Sharma, S. & Brooker P. (2016) #notracist: Exploring Racism Denial Talk on
Twitter, in J. Daniels, K. Gregory & T.M. Cottom (Eds.) Digital sociologies.
University of Bristol: Policy Press.
#notracist: Exploring Racism Denial Talk on Twitter
The study of race online points towards not only extant forms of racism enduring on the
internet, but the emergence of new and unique practices (Daniels 2009; Nakamura & Chow-
White 2012). The development of 'Web 2.0' social media and networking platforms such as
Twitter, Facebook, Instagram and YouTube, have expanded user participation and intensified
online interactions. The rapid rise of social media appears to be proliferating racism and
racialized expression (in addition to forms of misogyny and homophobia). While it is difficult to
ascertain if social media is responsible for escalating practices of racism - see for example
Roversi (2008), Meddaugh and Kay (2009) - it has been central to increasing the visibility and
publicness of expressions of racialized discourse.
How may Digital Sociology approach the study of racism in ever-changing mediated spaces? Les
Back and Nirmal Puwar (2013) advance the discussion of a 'Live Sociology', by making the
important claim that innovations in research methods and developing new, critically reflexive
sociological devices are essential for grasping a digital landscape. Furthermore, Lisa Adkins and
Celia Lury (2009) contend that the digitization of everyday life is reconfiguring notions of
stability and social structure, meaning and signification, and the changing relations of
representation, experience and understanding. They contend that sociological research is
compelled to ‘break with representational models of the empirical…and…confront a newly
coordinated reality, one that is open, processual, non-linear and constantly on the move.’
Our essay offers an investigation of the phenomenon of racism denial on the micro-blogging
Twitter platform in the form of funded case study, which has a distinctive socio-materialist
methodological focus. Twitter has established itself as an influential online communication
medium for the dissemination of news and information sharing. Its ‘real-timeness’ and virality
of information diffusion have drawn attention to its capacity to intervene in the social world,
such as a means of co-ordinating emergency relief or influencing global political events (Murthy
2012). Breaking news stories and controversies dominate how Twitter is perceived to operate,
leading to issues propagating through its network and beyond, with the capacity to acquire
mainstream media status. While a burgeoning body of 'Twitter studies' research is emerging,
there has been limited research work studying racialized discourse (Demos 2014). Little is
known about the how modalities of everyday racial expression play out on the Twitter platform,
and particularly practices of racism denial.
Our account of Twitter race-talk aims to offer a unique intervention, by presenting a
methodologically motivated study. Its ambition is to highlight the significance of developing
critical race theory vis-a-vis engaging with the technological affordances of digital media. We
elaborate an instance of doing Digital Sociology from an approach that deploys the concept of
the assemblage (Langlois 2011; Lupton 2014) for understanding the constitutive relations
between the human (social media users), social phenomena (race and racism), and the non-
human (digital technologies and devices). More specifically, the study explores the techno-
cultural practices of Twitter by focussing on use of hashtag operators in creating the conditions
for the production of racialized meaning. Hashtags are notable for conveying more than
linguistic meaning, as they shape how users interact with the Twitter platform (Zapavigna 2011;
Sharma 2013). We empirically examine and analyze a relatively large corpus of tweets featuring
the #notracist hashtag that formulates one rivulet of the overall Twitter stream of racialized
discourse. This hashtag was selected on the basis that it makes apparent expressions of racism
denial. Moreover, the affiliative function of the hashtag is considered as means of exploring the
'imagined audience' (Marwick and boyd 2011; Zapavigna 2015) of users propagating
expressions of racism denial.
The first section of the essay briefly explores the significance of racism denial talk in relation to
the shifting nature of the private and public sphere. In a post-civil rights era, the public
expression of racism has become increasingly regulated and sanctioned, yet it has given rise to
covert forms of racialized expression which seek to deny racist intent (Bonilla-Silva 2010; Picca
and Feagin 2007). The current understanding of racism denial is limited to 'off-line' spaces, and
it remains an on-going task to explore distinctive on-line practices.
The case study research process has not been linear, involving flitting between theory, the
filtering and refinement of empirical data, and undertaking a grounded analysis. The second
section of the essay outlines our methodology, focussing on the significance of Twitter hashtags
and the Chorus software tool used to undertake the data collection and analyses. A dataset of
approximately 25,000 individual twitter messages (tweets) that included the hashtag #notracist
was harvested over a period of time, which formed the basis for analyses. We offer a discussion
of how working with Chorus - as a 'methodological device' (Lupton 2014) - formulates a
component of a socio-material assemblage in the production of visual analytics of Twitter race-
The third section presents a discussion of the dataset via Chorus analytics, by highlighting that
#notracist is not about any specific event or issue as such. Rather, it is characterized by a
steady, relatively low-volume of tweet activity, around a wide array of different sub-topics
which bubble away on Twitter without ever trending or becoming visible. In contrast to the
majority of event-based Twitter studies, we contend that an alternative approach is required
for investigating everyday types of racialized 'micro-aggressions', which are not necessarily
explicitly visible on social media. Furthermore, our analyses indicate that for the #notracist
dataset, multi-hashtagging is a key practice in the differentiation of types of Twitter race-talk;
and distinguishing between modes of racism denial can be achieved praxiologically rather than
focussing exclusively on semantic meaning. Our approach seeks to grasp the digital materiality
of hashtags, beyond text-based or linguistic-oriented accounts of Twitter talk that ostensibly
dominate the emerging field of social media analytics.
