2.1 Alternative media content on social media platforms

The existing research demonstrated that platforms play quite a significant role in news sharing and distribution of alt-right content [11]. In the Swedish context, alternative media readership was found to stand out from a sample of other Northern and Central European countries [11], which makes it hard to underestimate the role of this information source and the potential effects it may have on the readers’ attitudes and political orientations. On top of this, alternative media readership was associated with distrust in the mainstream media, however, it also needs to be mentioned that it was found to supplement rather than completely replace traditional news outlets [12]. On social media and, in particular, on Twitter, alternative media presence was described in terms of “echo-system” where the same alternative media content was disseminated by an ideologically diverse set of actors [13]. On Facebook, links to Swedish alternative media were found to constitute almost one-third of all URLs shared during the 2018 election period, with link engagement levels almost on par with that of traditional media [14]. Even more so, the existing research demonstrated that Swedish right-wing actors were more successful in engaging the public during the 2018 election in comparison with their established counterparts, in particular, by means of resorting to offensive language and negative emotionality [15]. A similar study in the US setting also found that ideologically extreme news sources enjoyed the highest engagement levels on Facebook, despite that well-known mainstream publishers had larger audiences, since the latter did not necessarily imply high engagement with their news content [16].

When it comes to the use of alternative versus mainstream media in right-wing online environments, the existing research pointed out that, quite surprisingly, the use of and references to mainstream media sources are just as popular as those to alternative ones [17, 18], suggesting that right-wing audiences consume different types of content despite the dominating narrative of mistrust into the established media. Thus, the research suggests, despite the circulation of illiberal or radical opinions inside fringe and echo chamber-like environments, such platforms still allow for the dissemination of cross-cutting content and can nevertheless serve as deliberative spaces [19].

With regard to the dynamics and patterns of alt-right content sharing, a multi-platform comparison of alternative versus mainstream media news spread demonstrated that some of the fringe communities across the platforms generate a disproportionately high volume of alternative links sharing events that also seem to influence the emergence of the same content on other platforms [20]. Another study by Wang et al. that considered trustworthy and untrustworthy news sources instead of mainstream versus alternative ones arrived at a similar result and suggested that small communities can become extremely effective in pushing news stories to other communities [21]. Another relevant contribution to the existing evidence was made by Luna et al. who highlighted the differences in the dissemination of mainstream versus alternative media on Facebook: while alternative news content sharing was steadily increasing with time after the publication, mainstream news content sharing, in contrast, followed a burst and decay pattern [22]. Despite the existing evidence, little is known about the impact of alternative media content on the dynamics of discussions in fringe communities and whether such content reinforces or even produces more negative attitudes and narratives about the discussion topics. Some of the existing exceptions is a study by Introne et al. who provided a description of how pseudo-knowledge comes into being as a result of collaborative effort in the forum discussions [23].

2.2 Collective emotions and emotional contagion

Quite logically, negative and emotional content was shown to catch more users’ attention and disseminate more intensively [24, 25]. Likewise, the content diffused by alternative media was shown to be more negative and emotionally charged in comparison with the mainstream media content [26] and was shown to be more likely to express such emotions as anger and disgust [27, 28]. On Facebook, immigration- and security-related posts by right-wing actors were found to be especially likely to cause “anger” reactions and further sharing of those posts [29]. However, Berger and Milkman [30] demonstrated also that the relationship between emotion and virality is more complex than that: in particular, some emotions, such as sadness, might actually have a reverse, deactivating, influence on content virality. Rather, the emotions that evoke high arousal (no matter if they are positive or negative) are the main driving force for the diffusion of any type of content in online networks.

Some of the existing studies, mostly based on the results of experiments and agent-based modelling, also suggested that emotional content leads to higher arousal [31], while the valence of posts published on Facebook was found to depend on the valence of posts previously seen by a given user [32]. Another observational study of Twitter users and their followees’ posts arrived at a similar conclusion, thereby giving evidence to the presence of emotional contagion [33]. Some works suggested also that negative emotions, such as anger, expressed in the initial social media content, are more contagious than positive ones, such as joy [34, 35]. Yet another study demonstrated that, although the emotional tone of users’ messages is usually adapted to that of the chat, chat’s emotion tone can be characterized as persistent, rather than fluctuating towards particular negative or positive emotions [36]. Finally, concerning the connection between emotions and online deliberation, the previous research has distinguished constructive and non-constructive emotions, with the former being the only type of expression that can positively affect the quality of online deliberation [37]. A case study of the Chinese online platform Weibo, in contrast, has demonstrated a limited possibility for the platform to serve as a space for deliberation, with emotional content being a limiting factor [38].

