficients were estimated by performing a factor analysis on the original dataset. The factors were rotated obliquely, allowing for the correlation between factors, and a Schmid-Leiman transformation was applied as recommended by Revelle & Condon (2019). This transformation enabled the separation of the general and group factors, providing a more detailed understanding of the internal structure of the instrument. The total omega coefficient was used to assess the internal consistency of the total score, representing the proportion of variance explained by the general and group factors. This coefficient provides a comprehensive measure of the reliability of the total score, taking into account the correlation between the factors. In addition to the total omega coefficient, the proportion of variance in the group explained solely by the general factor was evaluated using the omega coefficient general (hierarchical). This coefficient provides insight into the extent to which the general factor contributes to the variance in the group factors, allowing for a more nuanced understanding of the relationships between the factors. To further validate the internal structure of the instrument, a confirmatory factor analysis (CFA) was conducted using the diagonally weighted least squares (DWLS) estimation method and assuming correlated factors (DiStefano & Morgan, 2014). This approach allowed for the testing of the possibility of using a latent trait for multidimensional instruments, providing a more accurate representation of the underlying structure of the instrument. The results of the CFA provided support for the use of a latent trait for the multifactorial instrument, indicating that the general factor was a significant contributor to the variance in the group factors. The findings of this study suggest that the instrument has a robust internal structure, with a general factor that explains a significant proportion of the variance in the group factors.

We initiated the process of identifying outliers and missing values, and performed univariate descriptive analyses. This involved evaluating compliance with the assumption of univariate normality for continuous data using the Shapiro–Wilk test (Royston, 1982; Royston, 1995) and examining skewness and kurtosis values. Outliers were replaced with missing values if they exceeded three standard deviations from the mean. To assess multivariate normality, we utilized the Mardia test (Mardia, 1970) under the null hypothesis that the variables follow a multivariate normal distribution. Additionally, a correlation matrix was computed and described for the dependent, independent, and mediator variables.

The SEM model of serial mediation was estimated using diagonally weighted least squares (DWLS) (DiStefano & Morgan, 2014). This approach allows for the examination of the model’s parameters while accounting for potential issues related to non-normality and the use of ordinal indicators. Standard errors were calculated using nonparametric bootstrapping resampling with 200 replicated samples. This approach is particularly useful when the data exhibits non-constant variance, as it can lead to more accurate and reliable standard error estimates. In addition to standard error estimation, the Satorra-Bentler test was employed (Satorra & Bentler, 2001) as the omnibus statistical test. This test is a chi-square test that is specifically designed for use with SEM models, and it provides a way to evaluate the overall fit of the model to the data. The Satorra-Bentler test is a more robust alternative to the traditional chi-square test, as it is less sensitive to non-normality in the data. The use of DWLS estimation, bootstrapped standard error estimation, and the Satorra-Bentler test in the evaluation of the serial mediated SEM model provides a comprehensive approach to assessing the model’s fit and parameters. This approach can help to increase the accuracy and reliability of the results, and it can provide a more nuanced understanding of the relationships between the variables in the model.

3. Results

3.1. Preliminary Descriptive Analysis

Outlier analysis revealed no extreme univariate scores. However, the data exhibited slightly elevated low values for “building illusion of public support” (5, 6, and 9), “invisible strategic coordination” (5), “disguising the real source” (5, 8, 9, and 10), “algorithmic visibility manipulation” (5 and 7), and “critical reaction and digital ethics” (4, 6, 7, 8, and 9). These outliers were retained in the analysis; however, the assumption of univariate normality was not met for any continuous variable (see Table 1).

The distributions of all variables showed significant negative skewness, while those of “building illusion of public support,” “algorithmic visibility manipulation,” and “critical reaction and digital ethics” demonstrated significant leptokurtosis, necessitating the usage of robust analytical techniques. The assumption of multivariate normality was not met due to the Mahalanobis distances varied between 0.63 and 5.42, and as also suggested by the Mardia coefficient (Mardia, 1970). Significant positive multivariate skewness (Mardia=5.57, Skewness=209.77, p<.001) and multivariate leptokurtosis (Mardia=44.77, Skewness=8.78, p<.001) were also observed.

