Platformed antisemitism on Twitter: Anti-Jewish rhetoric in political discourse surrounding the 2018 US midterm election

Hatred of Jews in America

Antisemitism, as Shaheed (2019) writes, is commonly understood as “prejudice against, or hatred of, Jews” (p. 4), which is packaged into conspiracy theories, stereotypes, and honed through “religious doctrine and pseudoscientific theories” (p. 6). The International Holocaust Remembrance Alliance (IHRA) (2020: n.p.) has coined a popular working definition, referring to antisemitism as “a certain perception of Jews, which may be expressed as hatred toward Jews. Rhetorical and physical manifestations of antisemitism are directed toward Jewish or non-Jewish individuals and/or their property, toward Jewish community institutions and religious facilities.” Lipstadt (2019) describes contemporary antisemitism after the Holocaust, pointing out that “[w]hat should alarm us is that human beings continue to believe in a conspiracy that demonizes Jews and sees them responsible for evil” (p. x).

Antisemitism is a problem that permeates US society both on and offline. Antisemitic hate crimes in the United States rose to the highest level since 2008 in 2019 (Levin, 2020). Social media platforms are particularly popular vessels for antisemitic discourse (Ozalp et al., 2020; Zannettou et al., 2020). Despite this, a recent survey from the American Jewish Committee reports that almost half of Americans are unfamiliar with the term antisemitism (Mayer, 2020) and nearly 3 out of 10 Americans say, “they are not sure how many Jews died during the Holocaust” (Alper et al., 2020: 6). However, more than a third of American Jews say they have been a victim of antisemitism within the previous 5 years (Mayer, 2020). In 2019, the Pew Research Center (Doherty et al., 2019) found that 64% of Americans “say Jews face at least some discrimination” (p. 2), a number that rose by 20 percentage points from 2016. Despite an apparent lack of comprehension of the particularities of anti-Jewish hate, a majority of Americans appears to understand that antisemitism is a serious issue.

Antisemitism, as some researchers argue, needs to be understood as a “cultural category deeply embedded in collective memory” (Schwarz-Friesel, 2019: 315). It permeates all aspects of society, finding expression within the left, right, and center of the political spectrum, and within Christianity as well as Islam (Fine and Spencer, 2017). In its most outright forms, antisemitism includes Holocaust denial propagated by state actors such as Iran and by neo-Nazis in the West (Litvak, 2006). Holocaust denial reaches the core of right-wing thinking, such as Germany’s Alternative for Germany party or the American “alt-right” (Gallaher, 2021; Salzborn, 2018). Antisemitism is not a fringe issue, extending its purchase into mainstream political organizations such as the British Labour party under Jeremy Corbyn (Equality and Human Rights Commission, 2020). New antisemitism theory proposes that concealed versions of antisemitism use the vehicle of a critique of the State of Israel (Cousin and Fine, 2012). Importantly, criticizing Israel in and of itself is not antisemitic, but it can become “antisemitic under certain conditions” (Cousin and Fine, 2012: 177), depending on prejudice, stereotyping, and generalizations.

Hate speech and social media platforms

Hate speech conventionally refers to the communication of disparagement against people based on one or more identity factors. ² With this in mind, the concept generally refutes easy classification (Fish, 2019; Sellars, 2016) and is “equivocal [in] that it denotes a family of meanings” (Brown, 2017: 562). The United Nations define hate speech as “any kind of communication in speech, writing or behaviour, that attacks or uses pejorative or discriminatory language with reference to a person or a group on the basis of who they are” (p. 2)—based on a range of identity factors. Building on this definition, antisemitic speech squarely falls into the parameters and family of meanings occupied by the term hate speech.

Definitions of and repercussions for hate speech differ across national jurisdictions. For instance, in the United States, it may be legally permissible to issue antisemitic statements in public. This is different from countries such as Germany, France, Austria, and Israel, in which specific laws outright ban, for instance, certain speech that denies the existence of the Holocaust (Carmi, 2008).

Social media companies, to protect their brands (Roberts, 2019) and their users typically employ community standards and content moderation with the intent to weed out certain types of speech (Gillespie, 2018a; Nurik, 2019). Since social media platforms are private operators, they have license to eliminate as much and any which types of speech that they choose to, including antisemitic speech. Platforms such as Google’s parent company, Alphabet, have sought to create machine learning tools like Jigsaw’s Perspective API (Wakabayashi, 2017) to weed out toxicity. The solutionist premise of such propositions is that “there is nothing wrong with the tech that can’t be fixed by the tech” (Seymour, 2019: n.p.). But tools are not panaceas and can be riddled with bias. For instance, when auditing Perspective, the civil society organization AlgorithmWatch found that text that included “as a Black person” or “as a gay” had a higher chance to be considered toxic than phrases with other descriptors (Kayser-Bril, 2020). This pinpoints that hate speech detection posits issues that are difficult to resolve. For instance, simple lexicon-based approaches may lead to many false positives as words are sometimes used in an antisemitic manner—and sometimes not. Therefore, researchers recommend approaches that are “based on the semantics of the problem rather than allowing them to be domain-agnostic” (Olteanu et al., 2017: n.p.)—which is precisely what the present study has set out to do.