What distinguishes our approach from the earlier effort is that we focus not only on the process of emotional contagion online in general (e.g. through news feeds and following/befriending other users) but rather on the group dynamics emerging as a result of users’ participation in online conversations that may lead to different dynamics and consequences of exposure to different types of emotions in news content.

2.3 Conversation dynamics online

With regard to the properties and dynamics of conversations, the existing research focused on several aspects such as the roles of individuals [39, 40], factors that make conversations interesting [41], conversations’ length and user participation prediction [42], or even the outcome of conversations in terms of demonstrated prosocial behaviours, such as social support, cohesion or mentoring [43]. Another piece of work employed a conversational approach to model content disputes on Wikipedia talk pages and developed a model for dispute escalation prediction task [44].

Since many of the existing works are based on the assumption that threads themselves represent distinct conversations, such a thread-based structure was found to enable higher reciprocity between users [45]. Further, work by Bagavathi et al. attempted to study conversations on Gab as an example of cascades and identified several types of cascades typical on the platform [46]. A similar approach to model conversation dynamics as a cascading process was proposed by [47]. Yet another approach to conversation reconstruction also took into account the linguistic features of the messages to determine the reply-to relationship between them [48]. The study most relevant to our empirical case, even though it does not explicitly focus on conversations or alternative news content, is the study by Caetano et al. [49] who explored structural, temporal and user engagement properties of “attention cascades” with falsehoods in non-political versus political WhatsApp groups. In particular, they found that the latter generate deeper and wider cascades [49].

Finally, it is also worth mentioning the effort to adapt Randall Collins’ micro-sociological approach to account for mediated forms of interaction and online communication, even though it was initially created with only offline interactions in mind. For instance, DiMaggio et al. applied Collins’ IRC theory to explore users’ behaviours online. This study also used a thread-based approach to online interactions between the users and explored the predictors of thread length, with the latter serving as a way to evaluate the interactions’ success [50]. Another interesting example is the work of van Harpenen et al. who applied methods of image recognition to develop a typology of mediated interaction rituals using a dataset of Instagram pictures related to the Black Lives Matter movement [51]. They proposed that such visual content served as a tool to connect with the movement despite geographical and physical separation, which also speaks in favour of the theory’s relevance for mediated types of communication.

4.1 Conversation reconstruction

Our two primary assumptions for the extraction of conversations is that, on the core level, a) a discussion thread consists of one or several conversations, depending on the frequency of interactions, and b) that such conversation can be represented either as a temporal network consisting of interconnected messages or, on an aggregated level, of users and links between them. However, in contrast to offline conversations that share the same spatial and temporal dimensions, online interactions are asynchronous. As a consequence, not all forum users participate in the conversations simultaneously. We note that more than two-thirds of all messages (third quartile) are published within a rolling window of 27 minutes (Δt < 27), and we create a link between any pair of messages in a thread that is published within this time span, also known as Δt-temporally adjacent messages [66]. We validate the robustness of this rolling window choice by testing alternatives: the thresholds that are two times shorter (13 minutes) and two times longer (54 minutes) to account for the fact that some of our results may depend solely on the choice of the rolling window. Further, we replicate the whole analysis for the conversations created using the three rolling windows and confirm that our results hold irrespective of the chosen threshold (unless specifically mentioned in Section 5).

To provide an example, if user A posts a message at time t₁, and user B at time t₂ (within 27 minutes after A’ message), then we assume that user B has read user A’s message and potentially taken its content into account while writing their own message. Thus, we create a link from B’s to A’s message. Additionally, we create also a link from B’s message to any other message quoted in it (similar to a reply on Twitter), irrespective of the amount of time that has passed, since this indicates a direct reference to the message previously published.

Thus, we may assume that a conversation, on an aggregated level, tends to capture the amount of information from a given thread that a given user will consume. Moreover, our approach to conversations takes into account the sequencing of messages, which helps to account for the co-evolution of users’ narratives. Lastly, since we are also interested in studying the properties of conversation networks, we explore a range of the resulting networks’ basic characteristics. Those include: average clustering coefficient, total number of edges and nodes, and edge density, all available as part of the NetworkX package in Python [67].