3.2. Corelation Analysis

The Bravais-Pearson correlation coefficient was not applicable; therefore, the matrix derived from the Spearman correlation coefficient (r_s) revealed statistically significant positive associations between all variables, ranging from .32 to .68, and was characterized as positive (see Table 2).

Building illusion of public support was positively associated with invisible strategic coordination (ρ=.59, p<.001), disguising the real source (ρ=.62, p<.001), algorithmic visibility manipulation (ρ=.56, p<.001), and critical reaction and digital ethics (ρ=.39, p<.001). Invisible strategic coordination was positively associated with disguising the real source (ρ=.62, p<.001), algorithmic visibility manipulation (ρ=.62, p<.001), and critical reaction and digital ethics (ρ=.32, p<.001). Disguising the real source was positively associated with algorithmic visibility manipulation (ρ=.68, p<.001), and critical reaction and digital ethics (ρ=.44, p<.001), and algorithmic visibility manipulation was positively associated with critical reaction and digital ethics (ρ=.44, p<.001).

3.3. Model Analysis

Further analysis of the proposed research model revealed convergence after 60 iterations, with the estimation of 27 parameters based on 226 data points, yielding very good fit indices without significant modification indices (χ²=31.363, df=26, p=.215, CFI=0.991, SRMR=0.023, RMSEA=0.03, p= 748, 90% CI [0, 0.071]; scaled and robust statistics were used).

Our findings (see Table 3) indicated that there was no significant probabilistic total effect of media literacy on critical response and digital ethics, encompassing all direct and indirect pathways (B=0.31, z=1.17, p=.241, β=0.08, 95% CI [-0.27, 0.99]). The analysis revealed that critical response and digital ethics were not found to have a statistically significant associations with media literacy (B=0.30, z=1.16, p=.248, β=0.07, 95% CI [-0.27, 0.97]). Conversely, a statistically significant association was observed between critical response and digital ethics and astroturfing (B=0.60, z=6.00, p<.001, β=0.56, 95% CI [0.34, 0.76]), while a marginally significant relationship was identified between evaluation style and critical response and digital ethics (B=0.37, z=1.66, p=.097, β=0.12, 95% CI [-0.16, 0.81]). Therefore, high scores on critical response and digital ethics were associated with high scores on astroturfing and, marginally, with high scores on evaluation style, oriented towards the rational-investigative pole.

On the other hand, high media literacy scores exhibited a statistically significant and positive correlation with both high astroturfing scores (B=0.90, z=3.05, p=.002, β=0.24, 95% CI [0.24, 1.44]) and high evaluation style scores (B=0.23, z=2.03, p=.042, β=0.17, 95% CI [-0.02, 0.49]). Consequently, it can be inferred that astroturfing fully mediated (B=0.54, z=3.17, p=.002, β=0.13, 95% CI [0.18, 0.97]) the relationship between media literacy and critical response and digital ethics, with a statistically significant effect (see Figure 2). A high level of media literacy correlated with high levels of astroturfing, which was also indicative of high levels of critical response and digital ethics. No other statistically significant relationships were identified, and of the three controled variables, only age was marginally significantly and positively associated with evaluation style (B=0.01, z=1.88, p= 0.059, β=0.12, 95% CI[0.00, 0.02]). It appears that as age increases, individuals tend to lean more towards a rational and investigative approach, whereas younger people are more likely to perceive themselves as having a more emotionally driven style.

4. Discussion

Astroturfing has been theorized as a distinct mode of strategic communication that manipulates the cues individuals rely upon to judge social consensus, rather than the factual content of messages themselves (Rauchfleisch & Chi, 2020). This theoretical insight proved essential in interpreting the relationships observed in the current study among media literacy, perceived manipulation, and ethical online behaviour.

Respondents who regarded themselves as skilled in navigating digital information did not automatically demonstrate a more critical stance toward content. Rather, critical engagement intensified only when these users also identified cues suggesting that the popularity of a message might be artificially engineered - through patterns such as synchronised posting or endorsements lacking transparent sponsorship, phenomena repeatedly observed in covert communication campaigns (Chagas et al., 2022; Keller et al., 2020). These findings challenge the pedagogical assumption that knowledge alone inoculates individuals against manipulation (Giovanelli & Persigo, 2022). Literacy, in this framework, is less a constant shield than a latent capacity, activated in the presence of contextual signals indicating procedural irregularities.

