State-aligned trolling in Iran and the double-edged affordances of Instagram

Repertoires of information control and suppression in post-revolutionary Iran

Any authoritarian regime has a unique history and utilizes context-specific instruments to remain in power. As such, Iran is no stranger to suppressing voices of dissent. Simultaneously, these regimes often learn and “adapt their own repertoires of suppression in response to developments on the ground, regionally and at home” (Heydemann and Leenders, 2011: 649). In addition, authoritarian regimes often learn from the experiences elsewhere and cooperate with one another. For example, Iran’s largest telecommunication company purchased surveillance equipment from ZTE, a Chinese company, in 2010 (Vatanka, 2015).

For the purpose of this study, we address current strategies within three main categories: censorship in publishing (Small Media, 2015), Internet filtering (Aryan et al., 2013), and harassment and persecution (Golkar, 2013).

Censorship in publishing

Censorship in Iran is not exclusive to the post-revolutionary era. From monarchs to clerics, politicians have continuously attempted to control the flow of information to consolidate power (Small Media, 2015). Iran’s current Press Law and other guidelines (e.g., Supreme Council of Cultural Revolution, 1988) do not tolerate (1) any materials “undermining the political system” or attempts at “infiltrating the pillars of the Islamic Republic” and (2) “unethical artistic products” with “adverse ethical effects” that promote “sexual freedom and indecency” and materials that lead the youth “astray to corruption” (Mollanazar, 2011). Those involved in any form of expression, from journalism to art, are therefore no stranger to state-imposed “redlines.”

Internet filtering

As the Internet penetration rate grew in Iran, it became one of the dominant sources of information and a medium for publishers to reach a much broader audience. This led to an extension of the control repertoire over cyberspace. As a result, new regulations on digital communications were introduced (e.g. Cyber Crimes Law passed in 2010; ARTICLE 19, 2017) and several regulatory and monitoring bodies were established (Bowen, 2015; Center for Human Rights in Iran, 2015). With the emergence of social networking platforms, Iran’s stance on censorship and Internet access has become further complicated. At politically and socially sensitive momentums, such tension has repeatedly surfaced (e.g. after the disputed presidential elections in 2009: Burns and Eltham, 2009; Morozov, 2011). Utilizing social media to organize street protests was an unprecedented form of resistance against the political establishment and immediately led to Facebook and Twitter being banned in Iran. Despite various restrictions on access to online content and platforms, Instagram curiously remains unblocked (Khodabakhshi, 2017) and continues to grow in popularity in unprecedented ways. ² Given these circumstances, this study will investigate how Instagram is utilized to disseminate harassment and propagate intimidating messages against the Iranian opposition and other vulnerable communities.

Harassment and intimidation of civil society and other dissidents

In addition to the limitations on access to information, politically charged intrusion campaigns against the opposition and different forms of surveillance remain prevalent (Guarnieri and Anderson, 2016). Political opponents and activists, journalists (Feinstein et al., 2016), members of civil society, religious minorities (Baha’i International Community, 2011), even artists (ARTICLE 19, 2015), writers, and poets have remained primary targets of intimidation, castigation, arrest, and, in some cases, execution (Soleimani, 2016). In recent years, journalists affiliated with international media, diaspora artists, and activists have endured recurring cyberattacks (Anderson and Sadjadpour, 2018) and some form of harassment. These new forms of abuse are utilized to deter these influential communities from pursuing professional activities that are perceived as “anti-revolutionary” by the Islamic Republic. Like other forms of information control, the extent and duration of harassment typically correlate with the political and social contours (for example, during election periods: Anderson, 2013).

While numerous studies have previously analyzed the role of social media during protests in Iran (e.g. Burns and Eltham, 2009; Morozov, 2011), very few have investigated the repertoires of suppression on an individual level (cf. Feinstein et al., 2016) and their potential impact on online activism in authoritarian regimes. To address this gap, this study investigates strategic communications that often emerge in tandem with, or following, intrusion campaigns or other forms of abusive treatment of specific targets. These in particular include high-profile members of Iranian diaspora pursuing activism in different domains such as sexual minority and women’s rights, thereby maintaining remarkable capacity to reach, influence, and even mobilize large audiences inside Iran (see Part 1: Qualitative Interviews and Index 1).

