Unmasking coordinated hate: Analysing hate speech on Spanish digital news media

Abstract

This study examines the characteristics and behaviours of accounts that propagate hate speech through their responses to articles posted on five leading digital news media in Spain on Platform X (previously Twitter). Using non-experimental quantitative research, we analysed 1345 hate-expressing messages from 173,449 user comments on content shared in five leading digital news media during January 2021. Network analysis, the Homophilic Exposure Index (HEI), regression analysis and the k-means algorithm were used to identify features that characterize accounts that disseminate low-intensity hate expressions in a coordinated manner, undermining the moderation efforts of digital news media. As a result, digital news media must develop strategies to reduce the presence of this type of expression and confront accounts that operate covertly in a coordinated manner, using Astroturfing to manipulate debates around the content published on X.

Keywords

Digital media digital news media features hate speech haters profile social media Twitter users X

Introduction

The growing prominence of social media platforms such as X (previously Twitter), TikTok and Facebook has given rise to a hybrid communication model that has affected and limited the influence of traditional media channels. Communication strategies on these social media platforms tend to be impersonal, with users accessing and sharing content and opinions on various topics. However, the emphasis on emotions and personal convictions has resulted in a situation in which verifiable information is often relegated to a secondary role, leading to its progressive loss (Chadwick, 2017; Matamoros-Fernández and Farkas, 2021).

The implications of social media and the hybrid communication model that facilitated their growth were examined. Several studies have focused on comprehending the specific contexts, behavioural patterns, motivations and gratifications associated with the diverse methods of accessing and using news provided by news media in digital settings (Bright, 2016; Gottfried and Shearer, 2016; Swart et al., 2019). In addition, progress has been made in understanding the role of these contexts in detecting and identifying strategies for the dissemination of disinformation content, including expressions of hate that facilitate its viralization, to influence public opinion and electoral processes (Howard et al., 2017; Van-der-Linden et al., 2017; Zerback and Töpfl, 2021; Zhao et al., 2020). The effects of the echo chamber and bubble filter on social media users in digital settings have also been assessed, including constant message transmission and repetition, persistent censorship of non-majority positions and information saturation (Flaxman et al., 2016; Guess et al., 2018; Arce-García et al., 2023).

There has been a significant transformation in our societies’ methods of producing, distributing and consuming information content (Swart et al., 2019). This has significantly affected the relationship between news media, social media and audience members. These audiences have become increasingly reliant on social media platforms, which require news media to adapt their messages and formats to the various algorithms used by these platforms to reach more users (Karlsson et al., 2022; Meese and Hurcombe, 2020). As a result, news media have opened official accounts on digital platforms, which do not always effectively manage and moderate the comments generated.

It is necessary to reconsider the role of media ‘gatekeeper’ and replace it with ‘gatewatchers’ (evaluators of the reliability and appropriateness of information), taking into account individual criteria, professional and organizational routines, which shape the strategies (discursive–interactive, authoritarian–interactive, laissez-faire moderation and highlighting of comments) applied by moderators (often journalists). The responsibility falls on these moderators to ensure a minimum level of discursive quality among users and reduce the prevalence of what they view as hateful expressions (Paasch-Colberg and Strippel, 2021; Wintterlin et al., 2020; Wolfgang et al., 2020).

Social media has emerged as a significant venue for the proliferation of disinformation and expressions of hate (Barud and Olalekan, 2022; Sánchez et al., 2022). These platforms frequently disseminate rumours, stereotypes and prejudices to specific groups. This situation presents a challenge for news media, as the proliferation of expressions of hate and their impact on how readers perceive and interpret messages contribute to the polarization of public opinion behaviour (Anderson et al., 2014; Russmann and Hess, 2023; Wintterlin et al., 2020).

The current discourse on expressions of hate lacks consensus on their definition (Paz et al., 2020; Rossini, 2022). Debate is often centred on the varying hatred between related but distinct concepts, such as uncivil, hateful, abusive, offensive and violent expressions (Culpeper, 2021; Esau, 2021; Pelicon et al., 2021). In other cases, different variables are considered, such as expressions that aim to incite hatred towards vulnerable individuals or groups based on various factors (ethnicity, gender, religion, origin and political beliefs, among others) (Sellars, 2016; Ziegele et al., 2018; Struck, 2019). It is generally acknowledged that these expressions are marked by their public scope and intention to cause harm. The primary point of contention lies in the severity of these concepts’ violation of social, educational and legal norms in different countries and societies (Culpeper, 2021; Pelicon et al., 2021; Struck, 2019).

The academic community has focused on developing indicators and detection models for hate speech (Papcunová et al., 2023; Thorley and Saltman, 2023). In addition, research has been conducted to understand the general perception of hate speech (Cáceres-Zapatero et al., 2023) and categorize it into specific types based on themes, such as racial motives (Molnar, 2021), sexual (Arce-García and Menéndez-Menéndez, 2022) or religious (Obermaier et al., 2021).

The study of hate expressions encompasses a range of academic approaches by analysing the features of messages associated with such expressions to explore the role of media, journalists and social agents in promoting disinformation and hate speech through social media (Abuín-Vences et al., 2022; Paz et al., 2021; Vidgen, 2019; Vidgen and Yasseri, 2020; Williams, 2021). Research has also advanced our understanding of the ability of users to recognize strategies for intensifying political and social polarization (González-Aguilar et al., 2023; Tayie et al., 2023). However, progress has been limited in understanding how users perceive and receive hate expressions and characterizing those who spread such expressions (Małecki et al., 2021). In addition, research has been conducted on the presence of hate expressions in news media, particularly in situations where anonymity facilitates their disclosure (Paz et al., 2021; Đorđević, 2020).