This mechanism aligns with a broader understanding of astroturfing as the strategic fabrication of social proof. Once the orchestration behind apparent public endorsement becomes perceptible, individuals tend to reframe the persuasive attempt as a breach of communicative norms, prompting defensive reactions such as source verification, disengagement, or peer warning (Boulay, 2012).

Perceived astroturfing thus functions as a psychological bridge, linking digital competence with concrete ethical behaviour. Users who rated their literacy as high were also more adept at noticing covert coordination, supporting the view that media education sensitises individuals to violations of communicative transparency (Pineiro-Otero & Martinez Rolan, 2020). Once such deviations became salient, persuasion knowledge moved from passive awareness to active resistance, with users adopting cautious and reflective strategies.

Self-reported evaluation style added a secondary, though meaningful, explanatory layer. Participants describing themselves as rational and investigative tended to behave more responsibly online, yet this disposition alone could not explain the marked behavioural shifts seen in response to astroturfing cues. This reinforces the idea that analytical motivation, while valuable, requires activation through explicit contextual triggers - a pattern mirrored in studies of influence operations reliant on synchronised amplification to obscure sponsorship (Baulch et al., 2024; Caro Castano, 2016).

Age introduced only a modest refinement: older participants showed a slight preference for deliberative evaluation, consistent with research suggesting that experience encourages scepticism towards algorithmic indicators of popularity (Boulay, 2012). Nonetheless, age did not alter the core pathway, confirming that contextual perception of manipulation - rather than demographic attributes - governs the activation of critical engagement.

These findings invite several conceptual and practical extensions. First, media-literacy education should expand beyond fact-checking and source authentication to include training in detecting orchestration. Heuristics such as recognising spikes in identical messages, analysing account creation timelines, or observing cross-platform diffusion can help users detect exactly the features that triggered critical engagement in this study. This recommendation echoes recent proposals for literacy models that highlight manipulations of visibility, not just of content accuracy (Zerback et al., 2021).

Second, platform design can support educational initiatives by making indicators of coordinated amplification more perceptible. As shown in studies of electoral manipulation via hashtags (Chagas et al., 2022), clusters of newly created accounts can distort trending metrics. User interfaces could flag anomalies - such as rapid, synchronised spread or coordination indicators - lowering the threshold at which users perceive manipulation and activating the dormant literacy capacities identified here.

5. Conclusion

The present study advances the understanding of digital resilience by demonstrating that media literacy becomes effective not by default, but when manipulation is perceptible. Users’ critical faculties were not continuously engaged, but rather triggered by cues of inauthentic coordination, suggesting that awareness of astroturfing is the necessary catalyst that transforms latent competence into active ethical behaviour.

This insight carries several practical implications. Educational efforts must go beyond propositional truth and emphasize recognition of engagement manipulation as a distinct threat. Equipping citizens with the ability to detect artificial amplification—through both training and supportive platform cues—can significantly enhance resistance to persuasive deception. At the same time, digital platforms bear responsibility for increasing the visibility of coordination patterns, enabling users to perceive what otherwise remains obscured.

Several methodological limitations should temper the generalization of results. The cross-sectional nature of the data restricts causal inference; future longitudinal and experimental studies are needed to test the temporal ordering of awareness and behavioural change. The convenience sample, composed largely of highly educated Romanian adults, may exhibit analytic habits not fully representative of broader populations. Replication across different cultural and educational contexts would clarify the generalisability of the mechanism. Additionally, the reliance on self-reported behaviour introduces a risk of socially desirable inflation; complementing survey data with behavioural traces—such as sharing decisions after exposure to known-content material—would bolster inferential validity. Finally, while this study focused on visible coordination, future research should explore covert, encrypted, or ephemeral messaging environments, which already shape political mobilisation in various regions (Baulch et al., 2024).

