The Swarm versus the Grassroots: places and networks of supporters and opponents of Black Lives Matter on Twitter

ABSTRACT

While activists have effectively used the #blacklivesmatter hashtag to organize protests against police brutality and racism, this success has also drawn out many who use the hashtag to express their opposition. How do supporters of the Movement for Black Lives and their opponents coordinate on Twitter? Drawing on a corpus of 18.5 million tweets, this paper compares coordination among supporters and opponents of #blacklivesmatter in terms of relations and spatialities. We elaborate two different models of coordination: the swarm and the grassroots. Compared to their adversaries, supporters of #blacklivesmatter are more strongly rooted in places and embedded in local relations, suggesting that their online activism builds on grassroots communities. Opponents can be differentiated into two categories. One group consists of conservatives that are rooted in places but in a markedly different geography than supporters; they are more often located outside of major cities and outside of the coastal states. A second group of digitally networked extreme right opponents coordinates more in the fashion of a swarm: they synchronize without being rooted in places or embedded in local relations. These findings demonstrate that movements and countermovements benefit from the affordances of social media in different ways.

KEYWORDS:

• Black Lives Matter
• grassroots
• networks
• social media
• social movements
• swarm

Introduction

In recent years, the hashtag #blacklivesmatter became a powerful demonstration of the potential of digital networking on platforms such as Twitter (Bonilla & Rosa, 2015; Clark, 2014; Freelon et al., 2018; Jackson & Foucault Welles, 2015; van Haperen et al, 2020). The fast and wide diffusion of protest frames catalyzed a movement, spurring thousands of people to fill the streets not only of Ferguson but of cities across the United States to protest police brutality (Freelon et al., 2016; Ray et al., 2017). The hashtag #blacklivesmatter enabled activists to relate specific and local conflicts to broader questions of racial and social justice. However, progressive activists have not been alone in capitalizing on the affordances of social media. A resurgence of online rightwing extremism has benefited from the same digital networking tools (Gallagher et al., 2016; Nagle, 2017; Sobieraj, 2017). Social media can be a space not only for activists, but also their adversaries. While the affordance of visibility can help to amplify the voice of otherwise marginalized activists, this same visibility may also mark them for backlash from opponents. Opponents may use hashtags like #blacklivesmatter to criticize, destabilize and delegitimate the Movement for Black Lives (Colson, 2016; Conover et al., 2011; Nagle, 2017).

Acknowledging that social media are not only conduits for mobilization but also arenas of conflict, recent research has examined framing contests between supporters and opponents (Freelon et al., 2016; Jackson & Foucault Welles, 2015), the formation of collective identities among movements and their opponents (Daniels, 2009; Ray et al., 2017), and the different geographical bases of supporters and detractors (Haffner 2019). We extend this nascent literature on digital conflicts between opponents and supporters of social movements by examining how they form different kinds of networks on the social media platform Twitter. As the hashtag #blacklivesmatter is used by supporters and opponents alike, we are specifically interested in how this movement and its countermovement organize digitally networked action. Drawing on the literatures on digital mobilization, movements and counter-movements, and geographies of social movements, we hypothesize that supporters of Black Lives Matter and their detractors use the same platform, i.e. Twitter, in different ways, forming different kinds of networks to coordinate contentious action.

To test our expectations, we draw on a corpus of 18.5 million Twitter posts containing the hashtag #blacklivesmatter in terms of relations and spatialities. We find differences in the locations and interactions among supporters and opponents of #blacklivesmatter. Supporters are more strongly embedded in local relations and places, suggesting that their online activism builds on grassroots communities. We contrast this with opponents of #blacklivesmatter, who can be differentiated into two groups. One group consists of conservatives with strong mutual relations, but a geography markedly different than supporters: they are more often located outside of major cities and outside of the coastal states. A second group consists of digitally networked extreme right opponents. Their mode of coordination follows not the model of grassroots mobilization but that of a swarm: they coordinate without any territorial embedding.

These findings show that online action and personal networks interact in ways that produce at least two different types of connective action (cf., Bennett & Segerberg, 2012): what we call the swarm and the grassroots models. While countermovements may benefit from the affordances of social media in ways that allow for quick, highly synchronized attacks on adversaries, synchronization will be more difficult to achieve in grassroots movements relying on personal relations, because such decentralized networks introduce more bottle-necks to coordination. In short, the affordances of social media appear to be stacked against the increasingly complex work of organizing social movements.