Limited academic attention has been paid to the strategies employed by promoters and detractors of online expressions and their impact on disseminating hate speech and disinformation (Malecki et al., 2021). Recent studies, such as those conducted by Said-Hung et al. (2023) and Sprejer et al. (2023), have focused on the role of extreme right political actors in driving public opinion polarization in Spain and the United Kingdom, respectively. These findings highlight the need to deepen our understanding of the structural characteristics of hate messages and develop research avenues to assess the role of influential and non-influential haters in propagating such content across various social media platforms. To effectively combat and reduce the prevalence of this type of expression, it is crucial to cultivate a culture of critical thinking and responsible online behaviour, as suggested by Benesch et al. (2018) and Garland et al. (2023). This can be achieved by implementing strategies that promote the verification of information before sharing it, such as consulting accredited sources or verification organizations. In addition, promoting media literacy can help individuals evaluate information, identify this type of expression and understand the importance of reliable sources. Encouraging active user participation in reporting and flagging this type of expression and fostering online communities that promote positive discourse and activism against this type of discourse can also be effective measures.

It is essential to recognize that the phenomenon of haters is not solely attributed to automated accounts but also involves human-managed accounts that operate in a coordinated manner to influence public opinion in specific directions (Alhazbi, 2020). These human operators, sometimes called the ‘Internet water Army’ or ‘elite sibylline corps’, are employed by political or economic groups and utilize diverse tactics (Zheng et al., 2017). Furthermore, corporate media accounts often function as channels for disseminating negative stereotypes, prejudices and polarizing content towards marginalized groups. Above all, considering certain characteristics associated with the media in Spain (Humanes and Valera-Ordaz, 2023; Padró-Solanet and Balcells, 2022), such as their high political polarization and the increasing number of channels through which citizens receive political information (including social networks and digital news media users), it can lead to the formation of ideologically driven communication niches among users (due to the echo chamber and bubble filter effect of social networks, as previously mentioned). This environment can also contribute to the polarization of public opinion and create an ideal setting for the spread of such expressions by other users.

According to Keller et al. (2019), Astroturfing is a ‘common strategy of disinformation, where the disinformation pertains less to the content of the campaign, which can be completely truthful, but to the false impression of independent popular support’. (p. 2). It is a tactic that has been employed in political communication, public relations and marketing for several decades and has now been extended to the digital domain (Arce-García et al., 2023; Elmas et al., 2021; Pérez, 2020; Sorensen et al., 2017). The promoters of this strategy use alpha and beta accounts to initiate viralization actions and operate in an alternative and nomadic manner by responding to the interactions on the messages they disseminate and retransmitting the content to influential users on social media platforms, such as journalists and media personalities. This approach aligns with the respective research interests.

The research conducted thus far, including Guess et al. (2018), Zhao et al. (2020) and (Arce-García et al., 2023) has focused on the analysis of tactics such as Astroturfing in the dissemination of disinformation on social media. These studies presumed that hate speech constitutes another form of disinformation and polarizing resources. The results indicate that most individuals promoting hate speech are not influencers characterized by many followers and high activity but micro or nano-influencers with up to 100,000 and 10,000 followers or less, respectively. These individuals reportedly employ their perceived anonymity and lack of organization to exploit the connections or vulnerabilities they establish on social media platforms (Granovetter, 1973). In other words, they take advantage of their ‘common character’ vis-a-vis the rest of the users to interact with them in a certain sense only when it is necessary to guide certain perceptions. According to Arce-García et al. (2023), these accounts are characterized by having minimal interaction with others, except during the initial stages of content dissemination, tend to have a limited number of followers, exhibit a high level of message-sending activity and possess a small degree of centrality in terms of proximity to users or accounts that are considered influential.

Method

This research aims to elucidate the characteristics and behaviours of accounts that disseminate hate speech through their responses to articles posted on five prominent digital news media in Spain on X. The goal is to verify the application of dissemination techniques employed by such accounts and discern the customary practices adhered to by those who habitually propagate hate on social media.

The stated general objective will be attained through non-experimental quantitative research, which aims to achieve the following objectives:

SO1: Point out the features that characterize these ‘common character’ accounts that spread hatred in digital news media.

SO2: Identify the processes of dissemination of hate expressions.

SO3: Determine the traits that help to understand better the behaviour of ‘common character’ accounts that spread expressions of hate.

SO4: Establish the number of profiles of ‘common character’ accounts that disseminate hate expressions.

This study aims to corroborate the following hypothesis:

The dissemination of hate speech through digital news media in Spain is carried out by coordinated accounts, utilizing strategies that are similar to those employed in the viralization of disinformation, as documented by researchers such as Arce-García et al. (2023) and Keller et al. (2019), among others.

The research in question pertains to the data collected during January 2021, which comprises 266,536 messages originating from 981 X accounts, which have actively participated in publishing messages with hate expressions in content (news) disseminated by the five most prominent digital news media in Spain, as identified by Statista (2020a, 2020b, 2020c) and the Digital News Report (2020): La Vanguardia, ABC, El País, El Mundo and 20 Minutos.

The acquisition of messages was accomplished by utilizing X’s API 2.0 and the airflow platform. The automation process employed to extract the data is illustrated in Figure 1. The messages were gathered between February and March 2021, encompassing all communications referencing or citing content from the chosen digital news media.

Figure 1.

Directed acyclic graph (DAG) of the data collection process studied.

The messages were scrutinized and refined to ensure standardization, coherence and practicality by eradicating null, duplicate and vacuous messages and eliminating URLs, emojis and references to the news media, as demonstrated in this study. This procedure enabled the extraction of 215,301 messages comprising content disseminated by the media and commentary provided by followers or third-party users. Among these, 173,449 messages representing comments (responses) contributed by 981 followers or third-party users on platform X were chosen as the focus of the study.

From February to April 2022, all collected messages underwent a thorough review, one by one, following the methodology outlined in Figure 2. Because they were collected sometime after their publication, some messages can be deleted by the users themselves or by X itself, following the current policy of this platform, upon receiving reports from its users for verification and possible elimination of messages with hate expressions (X Help Center, 2023). Despite this, we considered the deletion cases of these messages, but they did not affect the analysis shown in this study. Above all, if we take into account that the data extracted reflect a period in which institutions such as the European Union (CBS News, 2022) and the Center for Countering Digital Hate (2023) have reported failures in the removal of messages with this type of expression, about an 82% decrease in detection and elimination, compared with 2021, and 86% of the messages of this type of expression continued to be hosted in X by 2023. Therefore, this study assumes that most of the discourse was still present when the analysed messages were collected.