The findings suggest that astroturfing compromises deliberative integrity not by challenging the content of discourse, but by simulating public consensus. Addressing this challenge requires integrating mechanisms for detecting inauthentic coordination into both media literacy curricula and platform design. Such efforts can enhance the likelihood that users' latent competencies are activated in the presence of manipulation, thereby fostering more ethical and critically engaged digital behaviour.

Author’s Contribution

Tanase Tasante: Conceptualisation; Methodology; Project management; Investigation; Supervision; Resources; Data curation; Formal analysis; Visualisation; Writing – original draft; Writing – review and editing.

Cristina-Dana Popescu: Resources; Writing – original draft; Writing – review and editing.

All authors have read and accepted the published version of the manuscript.

Conflicts of Interest

The authors declare that they have no conflict of interest.

References

Arce-Garcia, Sergio; Said-Hung, Elías; & Mottareale-Calvanese, Daría (2022). Astroturfing as a strategy for manipulating public opinion on Twitter during the pandemic in Spain. Profesional De La Informacion, 31(3). https://doi.org/10.3145/epi.2022.may.10

Aust, Frederik; & Barth, Marius (2024). papaja: Prepare reproducible APA journal articles with R Markdown. https://doi.org/10.32614/CRAN.package.papaja

Bae, Kyun-Seop (2024). sasLM: ’SAS’ linear model. https://CRAN.R-project.org/package=sasLM

Barth, Marius (2023). tinylabels: Lightweight variable labels. https://cran.r-project.org/package=tinylabels

Baulch, Emma; Matamoros-Fernández, Ariadna; & Suwana, Fiona (2024). Memetic persuasion: And WhatsAppification in Indonesia’s 2019 presidential election. NEW MEDIA & SOCIETY, 26(5), 2473–2491. https://doi.org/10.1177/14614448221088274

Boulay, Sophie (2012). Which consideration(s) for ethics within the use of information and communication technologies by public relations? An analysis of astroturfing cases and a critical standpoint. Revista Internacional De Relaciones Publicas, 2(4), 201–220. https://doi.org/10.5783/RIRP-4-2012-09-201-220

Caro Castano, Lucía (2016). Microcelebrities and online astroturfing in spain. The case of # pentatrillones. Adcomunica-Revista Cientifica De Estrategias Tendencias E Innovacion En Communicacion, 11, 121–144. https://doi.org/10.6035/2174-0992.2016.11.8

Chadwick, Andrew (2013). The Hybrid Media System: Politics and Power. Oxford University Press. https://doi.org/10.1093/acprof:oso/9780199759477.001.0001

Chagas, Viktor, Carreiro, Rodrigo, Santos, Nina, & Popolin, Guilherme (2022). Far-right digital activism in polarized contexts: A comparative analysis of engagement in hashtag wars. Media and Communication, 10(4), 42–55. https://doi.org/10.17645/mac.v10i4.5622

Cronbach, Lee J. (1951). Coefficient alpha and the internal structure of tests. Psychometrika, 16(3), 297–334. https://doi.org/10.1007/BF02310555

DiStefano, Christine; & Morgan, Grant B. (2014). A Comparison of Diagonal Weighted Least Squares Robust Estimation Techniques for Ordinal Data. Structural Equation Modeling: A Multidisciplinary Journal, 21(3), 425–438. https://doi.org/10.1080/10705511.2014.915373

Elmas, Tuğrulcan; Overdorf, Rebekah; Özkalay, Ahmed Furkan; & Aberer, Karl (2019). Ephemeral Astroturfing Attacks: The Case of Fake Twitter Trends. arXiv. https://doi.org/10.48550/ARXIV.1910.07783

Espinoza, Armando; & Piña-García, Carlos A. (2023). Propaganda and manipulation in mexico: A programmed, coordinated and manipulative “pink” campaign. Journalism and Media, 4(2), 578–598. https://doi.org/10.3390/journalmedia4020037

Eysenbach, Gunther (2004). Improving the Quality of Web Surveys: The Checklist for Reporting Results of Internet E-Surveys (CHERRIES). Journal of Medical Internet Research, 6(3), e34. https://doi.org/10.2196/jmir.6.3.e34

Fox, John (2022). Polycor: Polychoric and polyserial correlations. https://CRAN.R-project.org/package=polycor

Genz, Alan; & Bretz, Frank (2009). Computation of multivariate normal and t probabilities. Springer-Verlag.