Conceptualizing coordination at the interface of geography and social media

Social media and digitally networked movements

Digital networks are integral to contemporary protest, prompting questions about the use of such affordances in building social movements (Evans et al., 2017; Tufekci, 2017, van Haperen et al. 2018). According to Bennett and Segerberg, digital networking facilitates rapid expansion through the sharing of personal action frames (Bennett & Segerberg, 2012). As opposed to collective action (cf., Olson, 1965), the sharing of personal action frames does not require overarching collective identities or organizations. Rather, individual views easily traverse personal networks through social media. These are frames ‘inclusive of different personal reasons for contesting a situation that needs to be changed’ (Bennett & Segerberg, 2012, p. 744). The inclusive nature of frames (e.g., ‘We are the 99%’; #metoo; ‘March for Our Lives’) allows a wide range of people to recognize the frame as meaningful and adapt it to make sense of personal experiences (Benford & Snow, 2000; Melucci, 1996). The resulting ‘logic of connective action’ has become particularly salient in a social context of individualization and social fragmentation (Bennett & Segerberg, 2012; Castells, 2007; Della Porta & Tarrow, 2005; McDonald, 2015).

The affordances of social media differ per platform, and each is used in a variety of ways by different actors. We focus on Twitter, which allows both public and anonymous accounts, offering personalized interaction in the form of direct messages, likes, mentions and retweets. The platform allows for follower relations between users, while hashtags can be used to index and easily follow conversations across all users. A number of authors have highlighted the possibilities that Twitter affords to form counterpublics and social movements. For instance, ‘Black Twitter,’ a networked counterpublic, affords Black users with the opportunity to curate news, set agendas, and start conversations on their own terms, jointly developing Black perspectives on current affairs that are absent or sidelined in mainstream news reporting (Graham & Smith 2016; Freelon et al., 2018). Such conversations, in turn, can be picked up in mainstream media, creating a feedback loop with Twitter (Freelon et al., 2018). Twitter further provides affordances for ‘distributed framing’ among different groups activists (Ince et al., 2017) and allows activists to share information in real-time and organize in decentralized fashion (Nummi, Jennings, & Feagin 2019). However, Twitter does not offer a space free of criticism, intimidation, or harassment. Jesse Daniels describes Twitter as an ‘ideal venue’ for white supremacists, with Twitter affording ‘new mechanisms for the furtive spread of propaganda and for vicious harassment with little accountability’ (Daniels 2017, 2–3). Women and people of color are disproportionally targeted, especially if they speak out against sexism or racism (Sobieraj 2020). For these reasons, it is unsurprising that hashtags by social movements like #blacklivesmatter are also used by opponents seeking to criticize, intimidate, or derail activists. Freelon, McIlwain, and Clark (2016) as well as Ince et al. (2017) and Van Haperen et al. (2020) identify different kinds of groups engaging with the hashtag #blacklivesmatter, including a sizeable minority of detractors. We thus turn to the question how supporters and opponents of the Movement for Black Lives organize online.

Two models of coordination: the grassroots and the swarm

To explain how supporters and opponents coordinate online, we suggest two contrasting ideal-typical models: the grassroots and the swarm. The notion of grassroots is elementary to the social movement literature, both as a normative ideal and as empirical focus (Chong, 1991; Della Porta & Tarrow, 2005; McAdam et al., 2001). While digital technology ostensibly allows for interactions without regard for distance, geography structures the networks underlying online interactions, meaning that social movements remain territorially rooted (Routledge, 2017; Van Haperen et al., 2018). Activists continue to derive advantages from local relations, as these provide material and intangible resources such as trust and commitment (Nicholls al 2020). Cities especially can form incubational spaces for marginalized groups to find and connect with each other (Nicholls & Uitermark, 2018). This suggests that coordination remains embedded in local grassroots communities (Castells, 1983). A key question, then, is how local communities form large movements while retaining degrees of autonomy. Bennett and Segerberg suggest such coordination is facilitated by digital networking, through personal action frames that serve as ‘transmission units across trusted social networks’ (Bennett & Segerberg, 2012, p. 755). Thus, local interactions allow for large-scale forms of collective behavior to arise, as frames branch outwards, bridging communities through personal networks without central coordination (Benkler, 2006). The grassroots model of coordination, in short, relies on networked local communities. As a drawback, synchronization may be difficult to achieve because individuals are oriented to local community goals rather than those of the broader movement (Nicholls et al 2016). In fact, connective action is likely to break down if transmission fails through personal networks (Bennett & Segerberg, 2012, p. 754). Grassroots movements therefore have to contend with the challenge of synchronizing action, for example, in reaching consensus between groups, or communications between local chapters (Khan-Cullors, 2016).