4.2 Classification of stances and sentiments

To quantify the sentiments on and evaluations of a specific topic (immigration and corresponding policies in Sweden), we use a notion of stance and follow the definition of stance-taking as a subjective evaluation or appraisal of a specific target [68, p. 142, 153], “explicitly mentioned or implied within the text” [69]. Although this task is closely related to that of classic sentiment polarity detection, the difference between the two is that sentiment analysis is concerned with identifying the sentiment polarity of a text in general, whereas stance classification requires a specific target (be it an entity, topic, claim etc.). For instance, the following statement “It is frustrating that we still need to fight against implicit ethnic discrimination of newcomers in our country” can be described as having negative sentiment polarity, but at the same time taking a positive or supportive stance with regard to the immigration topic.

Thus, in this work, we aim at identifying user stances with regard to the immigration issue. We use a two-step machine learning approach trained specifically on the data from the Flashback forum. The details on model training and selection are discussed in detail in [70]. In short, on the first step, our approach identifies on-topic and off-topic messages, since, as mentioned above, we are only interested in messages discussing immigration. In the second step, particular stances are determined (two classes: negative and non-negative, the latter includes neutral and positive messages). Choosing a two-class approach allows both to compensate for class imbalance (positive messages are under-represented in the corpus, accounting for only 5% of the training dataset), as well as distinguish negative rather than moderate expressions about immigration, whether positive or neutral.

Stance classification is further supplemented with an unsupervised sentiment analysis using the VADER dictionary [71]. Since this is a dictionary-based and unsupervised approach, it suffers from the limitations typical for this family of sentiment analysis methods such as lexicon narrowness and domain independence. However, our previously performed experiments demonstrated its reasonable performance for social media texts in Swedish, with an accuracy of 62%, which is on par with the previously performed tests (e.g. see [72]).

4.3 Identifying immigrant-critical alternative media

We use the definition immigrant-critical alternative media and a relational approach suggested by Kristoffer Holt who states that “Alternative news media represent a proclaimed or (self-) perceived corrective, opposing the overall tendency of public discourse emanating from what is perceived as the dominant mainstream media in a given system.” [73]. Other important traits include alternative news producers, in particular, non-professional actors, such as readers and activists; alternative news content, or the narratives that are perceived as counter-hegemonic and marginalised in the mainstream discourse; and, finally, alternative news organisations, in particular, in low-cost formats, such as blogs and webpages [73]. In contrast, we define mainstream media as established news organisations that rely on the work of professional journalists, adhere to ethical and professional journalistic standards and have an editor or an editorial board responsible for the content produced.

Our resource labelling process begins with the extraction of the most used domains mentioned by the users. In total, we extract 332 unique resources that comprise only 1.5% of all unique resources shared on the subforum in 2019–2022. Nevertheless, those resources account for 75% of all links shared on the forum. Further, we exclude the resources that do not belong to the media category (for instance, websites of private companies, non-profit organisations and government agencies). The rest of the resources are evaluated manually based on a range of criteria following the approach described in [73], in particular a) non-oppositional or corrective stance in relation to what is perceived as mainstream immigration discourse, b) availability in high-cost formats (radio, printed press, TV), c) adherence to the press ethical norms and standards, d) a presence of editor-in-chief or editorial board, or similar actors responsible for the published content, e) belonging to a larger organisation or publisher.

Further, the resources are assigned one of the three labels: mainstream/legacy media, immigrant-critical alternative media (or simply alternative media) and other media. The mainstream media label is assigned to the resources that are evaluated positively on the criterion a) (as providing non-oppositional framing of the public agenda), plus on at least three other criteria. The procedure for labelling resources as immigrant-critical alternative media is exactly the opposite: those are evaluated negatively on the criterion a) plus on at least three other criteria. The rest of the resources are labelled as other media, since the distinction between mainstream and alternative media is rather blurred and represents a continuum, as noted by the previous research [73], rather than a set of distinct media format categories. The details of resource evaluation criteria and labelling are presented in the S1 and S2 Files. Finally, it needs to be mentioned that social media groups on popular platforms (e.g. Twitter or Facebook) are excluded from our classification since they account only for a minor share of links disseminated on the forum.

4.4 The dynamics of collective stances and emotions and the role alternative media

As a first step, we compare the characteristics of a) conversations with alternative media versus mainstream media links shared, and b) conversations with any type of links versus no links shared. To that end, we use Mann-Whitney U test that fits well to compare samples where data is non-normally distributed. The conversations are evaluated on a range of criteria, including:

• duration in hours;
• the ratio of messages with negative and non-negative stances;
• average stance probability;
• average clustering coefficient;
• the number of nodes and edges;
• edge density coefficient.