Ghalanos, Alexios; & Theussl, Stefan (2015). Rsolnp: General non-linear optimization using augmented lagrange multiplier method.

Giovanelli, Letícia de Oliveira; & Persigo, Patricia Milano (2022). Public relations and astroturfing from a relational perspective. Revista Internacional De Relaciones Públicas, 12(23), 73–92. https://doi.org/10.5783/RIRP-23-2022-05-73-92

Harrell , Frank E. (2025). Hmisc: Harrell miscellaneous. https://CRAN.R-project.org/package=Hmisc

Heiss, Andrew (2015). Limer: A LimeSurvey r client.

Iwan, Nurul; & Hed, Norhafiza Mohd M. (2023). Definition and typology of the term political cybertroopers in malaysia br. Jurnal Komunikasi-Malaysian Journal Of Communication, 39(3), 81–100. https://doi.org/10.17576/JKMJC-2023-3903-05

Jensen, Arthur R.; & Weng, Li-Jen (1994). What is a good g? Intelligence, 18(3), 231–258. https://doi.org/10.1016/0160-2896(94)90029-9

Johns, Amelia; & Cheong, Niki (2019). Feeling the chill: Bersih 2.0, state censorship, and “networked affect” on malaysian social media 2012-2018. Social Media + Society, 5(2). https://doi.org/10.1177/2056305118821801

Kassambara, Alboukadel (2023). Rstatix: Pipe-friendly framework for basic statistical tests. https://CRAN.R-project.org/package=rstatix

Keller, Franziska B.; Schoch, David; Stier, Sebastian; & Yang, JungHwan (2020). Political astroturfing on twitter: How to coordinate a disinformation campaign. Political Communication, 37(2), 256–280. https://doi.org/10.1080/10584609.2019.1661888

Leerssen, Paddy; Dobber, Tom; Helberger, Natali; & de Vreese, Claes (2023). News from the ad archive: How journalists use the facebook ad library to hold online advertising accountable. Information Communication & Society, 26(7), 1381–1400. https://doi.org/10.1080/1369118X.2021.2009002

Mahbub, Syed; Pardede, Eric; Kayes, A. S. M.; & Rahayu, Wenny (2019). Controlling astroturfing on the internet: A survey on detection techniques and research challenges. International Journal of Web and Grid Services, 15(2), 139. https://doi.org/10.1504/IJWGS.2019.099561

Mardia, Kanti (1970). Measures of multivariate skewness and kurtosis with applications. Biometrika, 57(3), 519–530. https://doi.org/10.1093/biomet/57.3.519

Oguz, Mustafa Cem (2018). “Networked authoritarianism” or the relation of authoritarian regimes and cyber space: A literature review on Russia and China. Turkiye Iletisim Arastirmalari Dergisi-Turkish Review of Communication Studies, 31, 103–120. https://doi.org/10.17829/turcom.415025

Pineiro-Otero, Teresa; & Martinez Rolan, Xabier (2020). Understanding Digital Politics Principles and Actions. Vivat Academia, 152, 19–47. https://doi.org/10.15178/va.2020.152.19-48

Pinheiro, José C.; & Bates, Douglas M. (2000). Mixed-effects models in s and s-PLUS. Springer. https://doi.org/10.1007/b98882

R Core Team. (2025). R: A language and environment for statistical computing. R Foundation for Statistical Computing. https://www.R-project.org/

Rauchfleisch, Adrian; & Chi, Jo-yao (2020). Untangling Taiwan’s hybridity with structural dysfunctions. Social Media + Society, 6(3). https://doi.org/10.1177/2056305120947658

Revelle, William (2024). Psych: Procedures for psychological, psychometric, and personality research. Northwestern University. https://CRAN.R-project.org/package=psych

Revelle, William; & Condon, David M. (2019). Reliability from α to ω: A tutorial. Psychological Assessment, 31(12), 1395–1411. https://doi.org/10.1037/pas0000754