Second, some forms of online mobilization might benefit from the affordances of social media in a different way. A model of swarm coordination is based on insights into complex systems (Andersson et al., 2014; Crossley, 2010; Holland, 1992). This literature suggests that interactions between autonomous agents allow for system-wide synchronization, not only in the absence of centralized control, but without reliance on group relations (Bonabeau et al., 1999; Kennedy et al., 2001). In many instances of digital networking, local relationships are not prerequisites for being exposed to personal action frames. Affordances of social media such as visibility and accessibility allow for the engagement with others who are far away (Margetts et al., 2016; Sobré-Denton, 2016). Such interactions may be based more on shared interest than proximity (Rheingold, 2000). Indeed, independence from the restraints of local context may increase the resilience and flexibility of a swarm (Camazine et al., 2001, p. 36; Dorigo & Stutzle, 2004). In short, in a swarm, distributed interactions can generate large-scale synchronization in the absence of locally embedded relations.

To summarize, the key difference between the grassroots and swarm models is the place of the individual in the collective. This is a matter of relations and, by extension, place. The grassroots model refers to engagement with movements through personal networks, suggesting individuals are strongly rooted in places and embedded in local relations. While this benefits the sharing of personal action frames, it makes overall synchronization between communities difficult. The swarm model refers to coordination in the absence of agents’ local embeddedness. While less conducive to the sharing of personal action frames, overall synchronization is more feasible because it does not rely on strong local relations. This may be sufficient to coordinate reactionary attacks on social movements.

Digitally networked countermovements

The literature on social media activism has identified different models of coordination, but it provides limited insights into whether countermovements coordinate in the same way as the movements they oppose. However, social movement scholars have long called attention to adversarial dynamics, suggesting that these are essential to understanding the development of social movements (Lind & Stepan-Norris, 2011; Meyer & Staggenborg, 1996; Zald 1979; Zald & Useem, 1987). Countermovements are a network of actors that emerge in response to a social movement and can result in reactive interactions between the two sides (Tilly, 1975; Zald & Useem, 1987). Meyer and Staggenborg add that, ‘The emergence of one movement may precede that of its opponent and, early in such a conflict, it is appropriate to speak of the original movement and its countermovement. Indeed, some conflicts never progress beyond some preliminary challenges by an emerging countermovement’ (Meyer & Staggenborg, 1996, p. 1632). Most of these ‘emerging countermovements’ aim to disrupt and delegitimate the primary movement through loosely coordinated criticisms and attacks. When countermovements mature, they expand their goals and repertoires, accumulate more resources, and enhance their infrastructure and methods of coordination. They transition from emerging countermovement to full ‘opposing movements’, like the pro-life movement (ibid.).

We can draw upon the literature about countermovements to better understand how movements and their opponents use social media to coordinate their mobilizations. As noted above, this literature suggests that early or ‘emerging countermovements’ aim to stop and breakdown the mobilizations of their opponents by criticizing, antagonizing, and delegitimating their adversaries. Online, this can take the form of diffusing personal action frames through coordinated tactics, for instance, by attacking opponents with ‘tweet storms’. Personal relationships of trust are not a prerequisite for such relatively straightforward synchronization, which makes the swarm sufficient for a form of connective action similar to, but distinct from, organizing.

Fully developed grassroots movements have broader goals (i.e., mobilize resources, influence policy makers, organize local events, build local communities, among other things) and adopt a more diverse repertoire to achieve them. They are more likely to organize community events, street protests, mount civil disobedience campaigns, and lobby elected officials. The online sharing of personal action frames, amplifying issues, and exchanging information become increasingly intertwined with direct action on the ground. As the complexity of such interdepencies increases, social movements become increasingly reliant on grassroots coordination.

Expectations

Both models of coordination outlined above are ideal-typical and we should not expect movements and their opponents to operate fully according to either the grassroots or swarm model; activists may combine elements from both. We do hypothesize that, in general, supporters of social movements are more strongly embedded in local communities than their reactive opponents in emerging countermovements. Public activism takes courage, particularly for marginalized groups, requiring solidarity and comradery. Public protest and civil disobedience also require deliberations among activists and intimate knowledge of geography and political conditions. The work of building movements, and later the institutionalization of a movement, requires long-term commitment and trusted relations. By contrast, ad hoc and ephemeral relations may suffice for the counter-movement’s reactionary aims of criticizing or attacking another movement’s work. Opponents in emerging countermovements appear more focused on disrupting and blocking social movements by criticizing and harassing them. The circulation of frames and memes, and mobilizations such as ‘tweet storms’ require coordination, but social media platforms provide activists with sufficient capabilities to coordinate these tasks (Tufekci, 2017). Consequently, we expect that opponents in emerging countermovements are more likely to swarm than engage in grassroots activities.