Methodology and data

Nine out of 18 interviewees maintain public profiles on Instagram. These individuals agreed to be publicly identified and their Instagram accounts’ data be used for this analysis. In addition, we identified three other individuals whose work and “celebrity” status expose them to similar threats and aggression on social media platforms. Eventually, a total of 12 accounts were selected for our final analysis, all of whom represent civil society and public figures of Iranian diaspora. These individuals’ work is generally not welcome by the Islamic Republic, increases the possibility of state-affiliated harassment campaigns against them. To answer the research questions, in March 2017, we downloaded the information of all the publicly available posts (pictures and videos) that the 12 accounts in our sample had uploaded (n = 7893). ⁴ We then downloaded all comments that these posts had received. We trimmed the data set by focusing on users that had commented on at least two different sample accounts. This provided a final data set of 1,243,318 comments by 96,283 unique users. ⁵

As part of a multi-step process, we first identified the topics that these commenters had discussed for RQ1. Due to the large corpus of comments, a manual content analysis was not feasible. Therefore, we conducted an automated analysis to identify topics that contained harmful speech. This was possible through topic modeling, a statistical technique that identifies co-occurrence of lexical topics in a large corpus of documents (Blei et al., 2003; Roberts et al., 2014). In any topic model, two general parameters are estimated: the probability of each word belonging to a specific topic (per-topic word distribution) and the probability of the topics for each document (per-document topic distribution). The probabilities are usually normalized for each document and sum up to 1. We used the R package STM (Roberts et al., 2014) to calculate the topic model.

A challenge, however, is that for topic modeling to render reliable results, some information in documents is required. Single comments, in particular, are typically too short and consist only of a few words. For every unique user, we created one document that included all comments written by the same user (n = 96,283 documents). This is a common approach to addressing said challenge and often results in more comprehensive topic models (Hong and Davison, 2010). We removed punctuation marks and numbers as well as all stop words from the text corpus using different Farsi stop word lists (Kharazi, 2017; Taghva et al., 2003). As the use of emojis in Instagram commenting has become common, we included emojis in our analysis as single words. The number of topics (k) has to be defined before the analysis of topic models. We first identified the statistically optimum number of topics and checked all solutions between 10 and 100. We settled to use a topic model with 60 topics (Murzintcev, 2016).

To identify the temporal dimension of politically charged hate speech for RQ2, we identified pictures from our sample that had received the largest number of comments that contained these topics. Since we used a relatively high number of topics, there are more topics with a lower probability in our data set. Therefore, we decided to use a low threshold of at least 0.1. Users that have at least a probability of .1 for one of the hate speech topics were classified as disseminators of harmful speech. ⁶ For each user, we then combined the probability of two vulgar harmful topics and three politically charged hate speech topics. If a user had a probability of over .1 for one of the two forms of harmful speech, we classified him or her under the corresponding category. We applied the same threshold to the other topics. While the exact publication date of comments on Instagram is not available, we settled to use the publication date of the posts that had been commented on as a date.

For RQ3, we analyzed the co-commenting structure of the 96,283 unique users through network analysis. For each Instagram post, we confirmed the users who wrote a comment. In this network, every node is a user and every edge indicates that two users have at least commented once on the same post. The edge-weight shows the number of pictures that these users have commented on together. We then identified six distinctive co-commenting groups of users with the Louvain algorithm (Blondel et al., 2008). ⁷ The co-occurrence edge lists were created with an R script created by the authors. We then created an undirected graph and used the Louvain algorithm for the community detection with the R package igraph. With no additional information, defining these communities was rather convoluted. Therefore, we combined the results of the topic model with the information about co-commenting communities.

Finally, we downloaded for RQ4 all information for a subset of users who appear in a co-commenting community with higher frequency of disseminating politically charged harmful speech to detect any relation in regard to who they followed. This allowed us to further analyze how these users were connected with one another, whether they shared the same interests and ideology, and thus followed the same accounts (e.g. political leaders, singers, or football players).

RQ1: topics in the comments

Most of these 60 topics reflect typical communications on Instagram. Two topics consist almost entirely of emoji. The majority of topics reflect the general themes of what the interviewees publish on Instagram. These are typical Instagram discussions in which users may argue with one another; express reverence for a favorite celebrity, that is, one of the interviewees; or broadly talk about politics, religion, and/or society.