The findings and analyses presented here are not exhaustive, and nor do they fully attend to
the complexities of racialized expression on social media. Rather our aim is to offer an example
of a how a Digital Sociology of racism can develop an approach which brings together an
analyses of technology, language, race and power (cf. Brock, 2012).
An important body of academic research examining internet racism has become established
focussing on extreme right-wing/neo-Nazi websites and discourses (Daniels 2009; Meddaugh &
Kay 2009; Roversi 2008 ). While the field of internet research has diversified by exploring other
forms and spaces of online racism, in relation to social media and particularly Twitter, there are
currently a paucity of relevant studies. The majority of this work has been directed towards
investigating forms of racist 'hate speech', that includes abuse and insults towards minority
groups. Notably, Twitter is singled out to be the most popular platform for propagating forms
of hate speech. For example, a recent study (Kick It Out 2015), exploring online discourses
concerning UK Football, discovered that 88% of 'discriminatory language' (targeted at football
players and clubs) specifically circulates on Twitter, in comparison to other social media
platforms. The large-scale study, conducted by Demos (2014) entitled 'Anti-Social Media'
investigated the presence of 'hate speech' (in the form of ethnic slurs) on the Twitter Platform.
It found that that approximately 10,000 English language tweets per day include a slur.
The Demos study also points to challenges of identifying whether changes in modes of
communication are responsible for the apparent increase in hate speech. And it highlights that
the explosion of online communication enables the researcher to more readily access and
examine 'public' forms of racism:
[H]ate speech online...does appear to be increasing dramatically. This might reflect a
change in the way we communicate rather than an increase in the amount of hateful
speech taking place: communicating online makes it easier to find and capture instances
of hate speech, because the data is often widely available and stored (Demos 2014:11).
Researching online hate speech is important for gauging visible and public expressions
of overt forms of racism. Nonetheless, it does not directly address how phatic, everyday
and more indirect modes of racism are present; and which kinds of (rhetorical)
strategies are employed to negotiate the boundaries of acceptable public speech.
The fields of critical discourse analysis, linguistics and social psychology have developed a body
of work that explicates racialized discriminatory language in everyday and institutional public
talk (Augostinos & Every, 2007; Billig 1988; Potter & Wetherell 1987). Martha Augoustinos and
Danielle Every identify how these types of racialized discourse are invoked:
...patterns of talk around race ... can be seen to reflect not only interpretative
repertoires, that is, a set of descriptions, arguments, and accounts that are
recurrently used in people’s race talk to construct versions of the world ... but
also discursive resources that perform social actions such as blaming, justifying,
rationalising, and constructing particular social identities for speakers and those
who are positioned as other (2007: 125).
Discourse and language analysts have acknowledged the ambiguous and contradictory
nature of race talk. The unsettled and shifting meanings of racism have resulted in some
analysts refraining from making explicit categorizations and judgements ‘as to what
counts as racist but instead examine whether speakers themselves treat the talk as such
and analyse how it is managed and attended to in social interaction’ (Augoustinos and
Every 2007: 124-5). However, rather than merely acknowledging ambiguity and
contradiction in race talk, we can consider this kind of linguistic ‘indeterminacy’ as
symptomatic of contemporary forms of racism in a post-civil rights/'political correctness'
era: expressions and practices of racism can be more covert and obfuscated. Moreover,
from a sociological stand-point, it is crucial to maintain that racism is not simply a
question of individual prejudice or pathology. Expressions of racism - whether overt,
covert or contradictory - continue to reinforce racialized hierarchies and power
structures in society (Picca and Feagin 2010).
A post-Civil Rights era has resulted in the rise of legislation and social regulation against
certain forms of racist expression and 'hate speech'. Direct and explicit racist discourse
is less publicly and morally acceptable due to stronger anti-discriminatory social norms.
There is an increased public sensitivity towards avoiding inappropriate use of racist
language. Critical race scholars such as Eduardo Bonilla-Silva (2010) and Leslie Picca and
Joe Feagin (2007) maintain that the apparent decline in publicly (i.e. off-line) overt racist
discourse, has been substituted with subtler, covert and coded racialized expressions.
This has resulted in more strategic forms of public race-talk, particularly in relation to
practices in the ‘denial of prejudice’ which can pervade everyday racist talk (van Dijk,
1992). Strategies of denial can commonly take the form of a disclaimer:
Analysis of post-civil rights racial speech suggests whites rely on 'semantic moves,' or
'strategically managed...propositions'...to safely state their views. For instance, most
whites use apparent denials...or other moves in the process of stating their racial views.
The moves act as rhetorical shields to save face because whites can always go back to
the safety of the disclaimers... Phrases such as "I am not a racist"...have become
standard fare...They act as discursive buffers before or after someone states something
that is or could be interpreted as racist. (Bonilla-Silva 2010: 105)
Picca and Feagin (2007) develop a Goffman-inspired analyses of contemporary racialized
expression in terms of identifying differing 'frontstage' and 'backstage' racial
performativity. Rather than overt racist discourse disappearing, its articulation has been
mostly consigned to the 'private' backstage, generally hidden from public scrutiny. In
contrast, the frontstage performativity of covert racist expression can involve 'saving
face' via public disclaimers. These authors, alongside other scholars such as Nina
Eliasoph (1998) and Raúl Pérez (2013) also highlight the defensive role of joke-telling
and comedic performances, as means to continue to express more overt forms of racism
in public spaces.