Data and methods

The hashtag #blacklivesmatter catalyzed the Movement for Black Lives, a coalition of ‘more than 50 organizations representing thousands of Black people from across the country’ (The Movement for Black Lives, 2019). One of those organizations is the Black Lives Matter Global Network (colloquially: ‘Black Lives Matter’), a ‘chapter-based, member-led organization whose mission is to build local power and to intervene in violence inflicted on Black communities by the state and vigilantes’ (Black Lives Matter,2019). As such, we refer to #blacklivesmatter, from among other hashtags such as #Ferguson and #icantbreathe, as an instance of digital networking, noting that it captures only part of a multifaceted social movement described in more detail elsewhere (Bonilla & Rosa, 2015; Freelon et al., 2016; Ince et al., 2017; Jackson & Foucault Welles, 2015; Khan-Cullors & Bandele, 2018; Lebron, 2017).

To compare the online coordination of supporters and opponents of Black Lives Matter, we draw on data collected from Twitter. Tweets were collected through the streaming API, constructing a sample of tweets with the hashtag #blacklivesmatter¹ in the 30 months between 15 June 2015 and 15 December 2017. Due to practical restrictions on the use of historical Twitter data, earlier periods could not be included (Twitter Inc, 2014). However, using the streaming API for an extended period allowed for analysis of sustained engagement over time (Bennett & Segerberg, 2012, pp. 760–761; van Haperen et al., 2018). Covering multiple years also ensured a broad array of events and types of engagement are included in the dataset.

Identifying opponents and supporters

To identify opponents and supporters of #blacklivesmatter, we develop a three-tiered, semi-supervised method for content analysis. First, we draw and code a training sample from all tweets in the dataset. To make manual coding feasible while ensuring representativeness despite variable traffic volumes over time, this sample is constructed by randomly selecting 1% of the daily volume for each day in the data set. We then code this training set manually. Tweets are coded as positive or negative towards Black Lives Matter. In the interest of accuracy, only unambiguous sentiments are assigned to either category.

Second, the training sample informs classification of the full dataset. From the two subsets of positive and negative tweets we derive a lexicon of n-grams, i.e. combinations of characters typical for positive or negative tweets. Cross-referencing these n-grams, all tweets in the dataset are coded as either positive, negative, or unclear. This procedure yields 5.6% positive tweets (n = 1,027,554) and 1.9% negative tweets (n = 360,473). Given the stringent criteria for assigning positive or negative codes, many tweets in the dataset were assigned to the residual category (n = 17,113,758 or 92.5%).

Third, these tweets inform classification of opponents and supporters. Accounts are assigned either category when at least two thirds of their coded tweets are negative or positive. Only accounts with three or more coded tweets are included, in order to minimize misclassification from an incidental tweet (41,527 out of 4,240,861 unique users or 1%). Following this procedure, we identify 31,271 users as supporters and 4,791 users as opponents. To validate our procedure and assess degrees of automation or false representations, the self-reported biographies and account activity of both supporters and opponents are examined manually (Varol et al., 2017).

Since we use several stringent filters (users who employ the #blacklivesmatter hashtag; users who have sent at least three unambiguously supportive or antagonistic tweets), this sample represents a sizeable but relatively small subset of the people talking about #blacklivesmatter on Twitter. However, since we focus on comparing movements and countermovements, for our purposes precision is more important than recall, i.e. we prefer to have reasonably reliable data on users’ dispositions over covering as much Twitter activity as possible.

Examining rootedness in place and embeddedness in local relations

We expect the grassroots and swarm models have different levels of local embeddedness. We study this in two ways. First, we examine the degree to which users are rooted in places in terms of self-reported geographical location. Twitter offers the option to geotag posts, but users rarely do this (Jurgens et al., 2021). We therefore also extract self-reported information from Twitter bios to estimate locations. This biographical data comes with limitations. It is self-reported, not always accurate, and non-uniform. To deal with this, results are aggregated and manually examined (e.g., ‘Big Apple’ and Brooklyn is New York City, Long Beach is Los Angeles). The degree to which a user is rooted in place is measured by comparing the frequency of locations for both opponents and proponents of #blacklivesmatter. We interpret self-reported locations as a signal that the location is significant to the user. We compare various levels of specificity: coordinates, state, city, and non-pertinent locations. We interpret the optional geotagging functionality as most specific, as users have to opt in and coordinates are then programmatically recorded rather than self-reported. Self-reported state and city locations are deemed relevant, unless manually coded as non-pertinent.² To do so, we sample and manually code 600 locations, which are then compared to the full dataset, reported as the percentage of users with a (non-)pertinent location. Second, we study whether users are embedded in local relations by measuring the degree to which geographical proximity predicts relations among both supporters and opponents of #blacklivesmatter. When a shared geographical location predicts a tie, we interpret this as an indication of geographical embeddedness.