Figure 2.

Process of manual labelling of analysed messages.

The process was adapted to align with the hate labelling manual (De Lucas et al., 2022), which categorizes messages into six levels of intensity based on their content:

Intensity 0: Expressions without conveying a derogatory meaning or a clear intention to insult or incite hatred nonetheless carry negative connotations (e.g. using nationality or another characteristic as a derogatory reference).

Intensity 1: Expressions that, without resorting to verbal aggression, present or describe objectively, a fact that stigmatizes a specific social group (e.g. falsely accusing someone or a group of wrongdoing).

Intensity 2: Expressions that attribute negative stereotypes to disseminate a negative image of a person or a specific social group (e.g. portraying immigrants as criminals).

Intensity 3: Expressions that use verbal aggression (offensive, derogatory or humiliating language) to target a group that is despised (e.g. using the term ‘son of a bitch’ to insult someone).

Intensity 4: Expressions that contain veiled or implied threats as well as harassment or joy at death, aggression or physical harm to someone or something (e.g. taking pleasure in someone’s death).

Intensity 5: Expressions that call for specific acts of physical violence against someone, a group or something associated with them (e.g. suggesting burning down a church or mosque).

The study universe and population labelling identified 1345 messages with hate expressions (0.77% of the total messages that comprised the study population). A total of 981 accounts of X published them. The sample distribution included 335 messages about news from 20 Minutos, 248 about ABC, 369 about El Mundo, 311 about El País and 82 about La Vanguardia. The classification around the intensity of hate was 75 in Level 0, 7 in Level 1, 3 in Level 2, 1217 in Level 3, 40 in Level 4 and 3 in Level 5.

A series of methodologies were proposed for their study to determine the nature and behaviour of accounts within the hate speech broadcast. To this end, we propose their analysis using network theory to determine their position within the social network, as well as the determination of variables that determine their connection with other accounts, such as their homophily or importance within the group (eigenvector). Using statistical regression or clustering, we try to determine if there are patterns of behaviour on these variables on accounts that spread hate speech.

The sample data were analysed using machine learning techniques through the application of Python software and the Rattle interface available in R (Williams, 2011). This study merges the networks collected around the five digital news media into a single dataset. Each dataset was collected in different databases and merged under the same fields in R to be analysed to identify patterns between the accounts that spread hate among the leading Spanish digital media. The integration of the data was done to facilitate the processing of the information raised at the level of specific and general objectives, where the interest of this work is not the hate message itself but the nature and behaviour of the issuers of the hate message. Users can participate in different X accounts of these media simultaneously, as the analysis platform allows them to do so. For this reason, general analyses (set of users), such as network analysis, are required to identify the nodes identified as senders of the 1340 hate speech messages detected at the level of all the digital news media studied. To determine the data corresponding to the network, we analysed the nodes that sent 173,449 messages about the news broadcast by Spanish digital media news, establishing direct connections between users through mentions, forwards or comments made by them.

The following data analysis techniques are considered to achieve the specific objectives proposed. In the case of SO1 the following were used:

- The Louvain algorithm in the Gephi program (Bastian et al., 2009). This study involved determining the modularity and cluster analysis of the network and assigning values to each node or account. These values include the eigenvector (which represents the importance of the node in the network and identifies its influencers), closeness centrality (which measures the closeness of the account to the influencers of its group on a scale of 0–1, with 0 being the furthest and one being the closest), intermediation centrality (which represents the account’s ability to connect other accounts and serve as a backbone for the network), degree (which counts the number of connections received and issued), degree with weight (which assigns importance to the connections between accounts based on their relevance to the entire network) and average distance length for other accounts.

- The degree of homophily, or the inclination to connect only with similar accounts or nodes globally regarding the total number of users, is crucial. Furthermore, assortativity represents the inclination to establish connections between nodes that share equal modularity (high levels) or nodes that belong to different modularities or groups (low levels). The Homophilic Exposure Index (HEI) variable, which calculates the probability that two nodes or individuals share a common characteristic among the total links in the network, is used to measure the HEI. This index focuses on sharing specific characteristics, whereas assortativity measures connections between nodes with similar traits. The calculations for both variables were performed using the Python network library.

The analysis associated with SO2 and SO3 was performed using the Rattle library in R (Williams, 2011), to perform the following: (1) multiple linear regression considers the eigenvector as a dependent variable and develops a graphic study of residuals to observe the quality of the regression and (2) polynomial regression, taking the intensity of hatred as the dependent variable, with the remaining variables as independent variables.

Finally, to obtain data associated with SO4, an analysis was carried out to determine the subgroups of users based on their nature and behaviour by applying t-tests and clustering of users using a k-means algorithm. This involves grouping messages with similar characteristics and/or behaviours to identify potential patterns. In this case, the number of clusters to be formed was determined by determining the highest value in the Dunn index, which facilitates the evaluation of clustering quality. Values from k = 2 to k = 16 were tested, and the highest value was obtained at k = 7. As the different variables had different scales, a data normalization process was employed, followed by the return of the data to its original scale for presentation and better comprehension. This normalization enhanced the algorithm’s performance without altering this study’s results. In addition, the reliability of the cluster assignment was assessed by examining the coordinate discrimination graph sample for the two primary components contributing to the pattern establishment by the algorithm.

Results

The present study identified 1346 messages within a network of message tracking and forwarding (RT) connections. These messages were distributed across various clusters around the five digital news media users. The information presented in Figure 3 illustrates the distinctive traits of the 981 users who disseminated hate messages within the context examined (SO1), as evidenced by their intermediation and eigenvector values. They present values close to zero in all news digital media analysed. Accounts that disseminate hate speech tend to have limited followers and do not actively seek to amplify their messages by engaging with hubs or influencer accounts. Instead, they frequently employ nano- or micro-influencer accounts that typically have 10,000 followers or fewer. This observation, which involves using degrees or degrees with minimal weight, suggests that these accounts possess limited reach unless they are considered as collective entities.

Figure 3.

Distribution of position values in the network of distributions and comments around the analysed digital news media.