To date, no specific studies examining the practices of online racism denial on social
media platforms have been conducted. While there is research examining explicit
modes of internet racism (see Daniels 2012), the more coded practices of expressing
racist comments while simultaneously denying racist intent is far less understood in
terms of its online manifestations. What is of interest, is whether off-line racism denial
strategies are being reproduced on social media, and/or if new online practices are
emerging. Do the technological affordances of Twitter facilitate unique modalities of
racism denial? Moreover, online communicative practices, to varying degrees, can blur
the boundaries between public/private spaces and front/backstage performances (Baym
2010; Daniels 2012). The existing Twitter studies exploring hate speech indicate that
some of its users breach normative boundaries of acceptable speech. Somewhat in
contrast, as we shall discover in our analyses section, users in our study appear to
acknowledge the existence of these boundaries through their use of the 'disclaimer'
hashtag #notracist. In this respect, it may be the case that different sets of Twitter users
hold differing notions of their 'imagined audience':
Given the various ways people can consume and spread tweets, it is virtually
impossible for Twitter users to account for their potential audience, let alone
actual readers...Without knowing the audience, [users] imagine it. (Marwick &
Before turning to the analyses of our study, it is productive to discuss the
methodological approach we deployed, as it central to the developing a digital sociology
Notes on Methodology
Identifying racialized talk (including racism denial) on social media is a challenging task, because
there exists a huge array of linguistic terms and repertoires signifying variegated racist
expression. These can range from: extreme racist abuse; insults and micro-aggressions; and
obfuscated talk in which racism is covert, indirect or coded. As expressions become less
explicitly racist, they become increasingly difficult to identify and interpret by the social
researcher. This is particularly the case for expressions of racism denial, because of the
deployment of rhetorical and covert language in the act of refuting racist intent (van Dijk 1992;
Picca & Feagin 2007).
Our initial foray into identifying forms of racism denial on social media resulted in identifying a
handful of ‘anti-racist’ sites or accounts which exposed individual users' refutation of racism;
see Facebook public posts http://www.notracistbut.com/ and the tumblr site
http://imnotaracistbut.tumblr.com/. These indicated the popularity of permutations of the
phrase “I’m not racist but” on social media. Variations of this phrase were tested on the Twitter
search API, which led to locating the account @yesyoureracist. This account, included making
visible tweets which denied any racist intent. Examining these collated tweets indicated the
sporadic use of the hashtag #notracist within some messages. Concatenated in the form of the
hashtag, it appears that #notracist being included in Twitter messages echoed the "I'm not
racist..." strategy of racism denial. Investigating racism denial on Twitter via a hashtag such as
#notracist will exclude a whole range of potentially relevant Twitter data which does not
include this hashtag. However, our intention was not to undertake an exhaustive study, but
rather to focus our efforts by privileging the hashtag as a means to investigate particular
practices of racism denial which actively engage with the architecture of the Twitter platform.
Hashtags are a noteworthy phenomenon, because they have multiple uses on Twitter
(Zapavigna 2015). The practice of users attaching a label or ‘tag’ to online content such as a
message, document, image or video has become a central feature of ‘Web 2.0’ social sites.
User-based free-form tagging on social media platforms has been principally used for
information retrieval and recall, and in this respect is a posteriori. In contrast, tagging within
Twitter is primarily a priori, as it is commonly used for filtering and promoting messages in real-
time (Huang, Basu & Hsu 2010).
The hashtag − a single or concatenated term prefixed by the # symbol, for example, #obama or
#firstworldproblems − has become publicly synonymous with Twitter, although they feature in
less than 15% of messages of the whole Twitter stream (Liu et al. 2014). The Twitter platform
adopted this user-based ‘folksonomy’ practice by including it in its interface and rendering
hashtags as searchable hyperlinks. In particular, popular or trending hashtags are made visible
as part of the main Twitter interface (both web and mobile), and can collate hundreds of
thousands of disparate tweets, forming a networked sociality and enabling users to participate
in collective ‘conversations’. Many studies have focussed on hashtags ‘amplifying’ the
significance and findability of tweets, and generating ‘ad-hoc publics’ often with temporary or
shifting boundaries (Bruns & Burgess, 2011; Murthy 2012).
While the function of hashtags is variegated, they are significant in Twitter as ‘a form of “inline”
metadata, that is, “data about data” that is actually integrated into the linguistic structure of
the tweets’ (Zapavigna, 2011: 791). Hashtags can be deployed to categorize the content of a
message as 'topic-markers'; and as hashtags are user-created, this ‘bottom-up’ practice of
tagging can lead to both redundancy (many hashtags have the same meaning), and ambiguity (a
single hashtag has different meanings) (Garcia Esparza et al., 2010). Nevertheless, as discussed
by Thomas Vander Wal (2005), (hash-)tagging can be characterized by a ‘power law’
distribution which describes the phenomenon that a few tags are frequently used by many
people and in contrast, the majority of the remaining ‘long tail’ of hashtags are infrequently
Social researchers need to be careful not to circumscribe Twitter hashtags to principally acting
as online linguistic operators. One of the limits of privileging language-oriented analyses is that
'...text-focused methodologies deal with content in its linguistic and social aspects rather than
with the technological or material context that enables the production and circulation of signs'
(Langlois, 2011: 9). What is of interest in our study is how the techno-cultural affordances of
Twitter are generative of race talk in relation to the use of racialized hashtags. In this respect, it
is productive to deploy an alternative account of racialization, which does not only dwell on
semiotic meaning or the problem of representation. Conceiving race as a ‘digital assemblage’ -
which identifies processes of heterogeneous elements brought into sets of relations with one
another - facilitates an understanding of the emergence of racialization in online spaces by
exploring how it works and what relations it generates, rather than only the meanings it
produces (see Sharma 2013). This materialist approach of conceiving race (cf. Saldhana 2007),
considers the specificities of racism and how it is manifested in online spaces. Thus, specific
forms of racism denial can be grasped in terms of how it is formed in relation to a Twitter
techno-cultural assemblage, constituted by the informational logics of hashtags, software
interfaces and algorithms, networked relations, racial dis/ordering, and meanings and affects.