Online antisemitism in the Trump era

In the Trump era, antisemitic expression has appeared in communication from people in the highest echelons of the government, including within the president’s own tweets (Anspach, 2021). Researchers argue that this has contributed to a normalization of antisemitism (Tuters and Hagen, 2020). However, not all of Trump’s tweets about Jews are seen as antisemitic—some language he uses “humanizes Jewish oppression, demonstrates Holocaust remembrance, and advocates against the antisemitic hate crime” (Coe and Griffin, 2020: 7).

Antisemitism is especially virulent online (Schwarz-Friesel, 2019; Shaheed, 2019), marring mainstream social media platforms such as Twitter (Ozalp et al., 2020) and video games (Ingersoll & Anti-Defamation League, 2020), and roaming relatively unhinged on fringe platforms, such as 4chan and Gab (Zannettou et al., 2020). For example, in the lead-up to the 2016 US presidential election, more than 800 journalists were targeted with antisemitic tweets (Anti-Defamation League, 2016).

Outlets for hate have had a purchase on online life for perhaps as long as people have had access to the Internet (Borgeson and Valeri, 2004; Daniels, 2009). Sometimes disguised and not openly antisemitic, cloaked websites and online communities may “call into question the basis of what we say we know about ethnicity, racism and racial equality” (Daniels, 2009: 661), implicitly endorsing antisemitism. Research has found that major political events such as the 2016 US presidential election and the 2017 Charlottesville “Unite the Right” rally are often followed by surges in the spread of antisemitic content online, which makes its way from fringe forums onto mainstream platforms (Zannettou et al., 2020).

We focus on Twitter in this study since, for a long time, the platform has been defined by its “anything-goes approach” (p. 64), and has served as an avenue which fulfills a critical role for white nationalists and antisemites to mainstream their ideas and shift the Overton window of acceptability (Daniels, 2018). Seymour (2019: n.p.) goes as far as suggesting that “something about social media is either incipiently fascistic, or particularly conducive to incipient fascism.” Meanwhile, the relationship between Twitter and white supremacy has been coined a “love story” (Daniels, 2017: n.p.). Panizo-Lledot et al. (2019) have studied alt-right forums during the US midterms, identifying white supremacist, antisemitic, and islamophobic themes in the discourse. With these prior findings in mind, the 2018 US midterms serve as a political event with sufficient magnitude for our study during which one could reasonably expect antisemitic speech to emerge. We ask the following research questions related to both the nature and semantics of antisemitic discourse on Twitter during the 2018 contest:

Platformed antisemitism

When discussing antisemitism, it is crucial to locate it within a larger system of racisms (Cousin and Fine, 2012). The relationship between racism and antisemitism, however, is not a simple one to conceptualize. Both terms center around hostilities toward specific groups in society (Banton, 1992). Racism originally formed around the idea of racial antisemitism (Bernasconi, 2021), though contemporary understandings of racism typically focus on anti-Black racism. Fanon (1967) once referred to Black people and Jewish people as “brothers in misery”—with both antisemitism and racism as indicative of “the same collapse, the same bankruptcy of man” (p. 86). Du Bois (1952), upon visiting Germany under Nazi rule, reflected on his impressions of the Nazi terrors against Jewish people: “It had never occurred to me until then that any exhibition of race prejudice could be anything but color prejudice” (p. 251). Contemporary readings of antisemitism see it as a “form of racism” or “religious discrimination” (Moerdler, 2017: 1282). Rich (2016) summarizes that “antisemitism differs from other forms of racism because it uses conspiracy theories to claim that Jews are a powerful, controlling influence in society” (pp. 201–202). In this study, we build upon Matamoros-Fernández (2017) who coined the term platformed racism, which (1) casts platforms as ‘‘amplifiers and manufacturers of racist discourse by means of their affordances and users appropriation of them” (p. 940), and (2) argues that platforms reproduce inequality by way of content policies and how these rules are enforced (Matamoros-Fernández, 2017). The notion of platform, in the context of this study, refers to “sites and services that host, organize, and circulate users’ shared content or social exchanges for them” (Gillespie, 2018b: 254).