One characteristic of the profile of accounts on X is their resemblance to those employing Astroturfing tactics. Data analysis reveals that these accounts exhibit shallow closeness centrality values, indicating that they are at the periphery of their respective clusters (Figure 3). According to Granovetter’s (1973) theory of weak ties, these accounts serve as effective conduits for disseminating novel ideas in public discourse. Specifically, they achieve this by maintaining sporadic connections and residing outside the primary hubs of opinion groups (i.e. hub accounts or influencers).

Regarding SO2, the data suggest that the most prevalent intensity levels of these messages correspond to Levels 0–3 (De Lucas et al., 2022). This refers to expressions that go as far as violent verbal allusions, using terms that insult, offend, despise or humiliate individuals belonging to a hated group (e.g. ‘Son of a bitch’). However, these expressions do not reach levels that imply veiled threats or calls for physical violence. Most expressions of hate do not exceed the maximum intensity levels and are likely to stay outside the limits of crimes established by the Penal Code. The low presence of messages with hate expressions of high level of hate intensity (4 or 5 on the scale used in this work) could be due to the potential elimination of these messages by the users themselves or by platform X. It also allows us to reflect on the possible capacity and relevance that messages with low intensity of hate would have to reach a more significant number of accounts, from the evasion (hiding) of actions implemented against this type of message through strategies mentioned by authors such as Benesch et al. (2018) and Garland et al. (2023). Something that could allow the use of hate expressions that are not legally penalized is to disseminate this type of expression more and better among more users. This could help position certain narratives within public opinion associated with the notices published by digital news media against certain social groups.

The study of accounts that disseminate hate expressions (SO3) necessitates a close examination of their characteristics, particularly their tendencies towards homophily (a preference for interacting with other accounts within the same group) and assortativity (the preference of a node to connect with nodes that are similar to it) when grouped alongside the digital news media accounts under investigation (as shown in Table 1). The findings reveal disparate and contradictory patterns compared with those expected for typical accounts. The assortativity values indicate that the nodes prefer to connect with familiar others. However, the degree of homophily is minimal; these accounts prefer to connect with other accounts that belong to different groups than their own. Moreover, minimal, close or equal average separation lengths exist between nodes surrounding each digital news media account. For instance, La Vanguardia has an average separation length of precisely 1 between nodes. This suggests that there is swift and close communication between accounts and that small-world phenomenon patterns may be at play. Thus, it is likely that accounts that disseminate hate speech would significantly impact accounts belonging to other nodes beyond those they are affiliated with.

Table 1.

Homophily values of the accounts in their networks.

	Assortativity	Assortativity Homophilic exposure index (HEI)	Global connections in the network	The average length between nodes
20 Minutos	0.67	0.04	0.16	1.15
ABC	0.64	0.02	0.19	1.25
El Mundo	0.59	0.04	0.16	1.2
El País	0.78	0.02	0.24	1.68
La Vanguardia	0.79	0.08	0.19	1

Accounts that promote hate expressions exhibit high assortativity and low homophily, which is atypical behaviour for a standard social group whose primary relationships are with acquaintances. These accounts have a strong bond within their group but lack similarity. They resemble a global network of accounts organized and utilized to support various groups in a decentralized manner rather than a single group. This structure renders them challenging to identify while possibly coordinated hate messages are being spread.

The next step involved performing a linear regression analysis on the variables related to the eigenvectors (as shown in Table 2) of accounts that disseminate hate expressions. The objective was to determine their behaviours and common characteristics statistically. The results indicated a highly significant statistical correspondence between the position in the network, as indicated by its eigenvector, and the intensity variables (p-value < .0001), weighted degree, betweenness (p-value < .01) and closeness centrality (p-value < .05). This suggests that these accounts are located far from the centre of their clusters, have a low profile and a position far from the centre of their clusters, consistent with the findings of studies that have addressed the use of Astroturfing strategies in social media (Arce-García et al., 2023; Keller et al., 2019).

Table 2.

Analysis of variance using linear regression concerning the eigenvector.

	df	Sum Sq.	M Sq.	F value	Pr(>F)
Intermediation	1	0.01079	0.0107940	7.8957	0.005059	**
Close centrality degrees	1	0.00701	0.0070105	5.1281	0.023770	*
Grade	1	0.00328	0.0032777	2.3976	0.121857
Grade with weight	1	0.01389	0.0138911	10.1612	0.001482	**
Intensity	1	0.02426	0.0242621	17.7474	0.00002767	***
Residuals	935	1.27822	0.0013671

p-value = 05, *.01, **.001, ***0.

The analysis of the residual distribution graphs of the linear regression model, as presented in Figure 4, lends credence to the model’s validity. In particular, the normal Q–Q plot reveals that the fit is not entirely linear and that most residuals are randomly dispersed, with only a few located in the lower region. This conclusion was tested using different data normalizations and significant variables, and consistent results were obtained.

Figure 4.

Residual plots in the linear regression model.

It is important to note that this finding does not necessarily indicate a failure of the model adjustment but rather the possibility that certain accounts may deliberately avoid conforming to a standard behaviour model to evade detection. This could be seen as an effort to disguise themselves as ‘normal’ and carry out a premeditated plan, particularly at the level of public opinion. Therefore, despite their similarity in being off the radar, these accounts are considered ‘abnormal’ and do not exhibit a normal distribution. This suggests that they do not follow a normal distribution despite their weaknesses.

Using a polynomial regression ANOVA, it was determined that there is a statistically significant relationship between the dependent variable of intensity and the remaining values (p-value < .0001). This finding supports the notion that the intensity of hate is linked to the importance of the account within the network.

To determine the number of profiles associated with the accounts that disseminate hate expressions (SO4), the k-means algorithm was employed to create seven distinct groups. This decision was based on the fact that this number yielded the highest Dunn index, which measures each cluster’s quality of separation and definition. The Dunn value for this study was 0.018, which was relatively small. This finding is consistent with the observation that nodes or accounts within a network exhibit similar characteristics and behaviours. Discriminant representation further supports this conclusion, based on the two main components or variables accounting for 56.37% of the distribution. Figure 5 illustrates how the seven clusters identified by the algorithm largely overlap, indicating the possible similarities in the nature and behaviour of these accounts.