The dataset for our study was generated by collecting usages of the #notracist hashtag,
searched via Twitter's Search API between March - November 2013. This resulted in harvesting
24,853 tweets over the eight month time period.
The period was determined by the
constraints of the length of the funded research project, and based on monitoring whether
further harvesting led to data redundancy for the purposes of our analyses.
The empirical analyses of the dataset was developed through a visual analytic approach (Card,
Mackinlay & Shneiderman, 1999). This methodology has its origins in the fields of information
and computer science and has informed the development of Chorus,
a software suite capable
of collecting and visually parsing Twitter data. Chorus was deployed for generating the
#notracist dataset and assisting in its analysis. The primary tenet of visual analytics is that
visualisations should serve some functional purpose; as opposed to being merely images and
outputs, visual analytic representations are dynamic and interactive research tools. In our case,
Chorus was initially used to identify the frequency of the appearance of the #notracist hashtag
over the specified time period, and subsequently, to visualize the relationship between terms
(i.e. other related hashtags) in the #notracist dataset.
We are aware of the technological affordances of Chorus − it is not merely a method or tool for
analyzing a large corpus of Twitter data, because it governs what we perceive is possible to do
with this type of analytic approach. Chorus is a 'methodological device' (Lupton 2014) that
connects together both method (as technique) and the research object (hashtags). The data
visualizations produced by Chorus is a key step in studying the #notracist dataset. Moreover,
understanding how the software produces these visualizations is crucial towards developing a
meaningful analysis. Thus, Chorus constitutes an element involved in the production of a
Twitter assemblage that activates an analysis of racialized hashtags. While the technical work of
processing this type of Twitter data is accomplished by Chorus, a methodological understanding
of the workings of those processes and algorithms is necessary for explicating what is observed
in that data, and how it may be interpreted (see Brooker et al., 2015).
#notracist: Hashtagging Racism Denial
An initial exploration of the #notracist dataset via the time-line graph (Figure 1) generated by
the Chorus software, indicated that the most useful reading of the data would not come from
considering it as having a meaningful temporal dimension as a basis for analyses – little within
this data is found to change across time. Figure 1 presents a sporadic and diverse dataset with
few (if any) distinguishing features in terms of how the volume of usages of #notracist
fluctuates over time.
Figure 1: Timeline graph of the #notracist dataset. The grey bar chart shows tweet frequency in
daily intervals (with the dark grey bar showing proportion of tweets containing an URL link).
To give a sense of how voluminous the #notracist talk is on a day-by-day basis, it averaged out
at slightly over 100 tweets per day, with the least populated day in our data consisting of 36
tweets and the most populated day featuring 239 tweets. There is little in the dataset indicating
that #notracist captures a topic in a conventional sense, i.e. a visible issue or one that inspires
significant discussion between Twitter users around some focal event (such as the publication
of a news report or the broadcast of a TV show). The content of the tweets in the dataset
exhibit a wide variety of everyday commentary that appears difficult to organise into a
meaningful schema. Nonetheless, they share a commonality in the use of the #notracist
hashtag as a disclaimer that has a ‘distancing function’ (van Dijk 1992) from accusations of
racism. The inclusion of the hashtag exhibits practices of ‘interpersonal punctuation’ which is
declarative of a user’s ‘stance’ (Zapavigna, 2011). Individual users deliberately punctuate their
tweet indicating their supposed ‘non-racist’ disposition. For example:
MikepFennyy: finally got a new boss today. Hes under 50 good guy has social
skills totally white with zero accent. I am so pleased #notracist
rellavent: I Hate Basketball & Rap Music. #notracist
Brodyrey22: If its not white its not right #notracist
These tweets are exemplary for highlighting the diversity of banal racialized 'content' of the
dataset. It is interesting to observe that in the #notracist dataset the majority of users do not
have large numbers of followers, and rarely are messages with the hashtag re-tweeted. It is
difficult to ascertain the 'imagined audience' of these users when deploying #notracist.
Nonetheless, in addition to expressing a defensive stance, the inclusion of the hashtag suggests
an affiliative mode of communication. The interpersonal function of the #notracist hashtag
may invoke '...the notion that there are people who feel the same way as the
microblogger...regardless of the fact that it is unlikely that anyone would ever use the tag as a
search term' (Zapavigna 2015: 18). While the #notracist hashtag does not appear to beget
direct interactions between users, its deployment intimates a shared predilection of racism
In contrast to explicit racist tweeting which can gain social media visibility via high frequency re-
tweeting and/or @mention conversations,
the #notracist dataset lacks such traction;
#notracist tweeting generally occurs in isolation without any noteworthy presence. We can
speculate that the #notracist hashtag is indicative of a social media racism that follows a power
law distribution, that is, a racism of the 'long tail'. What is usually witnessed as social media
racism are those events that have gained significant traction and visibility. Arguably, there also
exists many more racist micro-events which are ostensibly inconsequential due to their
'invisibility' - for example, as background chatter - yet are symptomatic of forms of everyday
online racialized micro-aggressions (cf. Sue 2010). Conceptualizing a racism of the 'long tail' via
the hashtag, highlights #notracist as an element of a Twitter racialized assemblage: aggregating
(connecting) what appears to be spontaneously-occurring individual race-talk that materializes
seemingly coherent yet diverse practices of the denial of racist expression.