Examining synchronization among opponents and supporters

We expect the grassroots and swarm models have different levels of synchronization. The level of synchronization among supporters and opponents is measured in two ways: as simultaneous activity and as a mutual focus of attention (Collins, 2004). We study synchronization of activity among groups by comparing the peak times of post volumes on Twitter. A high level of simultaneous tweeting activity suggests a stronger level of synchronization in a group. We study the synchronization of a shared focus of attention by examining whether many users focus on the same target simultaneously (Collins, 2004). More specifically, we count the daily number of supporters and opponents as well as the targets they focus on, and then divide these numbers. This gives the average number of users from groups who simultaneously focus their attention on the same targets.

Ethical concerns and validity

Data is collected as per the Twitter terms of service stipulating uses of data and user consent. To address concerns about risk of harm to users, particularly in the context of individuals engaged in contentious actions, we subscribe to more demanding research principles (Moreno et al., 2013). All identifiable personal information is anonymized and aggregated. We note that bots are an integral part of activity faced by activists on Twitter, so rather than filtering out these accounts we opt for inclusion in our analysis. Nevertheless, our efforts to identify bots suggest that they are more prominent among adversaries of #blacklivesmatter. In general, bots represent only a small percentage of accounts³

Results

In the period between 15 June 2015 and 15 December 2017, we collected 18,501,785 tweets from 4,240,861 unique users. We first present the overall network. Next, we compare the local embeddedness of supporters and opponents. Then, levels of synchronization are compared.

Overall network of supporters and opponents

The interactions on Twitter result in a network structure of both supporters and opponents. Figure 1 shows how these networks are organized.

Figure 1. Interactions between supporters and opponents.

Colors represent how strongly someone supports (blue, n = 31,271) or opposes (red, n = 4,791); the darker red nodes are accounts with more negative content. To aid visualization, users with fewer than five relations are omitted.

Groupings in the network are based on interactions, and result from particular interests shared with others.⁴ Among supporters, there are many different such groups. For example, supporters in one group are avid hip-hop fans and artists, who mostly tweet about music and occasionally retweet news about #blacklivesmatter. Among opponents, there are roughly two groups, which we identified using content analysis: the first (left in Figure 2) can be characterized as conservative, the second as alt-right.⁵

Figure 2. Synchronized activity among supporters and opponents of #blacklivesmatter.

Daily volume of tweets sent, by day, from opponents alt-right (red), conservatives (orange) and supporters (blue). Point size indicates the number of tweets.

The first group of opponents is composed of individuals with conservative political views. Oftentimes, these accounts are not anonymous. Descriptions that stand out in user biographies in this group are related to family, veteran affairs, the second amendment, bible, and Donald Trump. Although not univocally supportive of the President, slogans such as ‘Make America Great Again’ are salient among these users. Twitter users in this group often criticize mainstream media viewpoints and share links to alternative news sources. One day, the topic may be Hillary Clinton, unemployment among veterans the next. They engage with the hashtag #blacklivesmatter in the same way: something triggers attention, which leads to the group focusing an intense but brief gaze. This happens, for example, by sharing a meme about crime statistics, reinforcing group membership and ideological views related to #blacklivesmatter.

The second group of opponents can be characterized as ‘alt right’, and is a source of extreme content. The thematic scope in this group is narrower, less news-driven, and more extreme than the previously described conservative group. Content and images can be disturbing; examples include images of KKK lynch mobs and idealizations of Nazism. Most accounts do not disclose pertinent information. Manual verification of user names and tweeting activity brought up slight name variations, suggesting multiple accounts operated as one. Many of the accounts in this group have been suspended by Twitter or deleted for violating terms of service soon after being included in data collection for this study.

Local embeddedness of supporters and opponents

We hypothesized that supporters of Black Lives Matter are more likely to be rooted in places and embedded in local relations than adversaries of #blacklivesmatter. First, we compare self-reported locations for supporters and opponents (Table 1).

Table 1. Local embeddedness: specificity of reported location.

	Group	Population (Sample)	City	State	Non-pertinent
Control		4,240,861	10.7%	4.9%	1.7%
		(1000)	(0.98)	(0.68)	(0.41)
Supporters		89,203	11.9%	4.3%	1.8%
		(1000)	(1.02)	(0.64)	(0.42)
Opponents	Conservative	8,703	8.1%	6.3%	5.9%
		(1000)	(0.86) *	(0.77)	(0.75) ***
	Alt-right	8,477 (1000)	7.2% (0.82) ***	7.6% (0.84) **	5.6% (0.73) ***

Counted are the number of self-reported locations among supporters or opponents. Group membership is determined by an estimated measure of support on the basis of at least three unambiguously coded tweets. Significance compared to a randomly drawn control sample, standard errors in parentheses; * p < .05. ** p < .01. *** p < .001.

