How does hatred spread in Italy? Mario Draghi's dimision case

Regular Article How does hatred spread in Italy? Mario Draghi’s dimision case Daria Mottareale-Calvanese * , Sergio Arce-García , Elias Said-Hung Universidad Internacional de La Rioja, Spain 1. Introduction Social media platforms such as X (formerly Twitter), TikTok, and Facebook have become significant venues for the dissemination of disinformation and hate speech (Sanchez et al., 2022; Barud & Olalekan, 2022). These platforms often disseminate rumors, stereotypes, and prejudices that target specific social groups. Although the current discourse on hate speech lacks consensus on its definition (Paz et al., 2020; Rossini, 2022), there seems to be a common point of agreement when establishing the features that characterize this type of expression, such as the public nature and the intention to harm an individual or a specific group for ethnic, gender, religious, origin, or political reasons (Sellars, 2016; Struck, 2019; Ziegele et al., 2018). As Magall´ on (2019) rightly points out, hate messages sometimes involve the production and dissemination of intentionally distorted information. This misleads the audience into positioning narratives based on factual inaccuracies about issues or events, as well as stereotypes and prejudices towards certain groups within public opinion (Wood & Porter, 2019). Political parties and leaders increasingly leverage social media to connect with potential supporters and voters. However, the spread of hateful messages on these platforms contributes to a growing disconnect between institutions and public opinion (Thelwall & Cugelman, 2017), ultimately weakening trust in democratic institutions, such as the press and political parties (Farkas & Schou, 2018). The political campaigns of Barack Obama (Chaves-Montero, 2017) and Donald Trump (P´ erez Curiel & Lim´ on-Naharro, 2019) exemplify this trend, where social media became fertile ground for disinformation and hate speech. On these platforms, users share and comment on content primarily to express their ideological positions on specific issues rather than to engage in meaningful dialogue. In this study, we analyze how certain groups of users can influence public opinion and foster polarization through coordinated actions. This occurs through the dissemination of hate messages by accounts organized in interactive networks, which amplify extreme narratives on the X platform, especially in political contexts in various countries. 2. Research context The role of social media platforms in creating and transmitting hate speech is increasingly significant, yet often concealed behind politically correct discourses and social desirability bias—our tendency to present ourselves positively while distorting our actual positions (Lanz, Thielmann, & Gerpott, 2022). Politicians frequently employ this phenomenon to attract users’ attention, inadvertently spreading rumors, stereotypes, and prejudices against specific groups (Poletto et al., 2021). Kalsnes and Ihlebæk (2021), Ghasiya and Sasahara (2022), Roslan et al. (2022), and Müller and Schwarz (2023) confirm that platforms like X and Facebook have become breeding grounds for hate speech and anti-minority sentiments despite moderation efforts, primarily due to inconsistent regulatory frameworks and platform-specific strategies (Stockmann et al., 2023; Wang & Kim, 2023), driving many users toward alternative platforms like Bluesky, which experienced substantial growth after the 2024 U.S. presidential election (Rogers, 2024). In the domain of political communication, distinguishing between hate speech and negative political campaigning is important. While both employ emotional and polarizing strategies, their underlying principles and objectives are distinct from each other. Hate speech is specifically directed against groups identified by attributes such as ethnic origin, religion, gender, or sexual orientation, leading to stigmatization, exclusion, and even incitement to violence (Ruttloff et al., 2024, pp. 352–369). Its aim is not to engage in discourse on political proposals but to delegitimize the social belonging of certain groups, thereby contravening the fundamental democratic principles of equality and pluralism. Negative political campaigning, while legitimate, remains contentious in electoral competition. It primarily involves the critique of political opponents, parties, or programs, aiming to undermine their credibility or question their governing capabilities (Sch¨ afer et al., 2024). Although it may employ emotional appeals such as fear or indignation, its focus is on the political adversary rather than a social community This article is part of a special issue entitled: AI in political communication published in Social Sciences & Humanities Open. * Corresponding author. Avenida de la Paz 137, 26006 Logro˜ no (La Rioja), Spain. E-mail addresses: [email protected] (D. Mottareale-Calvanese), [email protected] (S. Arce-García), [email protected] (E. Said-Hung). Contents lists available at ScienceDirect Social Sciences & Humanities Open journal homepage: www.sciencedirect.com/journal/social-sciences-and-humanities-open https://doi.org/10.1016/j.ssaho.2025.102148 Received 23 May 2025; Received in revised form 11 September 2025; Accepted 24 October 2025 Social Sciences & Humanities Open 12 (2025) 102148 Available online 6 November 2025 2590-2911/© 2025 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/bync-nd/4.0/ ). defined by sensitive identities. Nonetheless, this form of messaging often becomes entangled with other expressions of hate, thereby blurring distinctions and ultimately converging the two phenomena. Abuín-Vences et al. (2022) state that hate speech on social media platforms causes desensitization, increases prejudice, and fosters social exclusion of the subjects of hate, even triggering violent acts. This phenomenon, which seems to be on the rise due to the proliferation of social media platforms, has led to growth in research in this field, providing new perspectives to understand and address this issue. This is especially true if we take into account the effects of the echo chamber and the bubble filter that occurs in users of current digital environments, which end up encouraging the constant transmission and repetition of messages, persistent censorship of non-majority positions, and information saturation (Arce-García et al., 2023; Guess et al., 2018). Despite academic advances, the need to better understand the capacity that users have in current digital scenarios to recognize and carry out strategies aimed at intensifying political and social polarization from these types of contexts is still present (Samy-Tayie et al., 2023). The nature and behavior of online haters remain largely unknown, particularly regarding their potential coordination