The significance of the hashtag in relation to a Twitter assemblage can be further elaborated in
terms of how it functions alongside other (non-racialized) multiple hashtags in the #notracist
dataset, which is where our attention turns in the discussion below.
Visualising Multi-Hashtags: ‘Truth’ and ‘Humour’
The time-line visualisation points to a dataset that is not significantly event-based. As such, our
analytic efforts were directed towards the exploration of ‘topics' consisting of aggregations of
terms that are more commonly used together. Thus, an alternative line of inquiry was pursued
using Chorus’ Cluster Explorer modelling, which build sets of visualisations to represent and
facilitate navigation around ‘topical’ clusters. These models plot the relationships of terms
(which can be words or other fields such as hashtags) as they are used together in tweets,
where a relationship signifies the commonality, that is, the co-occurrence of the usage of one
term with another in a tweet (cf. Callon 1983; Danowski 2009; Marres and Gerlitz 2015). A
cluster map is built up from direct and indirect relations of terms which allows a spatial
mapping algorithm to plot the relationship of one term to another as a function of distance
(where the closer a term is to another term, the more strongly it is related). In clustering
together strongly-related sets of terms – for example, the likelihood that two hashtags are co-
occurring within a tweet - Chorus provides a method of identifying and mapping distinct topics
and their inter-relations (without relying on a priori categories defined by the researcher).
kind of visual parsing of the #notracist dataset by the software is only one step towards an
analysis. Chorus is not able to discern the sociological significance and meanings of the relations
between terms it visualizes. Nevertheless, it is important to grasp how a cluster map is
produced, as it influences the trajectory of a deeper exploration of the dataset.
For the #notracist dataset, aside from the original #notracist term there were a further 7717
hashtags in use. That is, approximately 30% of the entire dataset consisted of more than one
hashtag being included (along with #notracist), which is remarkable as multiple-hashtagging is
not a common practice in Twitter (Liu et al. 2014). The following examples of tweets illustrate
practices of multi-hashstagging in #notracist dataset:
helen_louise_: I literally cant stop eating watermelon. & Im not even black.
PaneKilla: How to say the alphabet in vietnamese #funny #notracist #accent
#alphabet #vietnamese #peace #lol http://instagram.com/p/**********/
Given our original search query, which aimed to find usages of a specific hashtag, we plotted a
model which used hashtags as ‘nodes’ in the Cluster Explorer map - see Figure 2.
Figure 2: Cluster map showing the topical relationships between all hashtags within the
#notracist dataset (not including #notracist). Labels are given to hashtags which feature in >1%
This visualisation indicates a topical cluster map of multi-hashtags occurring with #notracist
(each node being a different hashtag). Immediately observable in Figure 2 is a tight central
cluster of hashtags (including #funny and #lol), which are closely related to each other and
demarcated in the inner (solid-line) radial. Although, there are also a number of significantly
populated nodes that feature on the outer branches extending from this central cluster
(including #truth, #iswear, #fact, #justsayin/g), often appearing on the end of branches −
located in the outer (dotted-line) radial.
The difference between the two radials is significant in as much they illustrate different
tweeting practices. The operational tendency of the Chorus clustering algorithm is to plot all the
highly populated nodes towards the centre of the map so as to make room for less connected
outliers around the edge of the map; we do not see this occurring. Picking through the most
frequently populated nodes in either radial, we find a thematic difference between the radials
as identified by two distinct ‘categories’, which supplements and coincides with their
algorithmic difference. Firstly, the inner radial consists largely of ‘Humour’ hashtags which are
intended by tweeters to mark tweet content as containing jokes or other comedic material.
Secondly, the most frequently occurring hashtags in the outer radial form a category of ‘Truth’
hashtags, which tweeters use to clarify or qualify their tweet statements by referring to them as
so-called observations and facts. The ‘Humour’ and ‘Truth’ categories are inductively derived
from the cluster map of Figure 2, which the radials reveal more clearly.
‘Humour’ tags (Inner Radial)
‘Truth’ tags (Outer Radial)
Table 1: Humour and Truth categories of hashtags and the frequencies of co-occurrence with
Table 1 offers a means of continuing the analysis and drilling down towards further insights
about hashtagged racialized talk in relation to a more nuanced grasp of what each of the two
categories (‘Humour’ and ‘Truth’ hashtags) consist of. Table 1 identifies other hashtags co-
occurring with #notracist, which are judged as significant in the formation of the ‘Humour’ and
‘Truth’ categories throughout the dataset.
At this stage of the analyses, it is productive to briefly turn our attention to the word content of
tweets, (rather than only hashtags as visualised in Figure 2).
‘Humour’ Top Terms
(1131 tweets, 1884 terms)
Table 2: Top Humour terms within the #notracist dataset
‘Truth’ Top Terms
(1347 tweets, 2417 terms)
Table 3: Top Truth terms within the #notracist dataset
Tables 2 and 3 reveal that the two categories 'Humour and 'Truth' share (loosely) a ‘dictionary’−
a palette of seemingly common terms used in tweets as a way of doing racism-denial Twitter
talk. There are a number of key terms (words) which frequently appear in both ‘Humour’ and
‘Truth’ tweets, such as: ‘black’, ‘white’, ‘people’, ‘like’ and ‘just’.