We take this as a point of departure to propose the concept of platformed antisemitism, acknowledging that racism and antisemitism are distinct—although related—from one another. Platformed antisemitism is a specific form of platformed racism that discriminatorily targets Jewish people and puts a strong emphasis on conspiracy theories and tropes. Platformed antisemitism is similar to platformed racism in that it is contingent on platforms for its distribution and their affordances that allow it to flourish, but topically different in its subject matter and specific forms of discrimination. We understand platformed antisemitism as the particularities of antisemitic discourse, ridden with conspiracy theories, prejudice, stereotypes, Holocaust denial and more veiled forms of hatred aimed at Jews, amplified and moving within and through platforms, on the fringes and in the mainstream.

Platformed antisemitism often manifests during major hate events on the Internet. For instance, during the sexist attacks that defined #Gamergate, the harassment campaigns that were waged against Los Angeles-based video game journalist Anita Sarkeesian included accounts of platformed antisemitism such as slurs or caricatures (Braithwaite, 2016). Platformed antisemitism can also be discerned in semantic developments particular to specific social media platforms that then later spread across a range of platforms. For example, the use of three parentheses before and after a particular name was and continues to be used by antisemites on social media platforms to bring attention to a person’s Jewishness (Tuters and Hagen, 2020). These parentheses have taken on a life on their own, at times decoupled from an antisemitic context, a phenomenon which Tuters and Hagen (2020) describe as a floating signifier. This means that although some people may not even know what the parentheses originally entailed, their usage normalizes and propagates antisemitic symbolism.

Affordances of platformed antisemitism: the case of Twitter

To understand platformed antisemitism, it is imperative to investigate particular social media affordances that are used toward its ends. The concept of affordances in communication research generally refers to the range of opportunities and actions one can potentially derive from something. According to Davis and Chouinard (2016), affordances are “the functions attached to a given object—what, potentially, that object affords” (p. 242). Social media, for instance, afford human interaction (Evans et al., 2017). Individual buttons—likes, hashtags, retweets—can also be understood as affordances (Bucher and Helmond, 2018). Ichau et al. (2019), when studying representations of Jewishness on Instagram through hashtags such as #jew, #jews, and #jewish, have identified antisemitic memes propagating stereotypes. Indeed, hashtags are relevant affordances as they can “create an interest-based micro-network” (Ozalp et al., 2020: 8). They can be understood as “tools of activating certain interpretative frames” (Lindgren, 2019: 421). In a similar vein, researchers have argued that the acronym of President Trump’s 2016 US presidential election campaign MAGA—“Make America Great Again”—when used as the hashtag #maga—has served as an affordance for community building among white supremacist organizations (Eddington, 2018).

Twitter as a platform allows users to retweet or quote-tweet other people’s posts. Retweeting is, in essence, “the Twitter-equivalent of email forwarding where users post messages originally posted by others” (boyd et al., 2010: n.p.). Retweeting allows other people to join threads and expands visibility beyond single publics. Aside from simple retweets, the platform also affords quote-tweets, a feature in which one can add their own contextualization to a retweet (Molyneux and Mourão, 2019). The possibility to add commentary when resharing content as a quote-tweet has been referred to as an “understudied, yet common” (Wojcieszak et al., 2021: n.p.) phenomenon. Retweets have been shown to play an important role in the formation of alt-right networks for “identity construction” (Shahin and Ng, 2020: 2425). Against this background of affordances—tweets, hashtags, and retweets/quote-tweets—with the potential to allow platformed antisemitism to flourish, we ask the following three additional research questions:

These questions help us parse apart which are the most popular hashtags that appear across content that is exclusively used in an antisemitic manner (RQ3), hashtags that yield high percentages of associated tweets that are antisemitic (RQ4), and differences in the degree of how antisemitic tweets are between a set of original tweets and retweets (RQ5). Contrasting datasets with original tweets and original tweets plus retweets/quote-tweets is important: The context that a quote-tweet adds may be able to reframe an originally antisemitic tweet so that—when considering the original tweet plus the comment in its totality—it is no longer antisemitic in its potential intent.