However, five topics could be classified as potential harmful speech to which various users contributed. Two of these used derogatory and explicitly sexist language. Particularly, the names of two women interviewees appeared with high probability in one of these topics. The remaining three topics were far more implicit in their lexicon and used less directly identifiable derogatory terms (see Figure 3 left). These topics mainly included nationalistic and religious vocabulary. One centered on the moral behavior of women (e.g. “promiscuous” and “engaged in wedlock”). The second includes explicit words referring to the sexual behavior of women (e.g. “prostitute,” “whore”). The third topic mainly includes nationalistic language (e.g. homeland, nation) and some insults (e.g. “dirty,” “treasonous dog”). Separately, we also identified one counter-speech ⁸ topic and a loosely defined “religion and society” topic, covering politics, religious debates, and social debates (see Figure 3 right).

OPEN IN VIEWER

Figure 3.

One of the harmful speech topics specifically targeting women (left). General religious topic (right). The size of the words indicates the probability of the word in the topic.

RQ2: the temporal dimension of harmful speech

We identified two forms of temporal patterns that are illustrated in the following two cases:

OPEN IN VIEWER

Figure 4.

Timeline of different types of harassment on the Instagram account of Sadaf Taherian and Aram Bolandpaz. Dashed line (above) indicates the moment Taherian’s unveiled pictures went viral. Dashed line (below) indicates the date Bolandpaz published the video of her coming out on Instagram.

RQ3: co-commenting patterns

With the help of the Louvain algorithm (Blondel et al., 2008), we identified six distinctive co-commenting communities. With no additional information, defining these communities was rather convoluted. Therefore, we combined the results of the topic model with the information about co-commenting communities. For each community, we calculated the average probability for the relevant harmful speech as well as a few other distinctive topics (see Figure 5). One particularly small community consisted of 1424 users and indicated the highest probability for hate speech and “religion and society” topics. This was a clear distinction from all other groups who had a much higher probability for the topics consisting of positive emojis, for example. Yet, the same community of 1424 users had a notable presence in counter speech topic. This can be explained through the methodological approach: it is highly likely that the content subject to hate speech simultaneously attracted support of the fan base of our interviewees (see more in RQ4: follower networks and backgrounds of users).

OPEN IN VIEWER

Figure 5.

Heatmap with the probabilities for the most relevant topics in the co-commenting communities.

RQ4: follower networks and background of users

To obtain more information about the relatively small community who disseminated political hate speech, we downloaded the follower network of all users in this community. This allowed for distinguishing these users and their interests and backgrounds more clearly. 17% of the accounts in this community had been deleted, and 53% were not accessible due to their privacy settings. Therefore, we could only analyze the follower relations of 429 accounts (30%) of said community.

The internal follower relations affirm a clear divide between two groups of users in this community (see Figure 6). The content posted by the most popular accounts in one group suggests conservative and religious interests of account owners. For example, one of the most prominent accounts in this community mainly posts pictures with negative references to the political opposition in Iran. ⁹ However, there are also popular accounts with religious usernames in this community that mostly posts pictures about their religious and political ideology. Notably, one of these users commented on three different accounts from our sample, using derogatory language, for example, “the cunt who lies,” “you and your goddamn family deserve hell.” The second group stood contrary to the first one. They publish content in support of the opposition and post critical content of Iran’s political establishment.

OPEN IN VIEWER

Figure 6.

Left: Internal follower network of the co-commenting community. Religious accounts are on the left side (yellow), liberal accounts on the right side (orange). Right: Follower network including the additional 200 most followed accounts. The yellow community shows strong religious orientation; the blue mainly consists of men interested in football; the orange is a more progressive community interested in art and culture; and the violet consists mostly of conservative politicians. Visualizations were created in Gephi.