in spreading “hate on the Internet,” “trolling,” or “hating” (Jędryczka et al., 2022). Recent research by Martínez Torralba et al. (2023) highlights the use of AI-designed propaganda techniques and microsegmentation to influence user attitudes and behaviors on key topics. These activities can be carried out by bots (automated computer algorithms), troll farms (groups of humans using fake accounts anonymously), and “Elite bodies” (actors serving specific political or economic interests). Astroturfing is a notable strategy that uses social media platforms to spread carefully crafted information, creating the illusion of grassroots expression while covertly influencing public opinion (Elmas et al., 2021). This approach employs microand nano-influencers to exploit loose social connections—aligning with Granovetter’s (1973) “weak ties” theory—to build extensive networks around targeted messages containing hate speech, negative emotions, and disinformation (Arce-García et al., 2023; P´ erez, 2020). These tactics intensify cyberspace polarization by leveraging the “selective retention” mechanism of the selective exposure theory (Turk, 2021), wherein users preferentially remember negative information about opponents (García-Orosa, 2021; T´ oth et al., 2023). The polarization arising from these discursive practices necessitates distinguishing between ideological polarization, which is rooted in programmatic and doctrinal differences, and affective polarization, which is characterized by hostility and the delegitimization of opponents. As noted by Díez et al. (2023), the interaction between these dimensions demonstrates how digital environments not only amplify the divergence between political positions but also exacerbate negative emotions towards those who hold them. The resulting orchestrated expressions, grounded in sociological foundations and leveraging both ideology and emotions to polarize political discourse, constitute what is now referred to as “Cognitive Warfare.” This concept aims to counter disinformation and develop tools to combat manipulation within digital communication channels (Casero-Ripoll´ es et al., 2023; Stato Maggiore della Difesa, 2023). It is essential to address this problem through strategies that promote informed and respectful public debate, encourage civic participation, and strengthen mechanisms to combat the spread of hate speech online. Effective regulation of hate speech on social media platforms and educational and awareness-raising initiatives can help mitigate its negative impact on public opinion and democracy in the long run. To this end, this paper focused, for the general objective set, on estimating the influence of hater accounts in shaping the debate on social media platforms such as X by promoting a polarized scenario of public opinion. The period of Mario Draghi’s resignation and the call for elections that preceded the victory of Giorgia Meloni and her far-right ideologically oriented Brothers of Italy Party was taken as a case study. This case study examines how political actors leverage social media platforms to advance their political-electoral interests, as Cloudy et al. (2023) documented. These strategies primarily involve promoting emotionally charged content with nationalist and populist themes designed to provoke fear and anger toward specific social groups (Tu˜ n´ on-Navarro & Bouzas-Blanco, 2023). This approach aims to attract, mobilize, and generate significant interactions with potential voters through viral messaging in digital spaces. Research by Berti and Loner (2023) on Matteo Salvini, leader of the far-right “Lega” party, demonstrates how these tactics undermine opponents’ credibility and influence public perception. Such strategies threaten national identity by disseminating difficult-to-verify content that heightens perceptions of insecurity and promotes divisive “us versus them” narratives, ultimately intensifying polarization in Italian society (Cervi, 2023; Denti & Faggian, 2021). This study examined the Italian case, considering that in 2021, the President of the Italian Republic, Sergio Mattarella, appointed Mario Draghi, a technocrat with extensive experience in international finance, as President of the government. Draghi, a former president of the Bank of Italy and the European Central Bank (ECB), officially accepted the task of forming a new cabinet in February 2021. The new president’s direction seemed to lean, on the one hand, towards technical experts in charge of the most strategic areas of the government and, on the other hand, towards the majority of parties eager to be part of a government that would be responsible for distributing resources on a large scale in the country’s economy (Garzia & Karremans, 2021). Despite favorable public criticism of the new technical government, on July 14, 2022, Draghi announced his resignation after Giuseppe Conti’s party, “Movimento 5 Stelle” (M5S) or 5 Star Movement in English, an important part of the government coalition, would not support a confidence motion in the Senate (Sica, 2021). At the European level, the most relevant aspect of the fall of the Draghi government was that the alignment with the European Union and the Atlantic Alliance was compromised (Charte, 2022), reflecting the division in Italian public opinion and the return to power of far-right populist figures such as Salvini and Berlusconi, the latter being the leader of the right-wing neoliberal party “Forza Italia” (FI) (Occhetta, 2022; Zulianello, 2020). In this context, and taking advantage of social media platforms, populist political movements began to use social media platforms to capture the attention and obtain the consent of users (Martella & Bracciale, 2022). Valenti (2023) observed that hate speech has become normalized within Italian political communication on social media platforms such as X, where it is frequently employed as a populist instrument. This form of hatred is expressed through a combination of verbal and nonverbal elements, including images and texts, which propagate stereotypical and discriminatory narratives. In this context, Ieracitano, Balenzano, Girardi, Gemmano, and Comunello (2023) elucidate that moral disengagement associated with hate speech on digital platforms like X is linked to conformity with established rules and authority, which partially accounts for the acceptance and perpetuation of these messages in digital environments. Consequently, political discourse on these platforms strategically utilizes multimodal and implicit linguistic structures to subtly yet effectively convey discriminatory content (Alonso, 2015). In the Italian context, political discourse on social media has increasingly manifested as hate speech directed at minorities, immigrants, and political adversaries (Pasta, 2019; Piangerelli, 2020). The Italian