Figure 5.

Discriminant coordinates concerning the two most notable components in the assignment to clusters according to k-means.

The data provided in Table 3 illustrate how the Dunn index effectively formed and displayed seven distinct clusters with nearly identical results. However, two clusters, particularly Cluster 6, comprise the majority of accounts, with an intensity of approximately 3. Cluster 6 has the lowest intensity of hate (2.81) and the most minor exponent activity (grade 9 and weight 12.34), indicating a lack of prominence. Most accounts within this cluster exhibited an activity profile that could be more notable, with minimal intermediation value. In addition, they have the second-lowest closeness centrality value, indicating that they are not closely connected to the network’s influencers. Cluster 1, the second largest group, has a slightly higher intensity of hate (2.98) but already exhibits more significant activity (grade 109.75 and weight 189.51), indicating a greater level of engagement. This suggests that these clusters desire normality because they are not highly active or influential within the network.

Table 3.

Cluster centres when applying k-means with seven groups.

Group	Intermediation	Eigenvector	Centrality closeness	Degree	Degree with weight	Hate intensity	Cluster size
1	0.00158	0.01125	0.12531	109.75	189.51	2.980	102
2	0.00000	0.06436	0.48715	1183.75	3305.75	3.000	4
3	0.00000	0.00874	0.72543	3020.00	6766.00	3.000	1
4	0.02300	0.01045	0.08306	231.39	475.55	3.020	51
5	0.09091	0.00848	0.37369	414.14	1050.55	3.000	22
6	0.00037	0.01506	0.02563	9.00	12.34	2.807	752
7	0.00000	0.00507	0.45868	680.22	2102.89	3.000	9

As depicted in Table 3, the other groups display a relatively small number of accounts yet exhibit a heightened level of hatred that falls short of reaching Level 4 intensity, as assessed in this study (group 4 with an intensity level of 3.02). This group demonstrated substantial network structuring or intermediation, albeit at reduced capacity. The most active account groups were 2 and 3, boasting grades 3020 and 6766, respectively, with a notable weight. Groups 5 and 7 exhibited an exact intensity value of 3 in hate, accompanied by intermediation values of 0 or nearly 0. Despite their minuscule eigenvectors, they are closer to influencers in their respective networks. Thus, the most active accounts are few but frequently engage in constant hate speech.

Table 3 presents data that support previous observations. Despite their similarities, all groups exhibited intense hatred around Level 3. Furthermore, these data suggest that highly active groups consist of a few accounts more central to their networks. These accounts could be called ‘kamikaze’ accounts, as they are prevalent on the platform, located at the clusters’ periphery and have minimal intermediation. They do not play a significant role in connecting users within a network. These accounts deviate from Granovetter’s (1973) weak link theory of introducing new ideas into public opinion, behaving differently and approaching the network’s centre for specific purposes, thereby explaining the deviations observed in the linear regression study.

Discussion

The data analysed in this study pertain to the specific objectives concerning the accounts of five Spanish digital news media on the platform, as investigated by Sorensen et al. (2017), Keller et al. (2019), Pérez (2020) and Said-Hung et al. (2023). These accounts are categorized as micro- or nano-influencers owing to their lack of a large number of followers, and they attempt to maintain a ‘low profile’ by staying away from the centrality of the different groups in which they engage in communicative actions, as indicated by the eigenvector values shown in Table 2. In addition, these accounts do not form structured networks through the studied digital news media because the intermediation between accounts is nearly zero.

The accounts that propagate hate expressions regarding the cases studied are likely to exhibit characteristics similar to those associated with the employment of strategies such as Astroturfing, as identified in studies on the dissemination of misinformation content on social media, including those conducted by Guess et al. (2018), Zhao et al. (2020) and others previously mentioned.

The dissemination of hate speech from accounts connected to digital media under investigation would adhere to Granovetter’s (1973) notion, which posits that such expressions originate from a supposed ‘common character’ distinct from the broader audience engaging with the content on social media platforms. This approach ensures that the intensity of hate speech does not surpass Level 3 on the scale used in this study (De Lucas et al., 2022), which avoids any implicit or explicit threats or calls for physical violence against specific individuals or social groups. Hence, these expressions do not violate legal norms, although they may still contravene social and educational standards. The perpetrators of hate speech aim to shape public opinion by transmitting insults, negative connotations and factual information to stigmatize or attribute actions to specific individuals or social groups. Concurrently, they can circumvent detection, monitoring and control systems for such expressions, as suggested by Vidgen (2019), Vidgen and Yasseri (2020) and Williams (2021). The connection between hate speech and an individual’s status within a network suggests that the dissemination of hate does not depend on influential figures but on individuals with a low profile who can connect with others on an equal footing. This type of discourse is more insidious and difficult to detect.

Based on the data analysed, the behaviour of users interacting with content disclosed by Spanish digital news media accounts would be anomalous or different. This behaviour deviates from what is typically observed in accounts that disseminate hate expressions. The analysis of assortativity, connection edge length, k-means and linear regression revealed that these users do not exhibit good association patterns favouring the echo chamber and bubble filter effect, as Pariser (2011) stated. Instead, the data allow us to recognize how their groups or clusters are related to many others, and their behaviour is similar to other accounts that disseminate hate expressions. Furthermore, there is a possible coordination between these ‘haters’ accounts around the content published by the digital news media through the accounts of X. This restriction on the ‘gatewatchers’ capacity of these media, as described by Russmann and Hess (2023), highlights the limitations of individual criteria, professional and institutional routines and strategies applied for this purpose.

Therefore, the hypothesis under consideration should be deemed acceptable within the confines of this study. Specifically, the dissemination of hate expressions emanating from accounts affiliated with digital news media in Spain appears to originate from profiles exhibiting traits and behaviours very similar to stratagems, such as Astroturfing (Arce-García et al., 2023; Keller et al., 2019). This scenario could amplify the dissemination of such expressions among the general public and, consequently, influence public opinion in favour of specific interests. These interests may aim to propagate particular themes and subject-specific social groups to the targeted attacks. The results obtained in each operation are sometimes based on measures with values that are not too categorical but taken as a whole; they raise serious doubts about their legitimacy and possible coordination.