Further, we explore conversation entrainment by analyzing the conversation pace and the evolution of users’ sentiments. In other words, a successful ritual, as defined by the theory, can be measured in two different ways: first, by conversation pace, measured as the frequency of interactions in the conversation, and, second, by the evolution of sentiments and stances, which in the case of fringe or right-wing environments, can be denoted by the dominating negative moods and evaluations of the immigration topic. Even more so, a successful conversation may also imply the alignment of users’ pace of interactions and subjective expressions, so that each new message arriving as part of the conversation will be dependent on the previous message(-s) posted as part of the same interaction chain. To measure such conversation entrainment, we use two sets of mixed effects models with conversations representing a random (grouping) effect, aiming to detect whether there is a relation between the subsequent messages in each conversation. Our sample includes a total of 6,674 conversations, all of which have at least one observation after obtaining time-lagged data. All analyses are performed using lme4 package in R [74].

5.1 Alternative media in the forum conversations

In total, we collected data from 4,692 different threads on the subforum that were started during the 2019–2022 election period. Those resulted in 9,304 conversations (with a rolling window of 27 min.), with an average conversation length of 28.4 and a median length of only 5 messages respectively, suggesting that the vast majority of conversations are rather short. As explained in the Methods section, our conversation networks include two types of edges, those between the messages posted in the time span of 27 minutes, and also those that denote replies to (or quotes of) other users, since the latter represent direct interaction between the conversation participants that can happen outside of the rolling window limit. Indeed, if we compare different types of edges, we find that user quotes take longer to be posted, and this is even more evident if we consider quotes that include any type of links (See Fig 1). Negative and non-negative messages take an equally long time to publish, while off-topic messages appear slightly faster in the conversations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Conversations’ pace depending on edge type, links shared and message stances.

(left): Interaction time intervals for different types of edges and message stances. (right): Interaction time intervals for different types of edges and links shared.

https://doi.org/10.1371/journal.pone.0294636.g001

Our results indicate that, quite unexpectedly, links to alternative media (hereafter AM) comprise only a minor part of the links shared on the forum: only 12% of all conversations on the forum has at least one AM link shared, and more than every third one (34%) at least one mainstream media (hereafter MM) link shared. Likewise, of more than 14,000 unique users in the dataset, only a small part (8% or 1,147 participants) is responsible for the dissemination of alternative media content.

Speaking of particular sources shared, we note that mainstream media are circulated far more actively than alternative ones, which is somewhat striking since one would expect the forum’s audience to favour the sources that would problematize the existing immigration discourses (see Table 2). Nevertheless, we find that the shared resources cover the whole ideological spectrum with regard to the immigration agenda: from the more “conventional” narratives articulated by the largest national publishers (such as SVT and Aftonbladet) and all the way to the extreme right resources that at times resort to hateful and extremist rhetoric when talking about immigration and immigrants in Sweden (for instance, Nordfront). In between these two extremes of the spectrum can also be found resources that were assigned to other category, such as, for example Nyheter Idag, since they do not fully comply with our understanding of legacy media, but which at times are successful at mimicking the behaviour and publishing strategies of the legacy outlets. Somewhere in the middle of this spectrum can also be found international resources such as Sputnik News and Russia Today that clearly stand out in their interpretation of political and policy issues.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Most frequently shared media sources.

https://doi.org/10.1371/journal.pone.0294636.t002

Our hypothesis is that conversations, where alternative media content is shared, exhibit higher user activity since such content can spark and provide some ground for users’ discussions. Besides, it may reinforce and, potentially, intensify users’ already existing negative preconceptions about immigration and immigrants. We use the Mann-Whitney U test to check this hypothesis and compare the means of conversation samples where alternative versus mainstream media content is shared. Since different conversations have varying ratios of links shared, we perform a range of tests using different link ratio thresholds calculated as the number of links of a given type divided by the number of messages in the conversation (0.2, 0.4 and 0.6).