Rosenberg, Joshua M.; Beymer, Patrick N.; Anderson, Daniel J.; van Lissa, C. J.; & Schmidt, Jennifer A. (2018). tidyLPA: An r package to easily carry out latent profile analysis (LPA) using open-source or commercial software. Journal of Open Source Software, 3(30), 978. https://doi.org/10.21105/joss.00978

Rosseel, Yves (2012). lavaan: An R package for structural equation modeling. Journal of Statistical Software, 48(2), 1–36. https://doi.org/10.18637/jss.v048.i02

Royston, Patrick (1982). An Extension of Shapiro and Wilk’s W Test for Normality to Large Samples. Journal of the Royal Statistical Society. Series C (Applied Statistics), 31(2), 115–124. https://doi.org/10.2307/2347973

Royston, Patrick (1995). Remark AS R94: A Remark on Algorithm AS 181: The W-test for Normality. Journal of the Royal Statistical Society. Series C (Applied Statistics), 44(4), 547–551. https://doi.org/10.2307/2986146

Santana, Luis E.; & Huerta Canepa, Gonzalo (2019). Are they bots? Social media automation during Chile’s 2017 presidential campaign. Cuadernos Info, 44, 61–77. https://doi.org/10.7764/cdi.44.1629

Satorra, Albert; & Bentler, Peter M. (2001). A scaled difference chi-square test statistic for moment structure analysis. Psychometrika, 66(4), 507–514. https://doi.org/10.1007/BF02296192

Sorensen, Anne; Andrews, Lynda; & Drennan, Judy (2017). Using social media posts as resources for engaging in value co-creation: The case for social media-based cause brand communities. Journal of Service Theory and Practice, 27(4), 898–922. https://doi.org/10.1108/JSTP-04-2016-0080

Strong, Catherine; & Tyler, Fran (2017). New Zealand media camouflage political lobbying. Pacific Journalism Review, 23(2), 144–158. https://doi.org/10.24135/pjr.v23i2.96

Venables, William N.; & Ripley, Brian D. (2002a). Modern applied statistics with s (Fourth). Springer. https://www.stats.ox.ac.uk/pub/MASS4/

Venables, William N.; & Ripley, Brian D. (2002b). Modern applied statistics with s (Fourth). Springer. https://www.stats.ox.ac.uk/pub/MASS4/

Visser, Ingmar; & Speekenbrink, Maarten (2010). depmixS4: An R package for hidden markov models. Journal of Statistical Software, 36(7), 1–21. https://www.jstatsoft.org/v36/i07/

Wickham, Hadley (2007). Reshaping data with the reshape package. Journal of Statistical Software, 21(12), 1–20. http://www.jstatsoft.org/v21/i12/

Wickham, Hadley (2016). ggplot2: Elegant graphics for data analysis. Springer-Verlag New York. https://ggplot2.tidyverse.org

Wickham, Hadley; Bryan, Jennifer; Barrett, Malcolm; & Teucher, Andy (2024). Usethis: Automate package and project setup. https://CRAN.R-project.org/package=usethis

Wickham, Hadley; François, Romain; Henry, Lionel; Müller, Kirill; & Vaughan, Davis (2023). Dplyr: A grammar of data manipulation. https://CRAN.R-project.org/package=dplyr

Wickham, Hadley; Hester, Jim; Chang, Winston; & Bryan, Jennifer (2022). Devtools: Tools to make developing r packages easier. https://CRAN.R-project.org/package=devtools

Zerback, Thomas; Toepfl, Florian; & Knoepfle, Maria (2021). The disconcerting potential of online disinformation: Persuasive effects of astroturfing comments and three strategies for inoculation against them. New Media & Society, 23(5), 1080–1098. https://doi.org/10.1177/1461444820908530

Zhu, Hao (2024). kableExtra: Construct complex table with ’kable’ and pipe syntax. https://CRAN.R-project.org/package=kableExtra

Zinbarg, Richard E.; Revelle, William; Yovel, Iftah; & Li, Wen (2005). Cronbach’s α, Revelle’s β, and Mcdonald’s ωH: Their relations with each other and two alternative conceptualizations of reliability. Psychometrika, 70(1), 123–133. https://doi.org/10.1007/s11336-003-0974-7