political system is characterized by pronounced instability, marked by frequent governmental changes, transient coalitions, and partisan fragmentation, complicating the formation of stable majorities. Consequently, the executive branch and its leaders have assumed a central role in decision-making processes (Fittipaldi & Musella, 2022). This institutional volatility, combined with escalating polarization, influences the strategic use of social media to mobilize support bases, disseminate emotionally charged messages and construct “us versus them” narratives. In this manner, platforms such as X exacerbate existing political tensions and reinforce extreme ideological positions, D. Mottareale-Calvanese et al. Social Sciences & Humanities Open 12 (2025) 102148 2 thereby posing significant challenges to social cohesion and the quality of democracy in Italy (Musella, 2024). During Mario Draghi’s government, digital communication channels were strategically used to generate opposition to his administration. This aggressive approach fostered social and political polarization in Italy (Capano & Sandri, 2022; Ieracitano & Centola, 2025), creating conditions favorable for Giorgia Meloni’s rise to power with her far-right party, Brothers of Italy (FdI), which represented the opposition during Draghi’s tenure (Imberti, 2021). Meloni’s electoral success may have benefited from the viralization of hateful content by coordinated accounts on platforms like X, which deepened political polarization. This strategy, combined with support from Salvini and Berlusconi and their respective parties, helped Meloni become Prime Minister, marking a rightward shift in Italian politics (Merlo, 2022). Based on the analysis conducted thus far, our working hypothesis posits that the political polarization observed during the examined period on platforms such as X was influenced by accounts that operated in a coordinated manner. These accounts may have employed strategies such as astroturfing to shape Italy’s political discourse. The hypothesis is based on what has been stated by authors considered in this work, such as P´ erez (2020) or Arce-García et al. (2023), who point out that the dissemination of messages with hate speech is not spontaneous or random, but that there is a previously designed strategy behind it, carried out by users who try to go unnoticed, to achieve the positioning of narratives that legitimize their ideological positions, in the case of Italy, in favor of the political project led by Giorgia Meloni and her political party, within the public opinion of this country. 3. Methodology This study aims to evaluate how hate accounts on platforms like X shaped public debate and promoted political polarization during Mario Draghi’s resignation and the subsequent election campaign that led to Giorgia Meloni’s victory with the Brothers of Italy (July–August 2022). Building on the context of Italian political polarization outlined in the introduction, we examine how this environment facilitated the rise of parties such as Meloni’s. Our analysis focuses on how these political actors strategically used social media platforms to gain significant representation in Italy’s political landscape, employing the tactics discussed earlier in this study. 3.1. Objetives of the study To achieve the general objective of the proposed work, the following is proposed: •SO1: Determine the form of grouping of hater accounts. •SO2: To estimate the hate load and intensity present around the case study. •SO3: Establish the general characteristics of hateful accounts. To achieve the proposed general objective, all messages were collected from December 7, 2022, to August 16, 2022, and associated with the hashtags #Draghi, #crisidigoverno, #crisigoverno, and #DraghiOut. This period saw the announcement of Mario Draghi’s resignation to the President of Italy, Sergio Mattarella (BBC News Mundo, 2022); the acceptance of the resignation and call for early elections by the Italian Parliament (HuffPost, 2022); and the debate generated as a result of these events, which led to the election results of September 2022. 3.2. Data collection Messages were collected using the RTweet library (Kearney, 2019) in R software, connecting to X’s API version 1.1, with both hashtag selection and geolocation parameters targeting countries identified by Bradshaw et al. (2021) as disinformation sources and European nations. While this method gathered a substantial sample of Italian-language messages containing specified hashtags (796,654 tweets total), it could not guarantee complete coverage. API 1.1 only permitted access to 7-day data windows, and collection occurred weekly with duplicates removed. Bradshaw classified countries by online campaigning capacity (17 high-capacity including China and Russia, 37 medium-capacity, and 27 low-capacity including Italy), with the geolocation technique offering 77.84 % worldwide reliability and 88.15 % in Europe (average error: 256 km) (Lopreite et al., 2021; Van-der-Veen et al., 2015). Notably, the collection yielded unexpected volumes from Thailand (22,085 messages) and Albania (17,525 messages). The sample size provided sufficient diversity to address the study’s general objective; however, it is essential to note that some hate speech messages may have been removed because of X’s content moderation policies (X Help Center, 2023). Research indicates that X’s moderation is limited in scope, with over 80 % of problematic messages remaining on the platform (CBS News Mundo, 2022; Center for Countering Digital Hate, 2023). From all the procedures applied in this work, it was possible to identify the groups that participated in the social media platform X around Draghi’s resignation and to see if there was hate speech, its main themes, and the characteristics that were observed in the speeches to detect the possible presence of astroturfing to condition the decisions and perceptions of users, as well as to dominate discourses and electoral processes on the Internet. 