It seems improbable that
there will be a linguistic or semantic means of consistently distinguishing between either
category, for example:
Tegan_Molly001: Black girls vs white girls in the club #lol #comedy #notracist
LENNYSGUY: THE HARLEM SHAKE IS A BLACK THING. THAT WHITE GIRL ASIAN
GIRL HARLEM SHAKE BULLSHIT IS WEAK. #JUSTSAYING #notracist
Both of the tweets above, despite being located in different categories, use the key terms
‘black’ and ‘white’, and are substantively about comparable topics - differentiating between
black and white people based on stereotypes of how they dance. Hence, it is difficult to see
how words alone - without multi-hashtags as 'topic-markers' (Zapavigna 2015) - may provide a
way of distinguishing which tweets are intended as 'jokes' and which are intended as ‘factual’
A key question at this point is: what do these mappings say about the ways people
communicate race-denial content with hashtags on Twitter, given that both ‘Humour’ and
‘Truth’ categories draw upon a broadly similar set of words? Arguably, analyses so far indicate
that both categories are generated by user hashtag tweeting practices, rather than only the
literal content of their tweeting. It is useful to explore these practices more qualitatively by
using Chorus to reduce the dataset - via filtering relevant tweets - to continue the investigation.
A distinguishing feature between the ‘Humour’ and ‘Truth’ categories is in the usage of
hashtags to achieve different purposes. To demonstrate how this is visible in the data, we note
that the majority of tweets featuring a ‘Humour’ multi-hashtag also feature a URL link which
has an additional function of embellishing the message, for example:
KokoBugz: RT @AlanCaravaggio: How white people react to black athletes
#funny #revine #loop #notracist #VineStar https://t.co/WG******
KoryBoolet: #whitepeopleproblems #howto #remake #notracist #comedy #funny
#cute #magic #loop #unPOP #see #drivingvine https://t.co/ZX********
It appears that ‘Humour’ hashtag usage promotes or shares an internet object of some kind −
typically a Vine video or Instagram picture
− and the utilisation of multiple hashtags seemingly
maximises the visibility of the link. The linking (or inclusion) of visual media is a common
practice amongst internet users in the sharing of online humour (Shifman and Blondheim
2010). Moreover, the juxtaposition of these kinds of humour hashtags alongside #notracist can
potentially mutate both sets of hashtags: the ‘Humour’ hashtags become racially charged, and
the #notracist hashtag acquires greater affiliative characteristics to construct an 'imagined
audience'. Moreover, the ‘Humour’ category is remarkable for the sheer number of multiple
hashtags included in a tweet, and the hashtags themselves (alongside possible links) can
become the primary ‘meaning’ (content) of the message. While the content of some of these
tweets is difficult to interpret due to both a lack of meaning- and content-carrying words and an
abundance of hashtags, Shawna Ross (nd.: 5) intimates: ‘as a tweet asymptotically approaches
contentlessness, the resultant tendency toward abstraction denotes increasing (not decreasing)
Notably, there are a small set of ‘Humour’ multi-hashtags such as #lol, #haha and #loop
are frequently used together, (thus producing the central cluster observable in the hashtag map
of Figure 2). The significance of the these ‘Humour’ multi-hashtags can be further explored in
relation to their co-occurrence. As indicated in Table 4 below, there is a high degree of
coherence with which certain key ‘Humour’ hashtags co-occur, such as #loop and #comedy. For
example, #loop features in slightly over 50% of tweets that also feature #funny .
these types of tweets pertain to objects not residing within Twitter such as Vine videos.
Hashtag Co-occurrences with
Table 4: Top hashtag co-occurrences with #funny, showing the strength of relationship between
#funny and hashtags to which it is most related.
‘Humour’ as type of racialized talk relies on an implicitly-agreed-upon − seemingly a priori − set
of general classificatory hashtags which users recognise and draw on in order to situate their
tweets as embodying racialized humour (and not, they may hope, actual racist intent). This
practice of humour-based multi-hashtagging does not necessarily seek to explain the meaning
of the tweet, because the hashtags themselves − as dense, self-referential meta-data (Ross, nd.)
− are the tweet.
The circulation of humour on the web has become a ‘ritualized social practice’ (cf. Perez), and
users of social media are well versed in its discursive conventions. The use of a relatively narrow
set of multi-hashtags and inclusion of links suggest that the circulation of racist texts (tweets,
images, videos etc.) is an intensely collective enterprise. The invoked ‘imagined audience’
shares the joke and participates in a racialized online culture that breaches social norms. While
the distancing function of the disclaimer #notracist is present, its imbrication with humour
complicates and legitimizes strategies of racism denial, and makes them more resistant to
critique because of the collectivizing function of jokes via their public sharing.
In comparison, in the ‘Truth’ sub-set of the data we discover a tendency to use multi-hashtags
much more sparingly, though from a much wider range of hashtag terms; and in ways which are
intended to clarify or qualify the semantic content of tweets, for example:
J3N5TT3R: Asian guys only have two volumes, quiet and shout. The ones on the
next table are stuck on shout #notracist #fact
christophe1435: This economics tutorial is like 95% Asian. #notracist #truth
Here, the usage of hashtags reflects a more semantic orientation to the convention, where
hashtags indicate how the tweeter intends the tweet to be interpreted − their ‘stance’ − for
example, as not representing any racist intent (e.g. #notracist), and justifying this disaffiliation
with racism because the tweeter is stating what they argue is a defensible or observable
everyday truth (e.g. #justsayin/g). Unlike the small set of general hashtags which are frequently
used in ‘Humour’ tweets alongside other multi-hashtags, ‘Truth’ tweets rely on a broad range of
multi-hashtags which do not co-occur with other multi-hashtags for at least two reasons. Firstly,
these multi-hashtags tend not to be used with other hashtags, and secondly, each tag tends to
be used relatively few times. This gives the ‘Truth’ cluster map (Figure 2) its distinctive outer-
density pattern − the wide variety of largely non-associated terms appear almost entirely
disconnected (and unrelated) from each other.