Conclusion

In the parameters that define this study (the top five digital news media in Spain and the collection period), a situation is depicted that allows for an exploration of the dissemination of hate speech in recent years in light of misinformation. Both concepts, inherently connected to quickness, are typically utilized as communication tools to elicit reactions and increased interactions surrounding specific topics, groups or social collectives. One attribute of social media is its exploitation (in this instance, the aforementioned digital communication context; Chadwick, 2017; Matamoros-Fernández and Farkas, 2021).

These actions would be carried out, at least based on the data provided, by potentially coordinated accounts exhibiting natures and behaviours similar to those observed in content dissemination strategies such as Astroturfing, as proposed by Keller et al. (2019), Pérez (2020) and Elmas et al. (2021). Its small number of followers, low closeness centrality and high activity are elements consistent with typical Astroturfing activities (Arce-García et al., 2023). This could create an atmosphere wherein the supposed spontaneity or randomness of the haters’ accounts does not exceed the social, educational and legal norms of those who manage them. Undertaking this endeavour would involve circumventing the legal stipulations in countries such as Spain and transcending the moderators in digital news media on social media platforms (Russmann and Hess, 2023). Although data collected through various statistical and machine learning techniques is not perfect for defining possible coordinated behaviour, it can still provide helpful insights. While this study has limitations, it is important to note that the data collected can be used to understand possible coordination between users. This understanding should not be used to make unfounded accusations or assumptions but rather to guide further research on the potential connections between accounts and their impact on disseminating hate speech against individuals or social groups. Future studies should aim to build on this preliminary understanding and continue to advance our knowledge of how accounts may influence the spread of hateful narratives.

Social media platforms have emerged as fertile breeding grounds for the proliferation of disinformation and hate content, as evidenced by the recent studies by Sánchez et al. (2022) and Barud and Olalekan (2022). In the case of digital news media in Spain, there appears to be a lack of viable strategies for confronting or mitigating the impact of such expressions. This study found that although alerts may be an effective means of addressing incidents of intense hatred involving humiliation or hate crimes, these measures were rarely utilized by the accounts examined.

Undertaking a reevaluation of the current moderation protocols and methodologies applied to determine the veracity and pertinence of remarks posted by users on news articles and shared content on social media platforms is a matter of great importance. Equally crucial is the need to critically examine and implement measures that facilitate identifying and effectively countering ‘low-intensity or moderate hate’ which originates within the confines of these digital communication settings. This is especially true in cases where the behaviour and nature of ‘haters’ studied in this research mandate a comparably indefensible context, akin to a scenario in which a hundred randomly selected individuals would spontaneously execute the same dance steps without prior training or coordination.

A potential practical application of this research is the need to reconsider the actions currently taken by moderators in digital social spaces in digital news media and other social entities interested in this phenomenon. They must acknowledge that expressions of hate might be disseminated by potential agents operating covertly. Therefore, it would be beneficial to advance research that enhances and supplements strategies already implemented by the digital news media to combat this type of expression, such as those proposed by Benesch et al. (2018) and Garland et al. (2023). To improve early detection processes that serve these purposes and identify accounts that could employ tactics such as Astroturfing, it is crucial to make advancements in future work that help better comprehend the nature and behaviour of these users, as well as other potential factors that could be influencing the apparent coordination detected between them in this study.

The need for further study of accounts that promote hatred has been emphasized by Małecki et al. (2021). This current endeavour is not intended to provide the final word on the subject but rather to stimulate contemplation and investigate novel approaches to analysis, with the ultimate goal of enhancing our comprehension of this subject, which has gained considerable attention in both academic and social spheres. Analysing a substantial amount of data and prolonging the study’s duration is essential. Employing both quantitative and qualitative methods is crucial for providing complementary insights. This study distinguishes between the long-term strategic and short-term tactical measures employed to promote hateful messages on social media platforms. It is crucial to analyse digital news media and other forms of media in order to understand the various types and intensities of hate speech. By leveraging this project’s growing volume of data, we hope that subsequent studies will assist news media and journalists in moderating their online platforms effectively.

Footnotes

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship and/or publication of this article: This article results from the Hatemedia Project (PID2020-114584GB-I00), funded by MCIN/ AEI /10.13039/501100011033.

ORCID iDs

Sergio Arce-García

Elias Said-Hung

Author biographies

Sergio Arce-García, PhD, is Associate Professor at the School of Engineering and Technology (ESIT) at the Universidad Internacional de la Rioja (UNIR) and a Member of the research group COYSODI at UNIR. His main research areas are communication about social media and digital media.

Elias Said-Hung, PhD, is Full Professor at the Education Faculty, member of the research group SIMI and Director of the master’s degree in Inclusive and Intercultural Education of the Universidad Internacional de la Rioja (UNIR). His main research areas are ICT applied in Education, Social Media and Digital Media.

Julio Montero-Díaz, PhD, is Professor at the International University of La Rioja (since October 2014) and surplus as Professor of History of Social Communication (since 2007) at the Complutense University of Madrid. His main research areas are movies, history and media, and hate speech in news media.

References

Abuín-Vences

Cuesta-Cambra

Niño-González

, et al. (2022) Hate speech analysis as a function of ideology: Emotional and cognitive effects. Comunicar 30(71): 37–48.

Alhazbi

(2020) Behavior-based machine learning approaches to identify state-sponsored trolls on Twitter. IEEE Access 8: 195132–195141.

Anderson

Brossard

Scheufele

, et al. (2014) The ‘Nasty effect’: Online incivility and risk perceptions of emerging technologies. Journal of Computer-mediated Communication 19(3): 373–387.

Arce-García

Menéndez-Menéndez

(2022) Inflaming public debate: a methodology to determine origin and characteristics of hate speech about sexual and gender diversity on Twitter. Profesional de la información 32(1) e320106. Available at: https://doi.org/10.3145/epi.2023.ene.06

Arce-García

Said-Hung

Mottareale-Calvanese

(2023) Tipos de campaña Astroturfing de contenidos desinformativos y polarizados en tiempos de pandemia en España. ICONO 14. Revista Científica De Comunicación Y Tecnologías Emergentes 21(1): 1890.