3.3. Data processing techniques To understand the network structures, identify the groups participating in the discussion, and study the hate speech uttered to help achieve the general objective proposed in this study, the following was conducted: •To meet Specific Objective 2 (SO2), this study analyzed hate expressions in the entire sample using natural language processing with the Syuzhet algorithm in R (Jockers, 2017). This technique identifies and quantifies hate speech using pre-established lexicons. The study used Hurtlex in Italian (2018 version), which was developed in the original language to preserve local meanings and nuances without translation issues (Bassignana et al., 2018). Hurtlex is a lexicon initially developed in Italian that encompasses offensive, aggressive, and hateful vocabulary. It was created by researchers at the University of Turin, building on a lexicon formulated by Italian linguists (Bassignana et al., 2018). The application of this lexicon in research involving the Syuzhet algorithm facilitates the determination of the intensity level of hate based on the occurrence of specific words, their combined usage, and the presence of other contextual words that may amplify or diminish their intensity. To prevent the misidentification of everyday, ironic, or ambiguous expressions as hateful, only messages exhibiting an intensity from the upper third quartile were examined. This approach ensured that only messages clearly propagating hate were considered. Although the original lexicon comprises various semantic categories of hate, for this study, they were consolidated into a single general category. After identification, the researchers established a hate average index by dividing the hate value by the total word count in each message. This word-count detection approach, previously employed in other hate detection algorithms (Pereira-Kohatsu, 2019), enables a fair comparison between messages of varying lengths, which is particularly important for social media content classification. D. Mottareale-Calvanese et al. Social Sciences & Humanities Open 12 (2025) 102148 3 •To comply with SO3, we conducted a network analysis using graph theory to examine user connections through retweeted messages (Barabasi, 2016). Data were processed in R and analyzed using Gephi 10.1. The ForceAtlas2 algorithm (Jacomy et al., 2014) provided a graphical representation, whereas the Louvain algorithm (Blondel et al., 2008) grouped users into clusters (Chen et al., 2020). Each cluster was characterized by examining the following key metrics: modularity, average edge length, average degree, lattice diameter, eigenvector centrality, betweenness centrality, and number of edges. •For SO3, a text mining process was conducted, eliminating words or stopwords that did not contribute sense or meaning to the message (e.g., articles and adverbs). The identified hate messages were graphically represented in two dimensions using a multidimensional scaling (MDS) plot. The words are represented by colors that allow for the differentiation of the different clusters, which are most frequently linked to one another. Clusters were identified using a kmeans algorithm with k =7 to ensure that a sufficient number of topics were identified. This technique was used to determine the primary issues of hate speech within each cluster identified in the initial network theory analysis. The R data file used for collecting, processing, and analyzing the messages is available in the Supplementary Material (see Supplementary Material XXXXXX for the R data file). All statistical analyses and figure creation were performed using R software, which enables the analysis of large datasets through programming. 3.4. Practical difficulties in the methodology During the implementation of the methodological design, several difficulties were identified that affected both the collection and analysis of data: •Limitations of the X API (version 1.1): The limitation of data access to a seven-day window necessitated weekly sampling, thereby increasing the risk of temporal overlap and the potential loss of pertinent tweets during download interruptions. Additionally, the inability to retroactively access messages removed due to moderation policies constrained the comprehensiveness of the sample, as up to 20 % of problematic messages may have been filtered before they could be recollected. •Geolocation accuracy: The geolocation technique based on location metadata demonstrated a reliability of 88.15 % within Europe, albeit with an average error margin of 256 km in the geolocation. This imprecision poses challenges for accurate territorial allocation, particularly in cross-border regions or for users with ambiguous geolocation data, which can potentially introduce noise into the analysis of the origin of fake accounts, such as those in Thailand and Albania. •Filtering non-explicit content: In applying the Hurtlex lexicon through the Syuzhet algorithm, it was necessary to establish a threshold within the upper third quartile of intensity to mitigate false positives arising from irony or colloquial expressions. While this approach effectively isolates overtly hateful speech, it may inadvertently exclude subtle or implicit expressions of hate, thereby limiting the capture of discursive nuances and potentially diminishing the representativeness of certain message types. •Network analysis with Gephi: Executing the ForceAtlas2 and Louvain algorithms on datasets comprising over half a million edges resulted in prolonged processing durations and occasional instability in the configuration of the modularity parameters. The adjustment of these parameters necessitated multiple manual iterations, thereby significantly extending the analysis duration and potentially introducing a bias in the detection of smaller cell clusters. •Integration of qualitative methodologies: Although this study does not primarily focus on qualitative methods, the absence of manual content analysis or user interviews precludes the ability to compare the findings of this research with the direct perceptions of participants or moderators of X. Consequently, this limitation constrains the understanding of the motivations underlying the observed coordination and the validation of automatically identified astroturfing strategies. 4. Results 4.1. Clustering of hate accounts Of the 796,654 tweets written in Italian containing the selected hashtags, 22,085 came from Thailand and 17,525 from Albania. Globally, the messages consisted of 161,613 original tweets and 634,951 retweets (RT), whereas the messages collected from Thailand and Albania did not include any retweets. The remaining countries were discarded because they did not reach 1000 messages. The temporal distribution of messages was primarily concentrated on July 13, 2022, and the week that followed, with the highest volume of tweets occurring between July 20 and 21, dropping drastically from July 23. These dates coincide not with the initial announcement but with the confirmation of the resignation in the Italian Senate. The sequence of messages was similar both globally and in the two countries. However, in Thailand, the highest activity was recorded on the 21st, whereas in the other cases, it was on the 20th. The network study shown in Fig. 1 shows highly interconnected groups with no clear separation between groups. The structure presented a modularity of 0.588, with 558 communities or clusters detected. Among them, six main groups were detected, representing 97.33 % of the total retweet traffic around the analyzed hashtags. Their political leanings are described by examining the main accounts in each