Supporters and opponents tend to disclose different kinds of geographical information. Opt-in geotagging with coordinates does not occur in any of the samples.⁶ Supporters tend to be more specific when they self-report locations and include fewer non-pertinent locations. Among opponents are more likely to self-report states as their locations, and fictional locations are more common, particularly among the alt-right cluster.

Our interpretation of these findings is that supporters are more transparent about the specifics of their location, and thus more locally embedded than opponents. The fact that opponents less commonly report city-level information may suggest less transparency about specific locations, or a greater tendency to identify with a home state rather than city. There may also be differences in urban and rural concentrations of supporters and opponents. We verified a sample of account information to assess such a selection bias, and the reported percentages are based on manual coding of this sample.

Second, we test how strongly supporters and opponents are embedded in local relations. We compare geographical proximity and tie formation among both supporters and opponents. Overall, given the large number of locations in our dataset, the percentage of ties within the same city is low (Table 2).

Table 2. Likelihood of online interaction in the same location.

	Empirical same location	Random same location
Supporters	563 of 8,431 (6.68%)	86 of 8,431 (1.02%)
Conservative	8 of 580 (1.38%)	7 of 580 (1.21%)
Alt-right	8 of 418 (1.91%)	6 of 418 (1.46%)

Reported are the number of network ties where both the source and target share a location.⁷ A bootstrap baseline for comparison is provided by randomly reassigning locations among the same number of nodes in an identical network configuration.

Supporters are six times more likely to interact online with other supporters in the same city than might be expected from randomized network configurations. By contrast, tie formation among opponents in the same location is about as likely as might be expected from randomized configurations. Note that there are many more supporters than opponents, meaning the range of locations is much larger. Despite this, supporters not only appear to be more rooted in as discussed above; they are also more likely to interact with others in the same location. These findings suggest that the online interactions of supporters are more strongly embedded in local relations than those of opponents.

Synchronization among supporters and opponents

We hypothesized that there would be more synchronization among opponents of #blacklivesmatter than among supporters. Synchronization is examined in two ways: as tweeting activity and as a shared focus of attention. We study synchronization of activity among groups by comparing time series of tweet volumes. Increased simultaneous tweeting activity indicates more synchronization (Figure 2).

Overall, both supporters and opponents are synchronized only at certain times. Such synchronized activity in large numbers occurs more often among opponents than among supporters. Among supporters, the range of daily volumes falls between 1 and 3786 tweets, the mean is 113.2, the median 49 and the standard deviation is 254.3. Among conservatives, the range of daily volumes falls between 1 and 177 tweets, the mean is 15.4, the median is 9, and the standard deviation is 20.7. Among alt-right, the range of daily volumes falls between 7 and 3233 tweets, the mean is 183.3, the median is 109, and the standard deviation is 284.2.⁸

Different dynamics are at the root of synchronized moments. The graph depicts proportions of overall activity, showing how evenly activity is spread out over the period in the different groups. Supporters tend to engage ongoing conversations in small groups, generating a baseline of daily activity. Because of this, as the grassroots model suggests, synchronization among large numbers of supporters is rare, but it does occur occasionally. For instance, this happens in July 2015, with the tragic death and funeral of Sandra Bland, instances of police brutality, or events like the Academy Awards ceremony. For opponents, synchronization occurs, for instance, by the circulation of links to alternative news sources. Among the alt-right, there is little activity related to #blacklivesmatter in between such moments of synchronization.

We also examine synchronized attention. The degree to which Twitter users synchronize their attention is measured as the average number of users from each group focusing on the same target simultaneously (Figure 3). Intuitively, a high average value suggests a large group of users directs attention to a relatively small number of targets.

Figure 3. Synchronized attention among supporters and opponents of #blacklivesmatter.

Average number of users from group focusing on targets, by day, from conservatives (orange),alt-right (red) and supporters (blue). Values are normalized for each group to membership as y = (x–x_min)/(x_max – x_min), the absolute sum of which is shown as point size.

Synchronization of attention is generally limited. This is illustrated by the low values throughout the period, ranging roughly between 0 and 0.25, with mean normalized scores: 0.08 for supporters, 0.06 for conservatives, and 0.11 for the alt right. For instance, on 9 October 2017 a fairly large number of supporters (369) align in support of Kaepernick.