We find substantial differences in user stances on the immigration topic after comparing the characteristics that differentiate conversations with alternative links from those with mainstream ones (Table 3). In particular, conversations, where alternative media content is shared, tend to have a higher ratio of negative messages. At the same time, quite a counter-intuitive observation is that conversations with alternative media content also tend to be shorter, with a lower number of nodes and edges between them, but, at the same time, higher edge densities (p < 0.05 in both cases irrespective of the chosen threshold for the link ratios).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Conversation properties conditional on the link types shared.

https://doi.org/10.1371/journal.pone.0294636.t003

We also decide to test an alternative hypothesis—for instance, one can suggest that link sharing in general, irrespective of link type, can affect the properties of conversations. Indeed, following the same procedure as the one described above, we find that link sharing in general makes the conversations longer, and leads to a higher edge density and a higher number of nodes and edges involved (Mann-Whitney U test, p < 0.05 irrespective of the chosen threshold for the link ratio in the conversations). Thus, one may suggest that content sharing can serve as a factor to spark conversations and users’ deliberations, irrespective of the type of the cited source.

Our last observation is also that conversations with any link types tend to have a lower share of negative evaluations of the immigration agenda in comparison with the conversations where no media content is shared (p < 0.05 irrespective of the threshold for the link ratios). Such conversations also tend to have a higher share of neutral and positive messages, which, however, holds only for conversations with a sufficiently high share of links in the conversation (p < 0.05 for hyperlink ratio over 0.4). Finally, we find that there are not only differences in the users’ stance polarities but also in their strengths denoted by the probabilities of belonging to a given category (negative or non-negative) calculated by the stance classification model. In particular, in the conversations with any type of links shared, stance probabilities are lower for the negative messages and higher for the non-negative ones (p < 0.05 irrespective of the threshold for the link ratios).

5.2 Conversation evolution and the role of alternative media content

Using mixed-effects models as explained in Section 4.4.1, we evaluate whether the previous conversation states at (t − 1) can be used to predict the subsequent conversation states at (t), which denotes conversation entrainment and users’ alignment with regard to the conversation pace and shared sentiments. On top of this, we evaluate whether link sharing (either alternative media, mainstream media or any links in general) has any impact on the evolution of conversations and the emotional states of its participants. In total, 6,674 conversation chains were included in the analysis. The minimal number of observations per conversation was 2 due to the need to include lagged observations in the models.

Most importantly, our results indicate that conversation entrainment is more visible if we consider the pace of interactions rather than message sentiments: in particular, the sentiments of the previously shared messages serve as significant (p < 0.05) but weak predictors for the subsequent message’s sentiments, suggesting a weak dependence between the sentiments of messages in the same conversation. The model that includes lagged message sentiments as a dependent variable performs only marginally better than the null model (see Table 4). This suggests that users’ emotional alignment is less discernible in comparison with conversations’ temporal alignment—conversation participants are not prone to picking up the emotional states of other participants. Note that, in this case, we use only the results of VADER sentiment analysis, which is a general sentiment analysis tool not capable of capturing the differences in users’ moods targeting the immigration topic specifically.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. The results of mixed-effects modelling.

https://doi.org/10.1371/journal.pone.0294636.t004

In contrast, our model for the conversation pace is more successful in capturing the temporal relationship between the messages, and we find that the pace of preceding interactions can be used to predict the pace of subsequent interactions. Our model with lagged message posting intervals as an independent variable performs significantly better than the null model (ANOVA test, χ² = 12, 294, p < 0.05, AIC 92,216 versus AIC 10,4507 respectively). In particular, lagged posting time intervals can serve as a significant predictor for the posting intervals of the subsequent messages (Estimate = 0.221, SE = 0.002, p < 0.05). We also note the positive estimate for the predictor, suggesting that conversations’ paces get longer with time, which also denotes a decaying pattern of the interaction pace. Nevertheless, our results indicate that we can infer the pace at which participants will be interacting by knowing how fast or slow the were communicating on the preceding steps.

Turning back to the question of the role of alternative media content in the conversations, we notice, once again, that link sharing, whether from alternative media or any other type of resources, does not have any sufficient effect on the conversation dynamics neither with regard to the conversation pace nor the dominating sentiments (p > 0.05 in both models). In our data, link sharing, irrespective of link type, has a significant impact on the conversation dynamics only when used as a predictor for conversation pace (p < 0.05), however, it does not improve model performance (AIC 92,193.8, conditional R² 0.135 for the model with time-lagged post intervals data only versus AIC 92,189.3, conditional R² 0.135 for the model also accounting for previously shared links).