cluster by eigenvector: •MEL Group (in green in Fig. 1): Revolves around followers of Giorgia Meloni and Matteo Salvini, with 22.19 % of traffic. •M5S Group (orange): Revolves around the 5 Star Movement (M5S), as well as anti-vaccine and conspiracy groups, with 11.41 %. •CD Group (pink): Italian center-right political tendency (23.48 %). •US Group (blue): Mainly ultra-left (17.4 %). •DC Group (brown): group of Christian Democrat politicians and journalists (16.87 %). •SIN Group (pink): Italian trade unions and the media (5.98 %). The remaining clusters were discarded owing to their low representativeness for the rest of the analysis. Table 1 analysis reveals that MEL, M5S, and CD clusters exhibit significantly higher average degrees between nodes, larger group diameters, and greater distances compared to US, DC, and SIN clusters, with self-descriptions aligning with the algorithm’s classifications through terminology frequently associated with right-wing ideology. These first three groups demonstrated more active accounts and extensive participant networks, characterized by very low modularity (minimal egocentrism), resulting in high interconnectivity among accounts despite apparent ideological differences between the accounts. Such characteristics reflect political strategies designed to maximize outreach through cross-cutting narratives such as nationalism or Euroscepticism, where users with clear ideological orientations dominated nearly all debates on the X platform during the study period. By fostering dynamic interactions and presenting themselves as open, civil movements rather than elitist entities, these parties facilitate the dissemination of mass messages and rapid response capabilities in polarized contexts, thereby strengthening the cohesion of their digital campaigns. Text mining analysis allows us to see the most used words for selfdefinition by the users that make up each detected group: D. Mottareale-Calvanese et al. Social Sciences & Humanities Open 12 (2025) 102148 4 •MEL Group: always, only, be, freedom, free, truth, never, Italy, world, right, politics, do, God, every, man, Italian, no green pass (No Vaccination movement). •M5S Group: always, world, just, politics, being, Italy, left, love, time, stars, communist, anti-fascist, never, war, music, do, people. •US Group: she/her, always, life, things, love, world, time, account, love, addicted, politics, like, write, fans. •DC Group: politics, journalist, antifascist, life, always, world, time, years, being, communication, love, passionate, Rome, never, you, only. •CD Group: always, politics, pro-European, italiaviva, I love, antifascist, Italy, only, never journalist, liberal, world, lover, freedom, being, history, art, every, engineer. •SIN Group: news, politics, news, social, journalist, and zazoom (Social Media News’ account, which has been very active on Twitter for many years.), Italy, web, so, time, site, world, all, account. 4.2. Presence of hate and intensity Once the groups were determined, messages containing hate were identified, totaling 136,887 messages, which represented 17.18 % of all tweets in the dataset. Among the messages from abroad, 5997 hate messages (1052 accounts) were detected in Thailand and 4553 (814 Fig. 1. Network map of the groups. Source. Own elaboration using Gephi 10.1. Legend. MEL: Group centered on followers of Giorgia Meloni and Matteo Salvini; M5S: Group linked to the Five Star Movement; CD: Italian center-right political tendency; US: Mainly ultra-left; DC: Group around Christian Democrat politicians and journalists; SIN: Italian trade unions and media. Table 1 Nature of the clusters detected. MEL (16) M5S (4) US (21) DC (48) CD (15) SIN (39) Mean grade 5,36 6,99 1,31 1,58 5,60 1,35 Diameter 14 13 6 10 21 5 Mean distance 5,32 4,25 2,43 3,11 6,55 1,68 Nodes (beads) 14.785 (22,19 %) 7.600 (11,41 %) 11.590 (17,4 %) 11.241 (16,87 %) 15.645 (23,48 %) 3.986 (5,98 %) Edges (RT connections) 79.219 (24,5 %) 53.127 (16,43 %) 15.204 (4,7 %) 17.753 (5,49 %) 87.634 (27,1 %) 5.372 (1,66 %) Modularity 0,28 0,33 0,72 0,65 0,31 0,77 Source. Own elaboration using R software. Note. The number of the clusters detected is shown in brackets. Table 2 Number of hate messages detected per cluster. MEL (16) M5S (4) US (21) DC (48) CD (15) SIN (39) Total 36.669 30.663 4.721 13.525 47.501 3.808 Albania 134 125 72 136 276 71 Thailand 171 184 101 173 344 173 Source. Own elaboration using R software. Note. The number of the clusters detected is shown in brackets. D. Mottareale-Calvanese et al. Social Sciences & Humanities Open 12 (2025) 102148 5 accounts) in Albania. The distribution by clusters is presented in Table 2, highlighting that the highest number of hate messages was concentrated in the MEL, M5S, and CD clusters. Hate messages from Albania and Thailand represent 3.3 % and 4.4 %, respectively, of the total, which, although not a majority, is significant since almost 8 % of hate messages in Italian come from these two non-Italian-speaking countries. The intensity of hate speech varied significantly among the different groups analyzed. The MEL (0.467), CD (0.355), and M5S (0.324) groups showed the highest average hate load per word in their messages, followed by DC (0.306), US (0.297), and SIN (0.167). However, the average values were very similar and low in all groups, ranging from 0.0311 to 0.0344. Fig. 2. Multidimensional representation of the messages of each cluster. Source. Own elaboration, using R software. D. Mottareale-Calvanese et al. Social Sciences & Humanities Open 12 (2025) 102148 6 For the messages from Thailand and Albania, their maximum (0.194 and 0.192, respectively) and mean values (0.0314 and 0.0316, respectively) were within the overall average, with no notable extremes compared to the overall maximum of 0.467 and the mean of 0.0316. Notably, 81.34 % of the retweets (RT) are on messages classified as hate speech. These messages had a median of 69 forwards and a mean of 234.4, compared to 47 and 194.7, respectively, for all messages. The text mining analysis, depicted in Fig. 2, reveals the predominant themes in the hate speech of each group: •MEL and M5S: They associated Draghi’s government crisis with the people, introducing themes such as the green pass and appeals to the country’s history. •DC: Focuses on the crisis of Salvini and Conte. •CD: Relates the crisis to Berlusconi, Salvini and Conte. •US: Criticizes Conte and Salvini, using derogatory language towards politicians. •DC: Accuses Salvini of a lack of responsibility and contrasts democracy with the “obscurantism” of Meloni and Salvini. •CD: Criticizes Meloni and Salvini for their alleged pro-Putin positions and lack of Atlanticism. •SIN: Focuses on historical responsibility after resignation and calls for elections. Table 2, therefore, does not show insults among the words most frequently used in hate speech, but it does establish the themes and objectives, which are mainly focused on emotional discourse (making history) and attacks on well-known political figures. Table 3 illustrates, by way of example, how hatred is primarily concentrated on the topics indicated: Draghi’s resignation, other political actors such as Salvini and Meloni, and ancillary issues including anti-vaccine and Putin-related tweets. Hatred is mainly manifested through accusations of incapacity, criminality, and dishonesty. It was observed that some insults contained double letters to avoid being easily identified by algorithms. A clear level of hate is therefore detected, although not of great intensity and without crossing legal boundaries, as other research has highlighted (Arce-García et al, 2023). 