Barud

Olalekan

(2022) Detection of fake news and hate speech for Ethiopian languages: A systematic review of the approaches. Journal of Big Data 9(1): 66.

Bastian

Heymann

Jacomy

(2009) Gephi: An open source software for exploring and manipulating networks [Conference]. International AAAI Conference on Weblogs and Social Media, San Jose, CA, 17–20 May. Available at: http://www.aaai.org/ocs/index.php/ICWSM/09/paper/view/154

Benesch

Buerguer

Glavanic

, et al. (2018) Danguerous speech: A practical guide. Dangerous Speech Project. Available at: https://cutt.ly/ew8f9T7l

Bright

(2016) The social news gap: How news reading and news sharing diverge. Journal of Communication 66(3): 343–365.

10.

Cáceres-Zapatero

Brändle

Paz-Rebollo

(2023) Stances on hate speech: Population opinions and attitudes. Profesional De La Información 32(4): 1–14.

11.

CBS News (2022) Twitter and other social media sites slipped on removing hate speech in 2022, EU review says. MONEYWATCH–CBS NEWS, 24 November. Available at: https://cutt.ly/qw8whZVk

12.

Center for Countering Digital Hate (2023) X Content Moderation Failure. Available at: https://counterhate.com/wp-content/uploads/2023/09/230907-X-Content-Moderation-Report_final_CCDH.pdf

13.

Chadwick

(2017) The Hybrid Media System. Oxford: Oxford University Press.

14.

Culpeper

(2021) Impoliteness and hate speech: Compare and contrast. Journal of Pragmatics 179: 4–11.

15.

De Lucas

Römer

Izquierdo

, et al. (2022) Manual Para El Etiquetado De Mensajes De Odio. London: Figshare.

16.

Digital News Report (2020) Se debilita la confianza en los medios, resisten las marcas periodísticas y emerge el periodismo local [Trust in the media weakens, journalistic brands resist and local journalism emerges]. Digital News Report. Available at: https://www.digitalnewsreport.es/2020/se-debilita-la-confianza-en-los-medios-resisten-las-marcas-periodisticas-y-emerge-el-periodismo-local/

17.

Đorđević

(2020) The sociocognitive dimension of hate speech in readers’ comments on Serbian news websites. Discourse, Context & Media 33: 100366.

18.

Elmas

Overdorf

Aberer

(2021). Tactical Reframing of Online Disinformation Campaigns Against The Istanbul Convention. Proceedings of the ICWSM Workshops. Available at: https://doi.org/10.36190/2021.42

19.

Esau

(2021) Hate speech (Hate Speech/Incivility). DOCA. Available at: https://doi.org/10.34778/5a

20.

Flaxman

Goel

Rao

(2016) Filter bubbles, echo chambers, and online news consumption. Public Opinion Quarterly 80: 298–320.

21.

Garland

Ghazi-Zahedi

Young

J-G

, et al. (2023) Impact and dynamics of hate and counter speech online. EPJ Data Science 12: 27.

22.

González-Aguilar

Segado Boj

Makhortykh

(2023) Populist right parties on TikTok: Spectacularization, personalization, and hate speech. Media and Communication 11(2): 232–240.

23.

Gottfried

Shearer

(2016) News Use across Social Media Platforms 2016. Journalism. Available at: http://www.journalism.org/2016/05/26/news-use-across-social-media-platforms-2016/

24.

Granovetter

(1973) The strength of weak ties. American Journal of Sociology 78: 1360–1380.

25.

Guess

Nyhan

Reifler

(2018) Selective exposure to misinformation: Evidence from the consumption of fake news during the 2016 U.S.Presidential campaign. European Research Council. Available at: https://cutt.ly/FOgUe1R

26.

Howard

Bolsover

Kollanyi

, et al. (2017) Junk news and bots during the US election: What were Michigan voters sharing over Twitter? [Conference]. Computational Propaganda Project-Oxford Internet Institute, 1–26 March, Oxford, United Kingdom. Available at: https://cutt.ly/kRihRoY

27.

Humanes

Valera-Ordaz

(2023) Partisanshipm, ideology, and selective exposure: A longitudinal analysis of media consumption in Spain (2008-2019). Media and Communication 11(2): 113–116.

28.

Karlsson

Couvering

Lindell

(2022) Publishing, sharing, and spreading online news: A case study of gatekeeping logics in the platform era. Nordicomreview 43: 190–213.

29.

Keller

Schoch

Stier

, et al. (2019) Political astroturfing on Twitter: How to coordinate a disinformation campaign. Political Communication 37: 256–280.

30.

Małecki

Kowal

Dobrowolska

, et al. (2021) Defining online hating and online haters. Frontiers in Psychology 12: 744614.

31.

Matamoros-Fernández

Farkas

(2021) Racism, hate speech, and social media: A systematic review and critique. Television & New Media 22(2): 205–224.

32.

Meese

Hurcombe

(2020) Facebook, News media and platform dependency: The institutional impacts of news distribution on social platforms. New Media & Society 23(8): 2367–2384.

33.

Molnar

(2021) The law of the jungle. The online hate-speech against the Roma in Romania. Victims & Offenders 16(8): 1108–1129.

34.

Obermaier

Schmuck

Saleem

(2021) I’ll be there for you? Effects of islamophobic online hate speech and counter speech on Muslim in-group bystanders' intention to intervene. New Media & Society 25(9): 2339–2358.

35.

Paasch-Colberg

Strippel

(2021) ‘The Boundaries are Blurry. . .’: How comment moderators in Germany see and respond to hate comments. Journalism Studies 23: 224–244.

36.

Padró-Solanet

Balcells

(2022) Media diet and polarisation: Evidence from Spain. South European Society and Politics 27(1): 75–95.

37.

Papcunová

Martončik

Fedáková

, et al. (2023) Hate speech operationalization: A preliminary examination of hate speech indicators and their structure. Complex & Intelligent Systems 9(3): 2827–2842.