Content analysis

We sought to capture political conversation on Twitter as it was unfolding leading up to the 2018 US midterm election. To construct our dataset, in a multi-step procedure beginning with the hashtags #2018 midterms, #vote, #bluewave2018, and #redwave2018, we first identified, by way of purposive and snowball sampling, 228 hashtags around the election. Delineating conversations by proxy of hashtags in such a manner is an established way of conducting research on political conversations on Twitter (Bradshaw et al., 2020). In a second step, we used Tweepy, a Python library through which one can access the Twitter API, to collect a population of 5,843,282 tweets that pertained to our hashtags in the time span between 31 August and 17 September 2018. Third, we compiled a list of 50 terms that were hypothesized to be used in an antisemitic manner. We drew these terms from known antisemitic forums on 4chan and 8chan, direct expert consultation with the Anti-Defamation League (ADL), and from literature on antisemitism (UNESCO & OSCE, 2018). Among our collection of more than 5 million tweets, 99,062 of them (1.7%) contained one or more of these hypothesized antisemitic terms. We conducted a content analysis of these 99,062 tweets to adjudicate whether each specific term usage was indeed antisemitic. Research recommends combining computational and manual methods in big data content analysis to preserve contextual nuance (Lewis et al., 2013). Our unit of analysis was a single tweet. Of the 99,062 tweets in our sample, only 14.5% are original tweets, whereas 85.5% are retweets (such retweets include quote-tweets in which someone else might have appended their own commentary to the quoted tweet). Among the population of original tweets 17.3% were human-coded as antisemitic, whereas among the population of tweets plus retweets, 25% were human-coded as antisemitic.

In accordance with what we conceptually outlined in the literature review, human coders assessed whether a tweet carried either (1) a reference to a specific Jewish person as well as a trope, prejudice, or conspiracy, or (2) a reference to Jewish people writ large in combination with a trope, prejudice, or conspiracy. Coders were asked to assess whether a Jewish person seeing a tweet would reasonably perceive it as antisemitic. Negatively charged references to Jewish people or a particular Jewish person (e.g. “Soros donates millions to Democrats!” or “the Netanyahu government are corrupt!”) were considered not antisemitic (if harsh) political criticism, but if references also invoked antisemitic tropes/conspiracies or generalizations about Jewish people (e.g., “ . . . the soros dream to your state of a country submissive to globalism and control” or “deep state cartel zionist”), they were classified as antisemitic. One co-author and one student research assistant coded 1000 tweets from the sample for training purposes, as well as a second round of another set of 1000 tweets, which yielded sufficient intercoder agreement (Krippendorff’s α = .86). These comments represent 1% of the sample size. Research suggests drawing pretest content from a separate sample to assess intercoder reliability (Lacy et al., 2015)—a practice that was not available to us. Therefore, another colleague provided tie-breaking votes to definitely assess ambiguous results during pretests. The rest of the sample was split evenly between the first and second coder, who carried out the coding process between 4 and 21 May 2020. This dataset with manual annotations forms the basis for further analytical pursuits in our study.

Analytical pursuit

The analysis consisted of two components: Quantitative analysis via logistic regressions based on human annotations to yield terms and hashtags most often associated with antisemitic speech, and qualitative textual analysis to parse the content of tweets—what references were made, topics that were touched upon, and more.

First, to analyze which terms were predictors of antisemitic speech, we ran logistic regression analyses, with human-coded antisemitism classification as the dependent variable. Predictor variables consisted of 50 terms identified by researchers as often used in an antisemitic fashion that were entered into the model as dummy variables. In line with recent developments in communication research, we heed recommendations to utilize penalized regressions (Hindman, 2015) typically used in computer science/machine learning (e.g. Beltran et al., 2021). Regression models such as the one used here—least absolute shrinkage and selection operator (Lasso)—are “especially useful when the number of available variables for a regression model is very high” (Scherr and Zhou, 2020: 205). The Lasso method is useful for improving model interpretability by reducing dimensionality of the model. It operates by shrinking regression coefficients toward zero and—unlike the ridge regression—some to be exactly equal to zero. It effectively performs variable selection, removing redundant predictors that yield no additional explanatory power.

Using the R package glmnet (Friedman et al., 2010), we performed a 10-fold cross-validation using deviance as the evaluation criterion. We chose the optimal value of the shrinkage parameter λ per the one-standard-error rule, that is the value of λ that yields the most regularized model such that the cross-validation error is within one standard error of the minimum. We then ran a logistic regression using the predictors selected from the Lasso (Xp where|βp| > 0 at λ = lambda.1se).

Second, we performed a qualitative textual analysis of all tweets. We employed coding techniques from grounded theory research such as open and axial coding to parse through the data (Corbin and Strauss, 2008). Open coding refers to a process in which researchers form associations and come up with abstract terms and heuristics to describe phenomena and observations in the data. Axial coding refers to interrelating these open codes in a next step, finding commonalities and dominant categories, and working toward the grouping of the most prevalent dimensions relevant to the research question. Constant comparison, a strategy which emphasizes the continuing re-evaluation of coding schemata, allowed us to identify three topics and themes representative of the discourse which we later present in the results.