It is fruitful to question why ‘Truth’ as a mode of online racialized talk of denial relies on a
diverse array of largely single-use hashtags, in comparison to ‘Humour’ which draws on a
relatively narrow set of hashtags that are used multiple times in tweets? The shared culture of
online humour suggests that the circulation of racist texts need not require an explicit
justification (e.g. #justjoking), and because for the user, the ‘imagined audience’ can be a ‘real’
one that shares the joke. In contrast, ‘Truth’-based statements include hashtags that attempt to
make explicit their semantic intentions (however misplaced or ignorant). These hashtags are
largely devoid of a shared online culture (apart from the possibility of #justsayin/g). As
Zapavigna notes, ‘The inline nature of #tag usage opens up the possibility of play with users
creating tags that are unlikely to be used as search terms and which instead seem to function to
intensify the evaluation made in the tweet’ (2011: 800). This strategy of intensifying a user’s
stance via adding another truth-type hashtag seeks to contain the ambiguity of racialized
meanings, and legitimize the possible breaching of the backstage of privatized racism (cf. Picca
& Feagin 2007; Bonilla-Silva 2010). Yet as indicated by the creation of many singular truth-type
hashtags, this practice is a fraught activity. The proliferation of different 'Truth'-based
justificatory hashtags is symptomatic of the dissonant registers of how race-denial is mobilised
in everyday online discourse, in which the ‘imagined audience’ in the final instance, remains
In summary, although the two categories, ‘Humour’ and ‘Truth’, share a lexicon - which is
remarkable given how little people appear to communicate with each other in the dataset - the
variations observed in the visualisations lie in the markedly different hashtagging practices that
tweets in each category display. Where ‘Humour’ tweets use many multi-hashtags for
propagation and dissemination of tweet (and often URL link) content, ‘Truth’ tweets use
singular multi-hashtags (i.e. #notracist plus one other hashtag) in order to rhetorically clarify a
potentially or purposefully ambiguous statement. Both types of tweeting practices are
modulated by a racialized digital assemblage. The ‘master’-hashtag #notracist organises and
racially charges other hashtags in so far as activating differential modes of racialization. In this
respect, race is not simply inscribed in Twitter messages, nor can it readily de-code their
meanings. Rather, modes of racialization emerge within and across tweets through the
aberrant connections elicited by multi-hashtagging practices. It is the variation of these
different hashtagging practices that may distinguish between the type of racialized talk being
published to Twitter, such that although the tweets themselves can broadly consist of similar
terms and semantic meanings, the adoption of hashtagging practices from one category or
another can change the affective meaning sufficiently to situate that tweet as joke-telling
and/or truth-telling. Hence, we find that racialized hashtagging on Twitter is, as a phenomenon,
not solely located in the words used by individuals, but in the evaluation of words by way of
hashtagging - a techno-cultural practice within Twitter that is influenced by societal modes of
This essay has advanced a research process for examining an intriguing type of racially-charged
social media data which is not structured temporally, but rather by an ambiguous ‘topicality’.
We explored the potential of ‘non-event based’ modes of analysis for investigating racialized
hashtagging as a practice, working to exploit the affiliative aspects of social media data and
offering sociological insights into one of society’s fundamental concerns: race and racism.
The empirical findings of this study point to on-line strategies of racism denial being complex
and diverse. In this respect, they resonate with the off-line world - after all, racism is a social
phenomena which has existed long before the advent of the internet - though from the
methodological standpoint of our approach, can only be adequately grasped by taking into
account the technological affordances of the medium they circulate in. Otherwise, we are liable
to simply import existing understanding of racism denial and fail to comprehend that online
modes of communication are mutating practices of racism.
The project has relied on Chorus, a software suite for collecting and producing a range of
visualisations of Twitter data. Our methodological approach has avoided fetishizing
visualisations or treat them as the end-point of analysis. The endeavour has been to think with
visualisations as part of an analytic process - deploying visualisations rather than merely
viewing them. Furthermore, we have grounded our analyses in our acknowledgment of the
limitations and constraints of the software. Our socio-materialist approach has been a creative
process involving intuitive insight and critical reflexivity, in addition to acquiring knowledge of
the workings of visualisation and co-occurrence algorithms.
We have treated this research dually as a methodological enterprise and as an empirical project
that informs conceptual ideas about online racism, beyond existing linguistic and text-based
approaches. Our study responds to the question ‘What kind of techno-cultural assemblage is
put into motion when we express ourselves online?’ (Langlois, 2011), by exploring how modes
of racialization modulate and is modulated by the Twitter social media platform. We discovered
that variegated informational logics and multi-hashtagging practices materialize online
The study aimed to develop an original account of Twitter race-talk which demonstrates how
hashtags work for users. This has been achieved by analysing multi-hashtagging by focussing on
what purposes the practice of deploying more than one hashtag (i.e. #notracist plus one or
more hashtag) might hold for those doing it. The resulting data visualisations and analyses
suggest two principal modes of multi-hashtag usage. These modes are distinguished by their
different methods of doing hashtagging. Moreover, the two multi-hashtagging practices of
‘Humour’ and ‘Truth’ closely correlate to a complex, racially-charged ‘topical’ distinction.