38.

Pariser

(2011) The Filter Bubble. London: Penguin Books Ltd.

39.

Paz

Cáceres-Zapatero

Martín

(2021) Suscripción a la prensa digital como contención a los discursos de odio [Subscription to the digital press as a containment of hate speech]. Profesional De La Informacion 30(6): 1–13.

40.

Paz

Mayagoitia-Soria

González-Aguilar

J-M

(2021) From polarization to hate: Portrait of the Spanish political meme. Social Media + Society 7(4): 2920.

41.

Paz

Montero-Díaz

Moreno-Delgado

(2020) Hate speech: A systematized review. SAGE Open 10(4): 3022.

42.

Pelicon

Shekhar

Škrlj

, et al. (2021) Investigating cross-lingual training for offensive language detection. PeerJ 7: e559.

43.

Pérez

(2020) “Yo fui un bot”: las confesiones de un agente dedicado al engaño en Twitter [“I was a bot“: The confessions of an agent dedicated to deception On twitter]. El País, 21 May. Available at: https://cutt.ly/wRihXGu

44.

Rossini

(2022) Beyond incivility: Understanding patterns of uncivil and intolerant discourse in online political talk. Communication Research 49(3): 399–425.

45.

Russmann

Hess

(2023) The management of uncivil and hateful user comments in Austrian news media. Journalism Practice. Epub ahead of print 16 March. DOI: 10.1080/17512786.2023.2189152

46.

Said-Hung

Moreno-López

Mottareale-Calvanese

(2023) Promotion of hate speech by Spanish political actors on Twitter. Policy & Internet 15(4): 665–686.

47.

Sánchez

Campo-Archbold

Rozo

, et al. (2022) On the power of social networks to analyze threatening trends. IEEE Internet Computing 26: 19–26.

48.

Sellars

(2016) Defining Hate Speech. Berkman Klein Center Research Publication No. 2016-20. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2882244

49.

Sorensen

Andrews

Drennan

(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.

50.

Sprejer

Margetts

Oliveira

, et al. (2023) An actor-based approach to understanding radical right viral tweets in the UK. Journal of Policing, Intelligence and Counter Terrorism 18(2): 139–157.

51.

Statista (2020a) Ranking de los 10 medios de comunicación con más seguidores en Twitter en España en abril de 2020 [Ranking of the 10 media outlets with the most followers on Twitter in Spain in April 2020]. Statista. Available at: https://es.statista.com/estadisticas/518802/Twitter-perfiles-de-medios-de-comunicacion-con-mas-seguidos-en-espana/

52.

Statista (2020b) Páginas oficiales de medios de comunicación más populares en Facebook en España en abril de 2020, por número de seguidores. Statista. Available at: https://es.statista.com/estadisticas/518580/paginas-de-medios-de-comunicacion-de-Facebook-con-mas-seguidores-en-espana/#statisticContainer

53.

Statista (2020c) Ranking de las principales marcas de medios de comunicación online según el porcentaje de población que las usaba de forma semanal en España en 2020 [Ranking of the main online media brands according to the percentage of the population that used them weekly in Spain in 2020]. Statista. Available at: https://es.statista.com/estadisticas/670743/principales-medios-online-porcentaje-de-usuarios-semanales-espana/r

54.

Struck

(2019) Digitale vorwärtspanik. Monatsschrift Fur Kriminologie Und Strafrechtsreform 102(1): 54–64.

55.

Swart

Peters

Broersma

(2019) Sharing and discussing news in private social media groups. Digital Journalism 7: 187–205.

56.

Tayie

Tejedor

Pulido

(2023) News literacy and online news between Egyptian and Spanish youth: Fake news, hate speech and trust in the media. Comunicar 31(74): 73–87.

57.

Thorley

Saltman

(2023) GIFCT tech trials: Combining behavioural signals to surface terrorist and violent extremist content online. Studies in Conflict & Terrorism. Epub ahead of print 19 June 2023. DOI: https://doi.org/10.1080/1057610x.2023.2222901.

58.

Van-der-Linden

Maibach

Cook

, et al. (2017) Inoculating against misinformation. Science 358(6367): 1141–1142.

59.

Vidgen

(2019) Tweeting Islamophobia. Doctoral Thesis, Oxford University Research Archive, University of Oxford. Available at: https://ora.ox.ac.uk/objects/uuid:3c32d29d-e2e4-4913-abf8-2a28886f55a7

60.

Vidgen

Yasseri

(2020) Detecting weak and strong Islamophobic hate speech on social media. Journal of Information Technology & Politics 17(1): 66–78.

61.

Williams

(2011) Data Mining with Rattle and R: the Art of Excavating Data for Knowledge Discovery. New York: Springer.

62.

Williams

(2021) The Science of Hate. London: Faber & Faber.

63.

Wintterlin

Schatto - Eckrodt

Frischlich

, et al. (2020) How to cope with dark participation: Moderation practices in German newsrooms. Digital Journalism 8(7): 904–924.

64.

Wolfgang

McConnel

Blackburn

(2020) Commenters as a threat to journalism? How comment moderators perceive the role of the audience. Digital Journalism 8: 925–944.

65.

X Help Center (2023) Hateful Conduct. Available at: https://help.twitter.com/en/rules-and-policies/hateful-conduct-policy

66.

Zerback

Töpfl

(2021) Forged examples as disinformation: The biasing effects of political astroturfing comments on public opinion perceptions and how to prevent them. Political Psychology 43(3): 399–418.

67.

Zhao

Sano

, et al. (2020) Fake news propagates differently from real news even at early stages of spreading. EPJ Data Science 9(7): 1–14.

68.

Zheng

Xue

Hao

, et al. (2017) Smoke screener or straight shooter: Detecting elite sybil attack. Social and Information Networks. Available at: https://arxiv.org/abs/1709.06916

69.

Ziegele

Koehler

Weber

(2018) Socially destructive? Effects of negative and hateful user comments on readers’ donation behavior toward refugees and homeless persons. Journal of Broadcasting & Electronic Media 62(4): 636–653.