In addition, we compared the average probability for the hate speech topics between the two groups. The community that hinted a strong religious ideology (yellow: M₁ = .019, SD₁ = .014; M₂ = .025, SD₂ = .014) had a significantly higher probability for both vulgar and politically charged harmful speech topics than the group of users supporting the opposition, orange: M₁ = .013, SD₁ = .008; M₂ = .017, SD₂ = .01; Welch’s t test: t(150.92) = 3.53, p < .001; t(116.84) = 3.97, p < .001. ¹⁰

To further identify the political orientation of these accounts, we identified who they followed and added 200 of the most followed accounts to the analysis. With this additional information, the political divide became even more evident. The results revealed that the majority of accounts that engage in politically charged hate speech follow the highest authorities of Iran. This in particular includes the Supreme Leader of Iran (khamenei_ir) whose ideological fanatics follow his lead on many topics (see Figure 6, yellow community). Further inspection of these accounts disclosed that these are mostly religious or were related to the security and intelligence forces of Iran.

As captured in Figure 6, next to the religious community (yellow), we identified a community with mostly men interested in football (blue) following the football players such as cristiano—the Portuguese football star, Cristiano Ronaldo—and aliiiiikarimi8—one of the most famous Iranian football players—and a conservative political community (violet) following iran_alimotahar—a conservative member of the Parliament—and hrouhani—the current president. The orange community consists of users interested in pop culture, including popular singers and actors. Users from the non-religious community in Figure 6 were identified commenting on the same pictures as the religious users did. But, they did not participate in disseminating hate speech. The presence of these ostensibly progressive users explains why the counter speech topic also appeared in this community.

Weaponized harassment and disinformation

State-imposed Internet filtering in Iran remains a key factor in choices that the public makes to adopt and use any social medium. Circumvention tools are widely accessible and their application is commonplace (ASL19, 2013). However, adoption of these tools is subject to a number of factors: (1) censorship apparatus shifts gears rather quickly to restrict access to circumvention tools that bypass censorship; (2) While using circumvention tools is not illegal, it is perceived as a means to accessing outlawed content. The availability of Instagram in Iran can therefore explain the appeal for state-aligned actors to coordinate harassment campaigns. Perpetrators prefer non-filtered platforms to demonstrate their adherence to the rule of law and the leadership of Iran.

Evidence of this form of abuse includes the terms and tone used by various accounts, timing of comments, and choice of profile pictures and captions, some of which signal a vivid religious or political affiliation. In this study, we identified a group of perpetrators that were highly likely to appear together in hateful attacks against the individuals in our sample. As many interviewees indicated, several attackers had signaled in their threatening messages that they had learnt about their targets in private forums, for example, Telegram groups. According to a few interviewees, several accounts had attempted to convince them that to avoid further harassment, they better remain silent about sensitive topics. Such evidence suggests that the abusive behavior documented herein may be an organized effort to collectively silence these targets.

Nevertheless, attribution of attacks remains a challenge. In some cases, it is highly complex to differentiate between patriotic fanatics and those officially organized by the state. In many cases, state-backed media such as the Islamic Republic of Iran Broadcasting (IRIB) or websites affiliated with IRGC incited violence against our sample by covering fabricated or manipulated news about these individuals. This may have triggered extreme reactions from the patriotic and religious audiences of these media.

While the financial motivation of perpetrators cannot be confirmed, there is evidence that the Islamic Republic has an abundance of volunteers to bolster its ideology on social media. In February 2017, Iran announced the recruitment of 18,000 “volunteers” to monitor and report “unlawful content” to the Iranian authorities (BBC Persian, 2017). This may signal other potential recruitments for participation in mobs and other abusive behavior in cyberspace. Our findings indicate that propaganda and ideology are potent drives for targeting of influential diaspora figures. And social media offers a more affordable and pernicious way of suppressing these voices in a manner that was not possible prior to the proliferation of networked communications.

The evolving dynamics of repertoires of suppression

Practices outlined herein demonstrate a shift in the way that the Islamic Republic engages in strategic communications. As critical voices extend their protest repertoires to nontraditional platforms like Instagram, the state also learns to adapt its suppression repertoire to these new mediums (Heydemann and Leenders, 2011). For example, faux reports of abuse or violation of terms of service by anonymous users are increasingly common practices that aim to take dissidents content down or have their accounts suspended. Multiple journalists and activists whom we interviewed confirmed that they had been subjected to similar silencing tactics. In response, targets often try to verify their accounts with social media platforms—one of the only available remedies against impersonation and false flags (Kargar, 2017). However, not many succeed due to the cumbersome, and sometimes contradictory, practices of these platforms to grant verification badges to applicants.