The level of focused attention is particularly low among conservative opponents, and a highest among the alt-right. We also note there is a degree of sequencing between opponents and supporters. Peak activity among alt-right opponents follows after, never precedes, supporter activity. While no expectations were formulated about coordination between opposing groups, this suggests more reactionary activity among alt-right opponents. One example where synchronization is achieved among alt-right opponents includes an incident in June 2016 when activist DeRay Mckesson’s social media accounts are hacked: messages allegedly obtained in those hacks are circulated widely by opponents in efforts to defame the prominent activist.⁹

Discussion and conclusions

To describe how digitally networked movements and countermovements coordinate, we draw on a dataset of 18.5 million tweets, comparing supporters and opponents of #blacklivesmatter in terms of relations and spatialities. To explain these differences, we elaborate two ideal-typical models of coordination: the grassroots and the swarm. The grassroots model refers to engagement with movements through personal networks, suggesting individuals are strongly embedded in local communities and places. Relationships within groups are strong (kinship, friendships, comrades, affinity groups, et cetera) while tools like Twitter facilitate synchronized action between such groups (Bennett & Segerberg, 2012; Granovetter, 1973; Tarrow & McAdam, 2005). By contrast, the premise of the swarm model is that synchronization can emerge from individual actions in the absence of strongly connected local communities or central coordination (Centola et al., 2007; Holland, 1992; Seguin, 2017). Noting that these models of coordination are ideal-typical, and the oppositional binary between supporters and opponents a generalization of diverse types of users (Lim, 2012, 2017), we expected that the grassroots model may be most suitable for the development of personal networks necessary for building social movements, while the swarm model may be ideally suited to reactions against movements.

Findings by and large confirm our expectations. With regard to local communities, supporters are more strongly embedded in place: supporters appear more forthcoming about their home location in their Twitter bios and geographical proximity drastically increases the probability of a relationship. By contrast, we find that opponents and interactions among them are less grounded in specific locations. Together, these findings indicate that supporters are more strongly embedded in local communities, confirming expectations of the grassroots model. Opponents, and the alt right groups in particular, are less rooted in specific localities, which confirms expectations of the swarm model. We should note, though, that Twitter activity for all groups, and for opponents as well as supporters, is largely deterritorialized. People generally do not communicate through Twitter to others in their localities and only rarely geotag their tweets. What is striking is not the degree to which Twitter activity is grounded in localities but the difference between groups. Considering that researchers consistently find that social movements networks are rooted in places (Borge-Holthoefer et al., 2014; van Haperen et al., 2018), these striking differences between different groups call attention to the need for careful differentiation of types of engagement with digitally networked movements.

Supporters of #blacklivesmatter are more strongly embedded in local relations and places, suggesting their online activism builds on local community relations, often rooted in specific cities. Opponents have very different relations and geographies. One group is, to a degree, rooted in localities but with a different geography than supporters, typically based in red states. We describe this group as conservative opponents. A second group of opponents is described as extreme right and is the source of the most extreme content. We identified a number of bad-faith actors and automated accounts, but note that these represent a comparatively small contingent of opponents’ activity. Moreover, we realize these are part of the daily reality faced by activists online. The swarm, as a whole, is highly synchronized without reliance on embeddedness in local communities.

Based on these findings, we suggest that online action and personal networks interact in ways that produce at least two different types of connective action: the swarm and the grassroots models. The work of building movements, and later the institutionalization of a movement, requires long-term commitment and trusted relations. By contrast, ad hoc and ephemeral relations may suffice for the counter-movement’s reactionary aims of criticizing or attacking another movement’s work. As noted by Tufekci (2017), such activity may be less likely to develop strong relationships compared to the complex work of organizing.

On the one hand, strong relationships may be more conducive to the formation of collective frames (as trusted friends, families and comrades seem more likely to come to terms with our shared understandings), while diffusion beyond such clusters would be hampered in a network with high modularity. Thus, counterintuitively, the personal relations prevalent in the grassroots model would be more conducive to collective action frames than what Bennett and Segerberg call personal action frames (2012). On the other hand, a network configuration relying less on strong ties among local clusters may have fewer bottlenecks, and thus be more conducive to the diffusion of personal action frames.

While we focused on supporters and adversaries of the hashtag #blacklivesmatter on Twitter specifically, we think similar dynamics between movements and countermovements occur in other instances of digitally networked action and on other online platforms. While the affordances of digital networking are leveraged by grassroots movements and their adversaries alike, they operate according to different kinds of coordination online. Given the importance of social media to social movements today, and given how pervasive forms of antagonism are online, questions of digital countermovement dynamics likely remain relevant for the foreseeable future.