4.3. Characteristics of the accounts disseminating hate An analysis of the accounts participating in the debate of hate messages, according to the clusters identified, shows (Table 4) that these are users with a nanoinfluencer-type profile, typical of prepared Astroturfing-type campaigns. Several aspects indicate this trend: a small number of followers and accounts followed, low proximity to the leaders or influencers of each cluster, almost null importance in the network (eigenvector), and nearly non-existent intermediation between users. However, differences were observed between the clusters. The MEL and M5S clusters showed very high activity in terms of message emission Table 3 Messages with higher hate load detected by cluster. MEL (16) ● The reading of the #crisisofgovernment of the declared and non-declared PD voters is as much as the most idiotic subnormal bebetoid idiot dishonest disloyal insincere hypocrite and fake I have ever seen. ● #inondala7 this bad people until yesterday with the MAFIOUS LEADER of COSA NOSTRA #Berlusconi now stay with the CenterLeft!!! EVERYTHING in order to REMAIN HANGED TO THE ARMCHAIRS AND KEEP THEIR FATHER DELINQUENT CRIMINAL FARABUTTO AGUZZINO PERSECUTOR OF THE ITALIANS #Draghi ● #coffeebreakla7 #lariachetira #inondala7 Citizens understand the game of these MISERABLES? Now they pretend Divisions … but then they replay to you again the #Duce DELINQUENT FARABUTTO AGUZZINO CRIMINAL OF THE ITALIANS #Draghi to Head of Government (“Europe asks us”) M5S (4) ● 30 years of Governments of whoremongers, mafiosi, midgets and dancers have brought us shit up to our necks. Now they all try to save themselves by blaming #M5S and #Conte for every nefariousness taking advantage of the usual Ox People with goldfish memories. We deserve it. #PD #Draghi #Renzi ● 37 miserable despised Di Maio, 79 dead dragon Draghi, 43 village idiot Renzi. Terno on Rome and all. #Draghi #buffoon #disappear ● only the #PD de mmerda stays in the courtroom to support this miserable criminal government. Laughing. #Letta puppet fucked up across the board. Idiot. Go to bed peacefully. #marathonmentana #crisisgovernment #draghi #DraghiGet OUT NOW US (21) ● “The parties that supported Draghi to the end showed sense of state and responsibility” this time I half agree! In democracy the sense of state is also of those who and in disagreement with #Draghi Maybe they are stupid, dumb, ignorant, but not irresponsible ● But please. Conte never did anything wrong to ally himself with Pd and 60/70 in a government led by a person who considered himself incapable, ignorant and conceited. #crisisgovernment #CalendAI #Calenda #15July #unaFollia #instagramdown #marathonmentana ● a fool is talking about #fdi and #draghi is looking at him like he’s a moron … actually this #fdi one is DC (48) ● Politics, in Italy, is in a state that to call pitiful is an understatement, but the responsibility also lies with an indolent, boorish and ignorant electorate. Mario Draghi you do not deserve. #crisidigoverno ● #crisidigoverno anyone who brings down a government in the current international context is simply an irresponsible moronic imbecile useless good for nothing ! Thank you on behalf of all citizens. ● That this was the most boorish and clumsy Parliament in republican history was already known. That it was also the dumbest was confirmed yesterday. Italy hostage to incompetent wafflers, this is. #crisisgovernment CD (15) ● I’m told that ignorant mentecatto Gianluigi #Paragone gave the usual braying ignorant buffoon show @gparagone the one who blocks those who criticize him politely #crisidigoverno ● @InOndaLa7 @corradoformigli the usual useless asshole. A moron who gave fuel to the fire of populism. Now he would like to know about #Draghi’s private. You moron, what do you care? ● Politics, in Italy, is in a state that to call pitiful is an understatement, but the responsibility also lies with an indolent, boorish and ignorant electorate. Mario Draghi you do not deserve. #crisidigoverno SIN (39) ● Regardless of who it will be and when they will be held, the next elections will have a winner. This one would be a mad, crazy, dastardly cretin if he decides to dispense with the authority, competence and seriousness of Mario #Draghi. Period. #pressconference ● @jacopo_iacoboni I think Draghi is not in it, I see it as bad. On Facebook, our newspaper’s colleague F.M. writes, “I hope this #crisidigoverno is not heterodirected by foreign forces.” There, one really has to think so. Either he is a moron or a scoundrel, I don’t know ● The best of Draghi’s qualities? Not being a politician. But we are a guiltily ignorant people and have a disaffected political class composed of embarrassing individuals. And now? The beginning of a horror movie #crisidiggovernment #Draghi #MarioDraghi #Salvini #Meloni #m5s #PD Source. Own elaboration. Note. The number of the clusters detected is shown in brackets. D. Mottareale-Calvanese et al. Social Sciences & Humanities Open 12 (2025) 102148 7 and bookmarking, significantly surpassing the other clusters. The number of daily messages and favorites for these accounts was three or four times higher than that of the different clusters during their entire existence. The MEL, M5S, and CD clusters show high connection degrees and support other accounts with minimal intermediation, suggesting that these accounts continuously follow main accounts and actively retweet content. Notably, accounts registered in Albania and Thailand exclusively sent direct messages without retweeting, indicating a potentially strategic approach that avoids public amplification of messages. This distinctive behavior pattern suggests either a different communication strategy or accounts being used for specific purposes. This study identifies a highly concentrated communicative dynamic within six macro-communities, which collectively account for 97 % of the network traffic. This dynamic is predominantly characterized by the MEL, M5S, and CD clusters, which are noted for their high interconnectivity, low