Polarized debates about the potential of the Internet in the context of authoritarian regimes are fairly common (Rauchfleisch and Schäfer, 2015; Sullivan, 2014). China, for example, has its own unique repertoires of suppression (for an overview, see Roberts, 2018). Optimists like Shirky (2011) consider the Internet’s potential and its affordances for protesters and dissidents living under authoritarianism. Skeptics like Morozov (2011), however, anticipate that authoritarian regimes eventually turn the Internet into the “Spinternet,” a “Web with little censorship but lots of spin and propaganda” (p. 117). The two positions are less contradictory if the dynamic nature of protest repertoires is considered: dissidents in authoritarian regimes continuously adapt their strategies—a process that can be best described as a “never-ending cat and mouse game” (Endeshaw, 2004: 41). In the absence of an official ban, Instagram has become an open battleground between the arms of propaganda, that is, state-aligned users and ideological fanatics, on one hand, and loyal fans and ordinary witnesses on the other.

In this context, weaponized harassment and disinformation seek two objectives: first, to strategically control the information flow to manipulate public opinion, garner support, and discredit opponents (Nyst and Monaco, 2018) and second, to curb “dissenting” speech and behavior. In these cases, political violence is instrumentalized to undermine trust in the work, ethics, and personalities of targets (Rid, 2013). Unlike other repertoires of suppression, for example, the Chinese model, propaganda, and state-approved content are unlikely to survive the diversity of communities and points of views on the Persian-language online spheres. Therefore, manipulation of public opinion may not have a potent chance to succeed.

However, the second objective, that is, suppressing critical voices is more complex to evaluate. While there is limited evidence on the emotional and psychological effects of online harassment (Feinstein et al., 2016), our interviews suggest that targets have internalized a long-term sense of precaution about their professional activities and the ways in which they engage in public debates. Several journalists in our sample emphasized that they have become more conservative in addressing socially and politically challenging topics such as lesbian, gay, bisexual, and transgender (LGBTQ) rights. They believed that exerting more caution would protect their families from suffering because of their sensitive work (Human Rights Watch, 2018). Curiously, none attributed such reservation to self-censorship.

As Rid (2013) argues, violence induced through computer codes in cyberspace is “physically, emotionally, and symbolically limited” (pp. 20–21). While this was immediately evident in the case of male interviewees, women still endured higher emotional and physical pressure as a result of weaponized, and in many instances sexualized, harassment. For women, cyber violence was not only virtual. It could have real implications for their physical safety. This gendered feature of cyber abuse observed herein corresponds to the experience of women in other authoritarian context and their encounter with weaponized harassment (Fincher, 2018; Sperling, 2014).

Final note

We explored the affordances of social media not only as they relate to dissidents but also in regard to the repertoires of suppression and how repressive regimes may utilize new media to their own interests. Further research on the insidious effects of weaponized harassment is required to shed light on the long-term, tacit consequences that targets, and in particular women, bear.

Our study is limited in scope to a specific subset of Iranian diaspora as targets of state-aligned harassment and to perpetrators who collectively engage in these campaigns. Future research can be illuminating for the formation of harmful speech, counter speech, and the communications that occur in between. While automated classification of users through topic models worked well in this study, it has some limitations. For example, more granular and implicit hate speech topics cannot be captured with our method. In multiple instances, users could not be appropriately identified as they used potential hate speech in only one of their many comments. As a result of using normalized probabilities, other non-hate speech topics stand out with a higher probability. In multiple other cases, distinguishing counter speech from hate speech could only be done through manual analysis. An example is when the same words are used by different commenters, some attacking the target, while others defending the same individuals. In such cases, only an extensive manual analysis can help to distinguish one use case from any other.

Finally, social media platforms have established policies, community standards, and algorithms to weed out harassment and abuse. However, it has become increasingly complicated to detect state-sponsored, state-executed, state-affiliated, or state-encouraged harmful speech that primarily targets the opposition groups abroad. Our study only partially explored the challenges that platforms encounter, in particular, when their technical solutions can be gamed by adversaries and their policies may not sufficiently cover the emerging threat models. Extensive multi-stakeholder collaboration can pave the way toward addressing these challenges.