Acknowledgments

The authors received no specific grants for this work. We are grateful for the valuable comments of the anonymous reviewers.

Disclosure statement

The authors declare no financial interest arising from the direct application of this research.

Additional information

Funding

The authors declare no financial interest arising from the direct application of this research.

Notes on contributors

Sander van Haperen

Sander van Haperen is Assistant Professor at the Erasmus School of Health Policy & Management. He studies governance on the intersections of collective action, complexity, and digital technology, with a focus on conflict resolution and leadership. To that end, he develops a critical relational approach, combining qualitative inquiry with geographic analysis and computational methods such as network analysis, image recognition, and natural language processing. Following a dissertation examining the role of social media in the development of social movements such as Black Lives Matter, he published in a variety of journals about social movements and governance. He is currently studying governance in health care to shed light on the role of leadership in complex networks.

Justus Uitermark

Justus Uitermark is Professor of Urban Geography and Sociology at the University of Amsterdam. His research is located at the intersection of urban studies and political sociology. From 2012 until 2016, he was the Gradus Hendriks Professor in Community Development at Erasmus University Rotterdam. In September 2010, he defended his PhD-thesis Dynamics of Power in Dutch Integration Politics (cum laude). Uitermark has published books and articles on a variety of subjects, including gentrification, state theory, social movements and urban governance.

Walter Nicholls

Walter Nicholls is Associate Professor of Urban Planning and Public Policy at the University of California, Irvine. He is interested in how cities are places that help empower (or not) marginalized groups. This interest led to the study of immigrant rights struggles in various countries including the United States, France, and the Netherlands. By closely looking into immigrant mobilizations in such different countries, certain cities appear to play important roles in politicizing marginalized people and providing the support needed for intensive political mobilizations. He plans to continue this research in the coming years by examining how urban planning can be used as a tool for building more socially just cities.

Notes

1. Including variations and misnomers: #blacklivesmatters, #blacklivematter, #blacklivematters, #blacklifematters, #blacklifematter.

2. Non-pertinent locations are shorter than two characters, or manually coded from a sample as either country-level or fictional locale, for example: ‘earth’, ‘woke’, ‘a galaxy far far away’.

3. We assessed the presence and activity of accounts that are operated through bots or by malfeasant actors (Varol et al., 2017). Recent reports have specifically addressed the role of Twitter accounts operated or directed by Russian intelligence to steer and escalate political conflicts. We cross-reference our dataset with the list of accounts released by Twitter in relation to investigations of campaign meddling used in official hearings (United States House of Representatives,2018). Of the 3,841 accounts in that release, 55 appear in our data (of 4,240,861 unique accounts originating tweets in our data). Together, these 55 accounts sent 2,993 of the tweets in our dataset (of 18,501,785 tweets). It is not unlikely that additional accounts or tweets in our data originate from inauthentic sources, but remain as of yet unidentified. In addition to the accounts identified in the release, we identified another 1,039 suspect accounts. To do so, we flagged accounts with ties to the original seed accounts and checked high tweet volumes. Interestingly, of the 55 accounts reported in the official documents, our procedure yielded 11 supporters, reflecting disinformation strategies to influence multiple sides of public debates (Stewart et al., 2017).

4. The layout algorithm used for the visualization in Gephi was ForceAtlas2, in which groupings are force-directed based primarily on degree (Jacomy et al., 2014). The striking alignment between spatiality and measure of support was not adjusted manually.

5. For reference we compared this structure with a random walk-trap algorithm and confirmed a strong overlap between group membership and cluster membership, that is: 76% of users that were identified independent of community structure as opponent or supporter turned out to share the same cluster algorithmically.

6. 0.06% of users in our entire dataset use this option.

7. Opting for accuracy over quantity, the following conditions were applied, explaining the relatively small numbers. Only ties are included of users unambiguously classified as supporter or opponent, which includes having sent at least three tweets. Only ties are included of users with unambiguous, pertinent locations. Ties are included only if the target also occurs in the dataset as a source, because otherwise a location could not be determined.

8. Note that the degree distributions follow a power-law distribution, with an alpha of 2.14 (Kolmogorov-Smirnov p = 0.67, xmin = 6) among supporters, 2.57 (KS.p = 0.95, xmin = 4) among conservatives, and 2.25 (KS.p = 0.73, xmin = 5) among the alt right (Clauset et al., 2009).

9. Circumvention of Mckesson’s two-factor authentication was achieved through sim card fraud. The hack itself became apparent from content posted on his account. Subsequently, screenshots allegedly taken from his (otherwise private) direct messages were circulated, leading to contention about their credibility between supporters and opponents. For further detail, see, (Dreyfuss, 2016).