6.1 Limitations of the work

One of the constraints with regard to the chosen methodological approach is of course that only a subset of the most frequently shared links was labelled, which might have left out some alternative media sources that are much smaller in terms of readership in comparison with established media. The same goes for social media groups and accounts that also generate online content. However, we note that their prevalence is much more limited in comparison with other media resources and that social media pages and groups do not necessarily follow the same publishing logic as media resources. Another constraint of our analysis lies in the fact that it is difficult to disentangle different conversations that are part of the same thread, and identifying the timeframe for conversation entrainment can be performed in different ways. Our solution to account for this limitation was to test alternative time limits. Finally, there is also a possibility to test more complex statistical models beyond linear mixed-effects models to validate the results of our analysis.

Another limitation relates to resource labelling, which is in many ways a problematic task, since, as mentioned earlier, mainstream—alternative represents a continuum rather than a distinct set of categories, which makes it difficult to draw the boundary between different kinds of resources, and especially given that some alternative resources try to mimic the behaviour of their mainstream counterparts. Manual labelling also leaves some space for annotator subjectivity, especially with regard to the identification of oppositional or corrective stances. The latter is an especially difficult task, since it presumes deep contextual knowledge of the current public discourse on immigration. We try to mitigate this constraint by developing a multi-criterion labelling pipeline that enables a more consistent resource evaluation.

6.2 Implications

The existing research has also suggested that online platforms may enable users’ radicalization [77, 78] and provide discursive opportunities for right-wing violence [6], however, particular low-level mechanisms through which online platforms enable users’ radicalization remain under-investigated. In this study, we evaluated conversation dynamics, as well as users’ emotional and temporal alignment in particular, as possible factors that may pave the way for users’ more extreme views with regard to immigration policies. However, we have not found any evidence that conversations on Flashback´s subforum generate more negative or more coherent expressions about the topic. On the other hand, this observation partly speaks in favour of the earlier evidence that the emotional tone of online discussions can be best described as stable rather than fluctuating towards particular extremes [36], which highlights the need to explore alternative mechanisms through which the users become indoctrinated into more negative narratives.’

Our observations also call into question the effects of exposure to and consumption of alternative media content, and, in our empirical case, we notice no such effect with respect to the development of conversations, which also has some implications for the debate about the effect of alternative media content on its consumers and the effect of participation in the online fringe political groups. Specifically, we have focused on the distribution of different kinds of content, including mainstream or alternative media, as a particular conversation driver. We found that sharing any type of links has a limited effect only on users’ temporal alignment in the conversations, which is the only case when this effect is significant. Thus, turning back to the explanation offered by Collins’ IRC theory, link sharing can primarily be represented as an element of the collective symbol system [79], which is enabled by Flashback as an online platform, on par with, for example, specific jargon used to talk about immigrants (see e.g. [60]).

Further viewing our results through the prism of Collins’ theoretical approach, we have been able to detect what Collins describes in terms of participants’ entrainment, which, for him, is a necessary condition for a successful ritual [64]. We have evaluated two different ways to approximate such conversation entrainment, namely temporal alignment denoted by conversation pace and emotional alignment denoted by the dominating message sentiments. Based on our analysis, we find that temporal alignment seems to be a more straightforward way to uncover ritual entrainment. In other words, our interpretation is that, in online interaction rituals, synchronizing conversation rhymes is much easier than generating common moods, while the former nevertheless serves as an important condition for the ritual to occur. This observation is supported by Collins’ own remarks regarding the limits to which online interaction rituals can generate emotional energy [62].

One may explain the lack of emotional alignment simply by the fluctuation of user moods in the conversations. Although we know that the majority of users’ messages have, in general, negative appraisals of the immigration topic, this does not exclude the possibility that users’ emotions fluctuate in the level of negativity, or that users may abide by different opinions about what the best immigration policies should be. This may leave some space for users’ deliberations, a possibility that has been suggested by the earlier research [19]. Speaking of alternative media more specifically, one may also want to explore whether immigration-related content in such resources is more controversial, and thus generating more diverse emotional responses and expressions on the forum.

Finally, our results to some extent correspond to those of previous work that primarily focused on conversation length as a successful predictor of ritual’s success [50], which, once again, speaks in favour of users’ temporal alignment as one of the important indicators for the development of online interactions. On a more general level, we conclude that, in our empirical case, online communication has a limited possibility to enable users’ emotional mobilisation and spill over to offline contexts, which relates to the debate about the potential effects of right-wing platforms on users’ radicalisation.