modularity, and employment of transversal narratives associated with the political right. The temporal pattern of digital activity corresponds to parliamentary milestones, underscoring the interdependence between political events and discursive production. The presence of hate messages, constituting 17 % of the total, is notable for its significant viral potential, as evidenced by an 81 % retweet rate, despite its moderate vocabulary and absence of explicit insults. The involvement of accounts from Thailand and Albania, characterized by a lack of retweets and direct broadcasting patterns, suggests astroturfing practices aimed at strategically influencing Italian discourse. Profile analysis indicates a predominance of nano-influencers characterized by low network centrality but exhibiting elevated levels of hyperactivity. This suggests a dissemination strategy reliant on swarms of micro-accounts rather than on individual leadership. The findings depict a polarized digital ecosystem in which hatred serves as a vector for community cohesion and a mechanism for political intensification. 5. Discussion The analysis of 796,654 Italian-language messages concerning Mario Draghi’s resignation identified coordination patterns that substantiate the initial hypothesis regarding astroturfing strategies in Italian political discourse. The six principal groups identified (MEL, M5S, CD, US, DC, and SIN) constituted 97.33 % of the total retweet activity, with a modularity score of 0.588, indicating significant interconnectedness among communities that appear to be ideologically diverse. This configuration implies strategic coordination rather than organic polarization, particularly given that the MEL, M5S, and CD groups exhibited distinct characteristics of low modularity (minimal egocentrism) and elevated message dissemination levels. The identification of 136,887 hate speech messages, constituting 17.18 % of the total, predominantly concentrated within these three groups, alongside the observation that 81.34 % of retweets pertained to messages classified as hate speech, underscores the coordinated nature of these campaigns. The intensity of hate speech varied significantly among the groups, with MEL (0.467), CD (0.355), and M5S (0.324) demonstrating the highest values. This pattern aligns with the populist strategies documented by Martella and Bracciale (2022), who used emotions as instruments to garner consent from influential users. Hate narratives were operationalized using the Syuzhet algorithm in conjunction with the Italian Hurtlex lexicon (2018 version), which classifies expressions into 17 categories of potentially harmful content. To reduce the incidence of false positives resulting from ironic or colloquial expressions, a filter based on the upper third quartile of intensity (≥0.324) was employed, ensuring that only expressions with a substantial emotional charge were identified as hate speech. The hateaverage index, determined by dividing the hate value by the total word count in each message, facilitates equitable comparisons between messages of varying lengths —a critical consideration for content analysis on social media. This methodology demonstrates that hate speech is not predominantly expressed through explicit insults. Instead, it is conveyed through highly emotional and personalized discourse, which, as noted by Abuín-Vences et al. (2022), fosters a polarized and negatively charged perspective. The primary themes centered on the personal attributes of politicians, emotional appeals to national responsibility, and associations with foreign leaders such as Putin—strategies that are characteristic of far-right populism, as documented in the literature. The findings confirm Granovetter’s (1973) “weak ties” theory, as identified nano-influencer accounts (with few followers and positioned on the network periphery) proved effective in introducing polarizing content. These accounts, with characteristics typical of prepared Astroturfing campaigns, leveraged loose social connections to build extensive networks around targeted messages containing hate speech and conspiracy narratives. Turk’s (2021) selective exposure theory is evident in the data, particularly through the mechanism of “selective retention,” wherein users preferentially recall negative information about opponents. This characteristic is displayed by the MEL, M5S, and CD groups, as evidenced by their retweets, which predominantly focus on hateful messages that encourage repetition, censor minority positions, and create information saturation. These findings align with Bobba and Roncarolo’s (2018) observations regarding heightened engagement with populist messages. The discovery of 22,085 messages originating from Thailand and 17,525 from Albania, constituting nearly 8 % of the identified hate Table 4 Nature of participating accounts. MEL (16) M5S (4) US(21) DC (48) CD (15) SIN(39) Median Media Median Media Median Media Median Media Median Media Median Media Followers 567 2051 653 2142 318 2667 233 2767 530 2316 736 36523 Friends 617 1468 726 1456 559 1507 460 1132 720 1462 671,5 1400 Broadcasted messages history 24667 66022 29867 69688 8789 34136 8056 35171 17581 52572 20786 93027 Favorites 27298 65650 33960 68746 13322 48576 13035 42028 25896 69991 1407 7494 Account creation 17/07/ 2016 15/09/ 2016 31/03/ 2014 21/06/ 2015 08/04/ 2014 18/04/ 2015 10/09/ 2013 13/12/ 2014 25/02/ 2014 16/04/ 2015 15/08/ 2013 07/09/ 2014 Degree 23 54,89 50 90,74 3 16,13 6 23,87 31 79,33 7 37,98 Closeness 0,208 0,506 0,227 0,558 0,170 0,727 0,156 0,964 0,199 0,284 0,174 2189 Betweenness 0 272003 0 354064 0 68490 0 46078 0 315525 0 52557 Eigen 0 0,005 0 0,097 0 0,002 0 0,002 0 0,007 0 0,006 Posts/day 13,26 36,68 11,07 30,94 3,27 14,74 2,78 14,40 6,43 22,68 7,10 36,64 Favorites/day 14,68 36,47 12,58 30,53 4,95 20,97 4,49 17,21 9,48 30,19 0,48 2,95 Source. Own elaboration. D. Mottareale-Calvanese et al. Social Sciences & Humanities Open 12 (2025) 102148 8 speech, indicates a transnational aspect to this phenomenon. These accounts exhibited a unique pattern: they exclusively transmitted direct messages without engaging in retweets, suggesting a potentially coordinated strategy aimed at avoiding public amplification. This behavior implies that accounts are utilized for specific objectives, possibly linked to foreign influence operations documented by Bradshaw et al. (2021) in their comprehensive global inventory of organized media manipulation. This research advances the theoretical framework of Astroturfing by elucidating how these strategies function within specific political contexts, thereby creating “cognitive warfare” (Stato Maggiore della Difesa, 2023). The data indicate that Astroturfing does not invariably manipulate public opinion directly; rather, its intent, coupled with individuals’ selective biases, influences the selection and consumption of news, ultimately shaping collective perceptions. The interaction between emotional content, information dissemination, and opinion formation highlights the complexity of the information ecosystem on social media and its potential to influence public perception. These manipulative efforts can subtly affect the selection of users to follow and how these choices influence their worldviews, thereby playing a significant role in shaping the media landscape and, consequently, in shaping opinions and social attitudes. In addition to confirming the specific objectives, these findings extend the weak-tie theory framework by demonstrating that peripheral nano-influencers can exert a significant influence through coordinated retweets. This suggests the need to revise information diffusion models that currently focus solely on central actors. Moreover, the high virality of messages classified as hate speech, despite their low lexical intensity, underscores the effectiveness of subtle emotional strategies and the inadequacy of existing moderation systems. This has important implications for both regulators and digital platforms. This phenomenon perpetuates a cycle of selective reinforcement and information saturation, which, according to selective exposure theory, exacerbates affective polarization by prioritizing harmful content over rational deliberation. 5.1. Limitations of the study and future lines of research Several methodological limitations constrained this study’s scope. First, the analysis period was confined to July–August 2022, which precludes the assessment of Astroturfing dynamics beyond an immediate electoral context. Second, reliance on the X API 1.1 necessitated data retrieval windows of only seven days, requiring weekly sampling that could result in temporal overlaps and the omission of pertinent tweets between download intervals. Additionally, up to 20 % of potentially hateful messages may have been automatically deleted before collection because of the platform’s moderation policies, limiting the comprehensiveness of the sample. Lastly, the geolocation technique based on location metadata provided a reliability of 88.15 % in Europe, but with an average margin of error of 256 km, which introduced noise into the territorial attribution of accounts, particularly in border regions or with ambiguous metadata. To address these limitations and enhance the comprehension of astroturfing and polarization strategies, it is recommended that future research be conducted as follows: •Extending the temporal scope of the study beyond electoral periods to enable differentiation between discrete electoral campaigns and continuous influence operations. •Employ multimodal methodologies that integrate text analysis with automated image and video detection to identify manipulative elements that are not solely lexical but also visual or audiovisual. •Conduct comparative analyses across multiple European nations to discern transnational patterns of media manipulation and the influence of foreign entities on national democratic processes. •Integrate qualitative methodologies, such as conducting interviews with platform moderators or performing manual content analysis of representative cases, to complement and enhance quantitative findings and substantiate automatically identified coordination strategies. •Assess the implementation of updated API versions and alternative data sources, such as academic APIs or tweet archiving tools, to address the platform’s temporal and moderation constraints. 6. Conclusions This study empirically validates the hypothesis regarding the presence of coordinated political polarization strategies in X during the crisis of Draghi’s government. An analysis of 796,654 messages revealed that 17.18 % contained hate speech, predominantly concentrated within three groups (MEL: 22.19 %, M5S: 11.41 %, and CD: 23.48 %), demonstrating characteristics consistent with astroturfing tactics. The substantial proportion of retweets in messages classified as hate (81.34 %) and the coordination patterns identified through network analysis corroborate the strategic rather than organic nature of the observed polarization. The key theoretical contributions of this study encompass the empirical validation of Granovetter’s weak-tie theory within the context of digital manipulation, the elucidation of astroturfing mechanisms through subtle emotional narratives that circumvent algorithmic detection, and the identification of transnational dimensions in influence operations, notably involving significant participation from Albania and Thailand. Furthermore, this study advances the understanding of hate speech as a strategic tool that, despite its low intensity, maintains sustained effectiveness by evading content moderation. These findings underscore the necessity for enhanced regulatory frameworks that acknowledge the nuanced nature of strategic hate speech. Additionally, there is a need to develop detection tools that account for patterns of temporal and geographic coordination, as well as to implement media literacy programs aimed at increasing awareness of emotional manipulation techniques. For digital platforms, these results highlight the importance of monitoring not only explicit content but also the patterns of coordinated behavior and suspicious cross-border activities. This research elucidates how specific political events act as catalysts for the proliferation of hate speech, thereby contributing to social and political polarization through the promotion of far-right narratives. The data reveal a definitive correlation between negatively charged messages and political polarization, particularly when these messages target political figures and sensitive topics such as the Russia–Ukraine war or anti-vaccine movements. This dynamic potentially facilitates the resurgence of far-right populist leaders in positions of power. The present findings illustrate that Astroturfing initiatives, even those characterized by minimal lexical intensity, sustain their effectiveness by circumventing content moderation and coordinating on a transnational scale. The capacity of these campaigns to disseminate polarizing narratives highlights the critical need to enhance the detection of behavioral patterns and fortify public media literacy policies. CRediT authorship contribution statement Daria Mottareale-Calvanese: Writing – original draft, Investigation, Formal analysis, Conceptualization. Sergio Arce-García: Writing – original draft, Validation, Methodology, Data curation. Elias SaidHung: Writing – review & editing, Project administration, Investigation, Conceptualization. Ethical statement Ethical approval is not applicable to this manuscript. D. Mottareale-Calvanese et al. Social Sciences & Humanities Open 12 (2025) 102148 9