Deploying visualisations and interrogating algorithmic data processes − and our consequent
depiction of the process of doing this work − is not trivial or irrelevant to sociology’s
programme. Rather, it reveals how such processes may come to make Digital Sociology a
feasible and fruitful task for social research.
We thank the British Academy/Leverhulme Small Grant Research (2012-13) for funding this
research project. Chorus development was supported in part through the MATCH Programme
(UK EPSRC grants GR/S29874/01, EP/F063822/1 & EP/G012393/1).
Adkins, L. & Lury, C, 2009. Introduction: What Is the Empirical? European Journal of Social
Augoustinos, M. & Every, D. 2007. The Language of “Race” and Prejudice: A Discourse of Denial,
Reason, and Liberal-Practical Politics. Journal of Language and Social Psychology, 26(2):
Back, L. and Puwar, N. 2013. A manifesto for live methods: provocations and capacities, in Back,
L. and Puwar, N. eds. Live Methods, Wiley-Blackwell.
Baym, N. K. 2010, Personal Connections in the Digital Age, Oxford: Polity.
Billig, M. (1988). The Notion of “Prejudice”: Some Rhetorical and Ideological Aspects, Text, 8,
Bonilla-Silva, E. 2010. Racism without Racists: Color-blind Racism and the Persistence of Racial
Inequality in the United States. Rowman and Littlefield
boyd, d. Golder, S. & Lotan, G. (2010). Tweet, Tweet, Retweet: Conversational Aspects of
Retweeting on Twitter. In: System Sciences (HICSS), 2010 43rd Hawaii International
Conference on. pp. 1–10.
Brock, A. 2012. From the Blackhand Side: Twitter as a Cultural Conversation. Journal of
Broadcasting & Electronic Media, 56(4), pp. 529–549.
Brooker, P., Barnett, J., Cribbin, T. and Sharma, S. .2015. Have We Even Solved the First 'Big
Data Challenge?': Practical Issues Concerning Data Collection and Visual Representation for
Social Media Analytics, in H. Snee, C. Hine, Y. Morey, S. Roberts and H. Watson eds. Digital
Methods for Social Sciences: An Interdisciplinary Guide to Research Innovation. Basingstoke:
Bruns, A. & Burgess, J. E. 2011. The Use of Twitter Hashtags in the Formation of Ad Hoc Publics.
Available at: http://eprints.qut.edu.au/46515 [Accessed December 13, 2013].
Card, S., Mackinlay, J. & Shneiderman, B. 1999. Readings in Information Vizualisation: Using
Vision to Think. London: Morgan Kaufmann.
Callon, M., Courtial, J. P., Turner, W.A. and Bauin, S. 1983. From Translations to Problematic
Networks: An Introduction to Co- word Analysis. Social Science Information, 22(2), pp. 191-
Daniels, J. 2009. Cyber Racism: White Supremacy Online and the New Attack on Civil Rights
Rowman & Littlefield.
Daniels, J. 2012. Race and racism in Internet Studies: A review and critique. New Media &
Society, 15(5), pp. 695 –719.
Danowski, J. A. 2009, Inferences From Word Networks in Messages, in Krippendorff, K. and
Bock, M.A. eds. The Content Analysis Reader, London: Sage.
Eliasoph, N. 1998. Avoiding Politics: How Americans Produce Apathy in Everyday Life,
Cambridge: Cambridge University Press.
Franzosi, R. 2010. Quantitative Narrative Analysis. London: Sage.
Garcia Esparza, S., O’Mahony, M.P. & Smyth, B. 2010. Towards Tagging and Categorization for
Micro-Blogs. Available at: http://irserver.ucd.ie/handle/10197/2517 [Accessed December
Goldberg, D.T. 2009. The Threat of Race: Reflections on Neoliberalism. Oxford: Willey-Blackwell.
Huang, Y., Basu, C. & Hsu, M. K. 2010. Exploring Motivations of Travel Knowledge Sharing on
Social Network Sites: An Empirical Investigation of U.S. College Students. Journal of
Hospitality Marketing & Management, 19(7):717-734.
Kick It Out (2015) Case Study: Kick It Out, April 2015. http://www.kickitout.org/kick-it-out-
Langlois, G., 2011. Meaning, Semiotechnologies and Participatory Media. Culture Machine, 12,
pp. 1–27. Available at:
[Accessed December 1, 2013].
Lewis, P., Newburn, T., Taylor, M., Mcgillivray, C., Greenhill, A., Frayman, H., & Proctor, R., 2011.
Reading the Riots: Investigating England's Summer of Disorder.
December 3, 2013].
Liu, Y., Kliman-Silver, C. and Mislove, A. 2014. The Tweets They Are a-Changin’: Evolution of
Twitter Users and Behavior, AAAI Publications, Eighth International AAAI Conference on
Weblogs and Social Media.
Lupton, D. 2014. Digital Sociology, London: Routledge.
Marres, N. & Gerlitz, C. (2015) Interface methods: renegotiating relations between digital social
research, STS and sociology. The Sociological Review. http://dx.doi.org/10.1111/1467-
954X.12314. [Accessed December 5, 2015]
Marwick, A. E., and d. boyd. 2011. “I Tweet Honestly, I Tweet Passionately: Twitter Users,
Context Collapse, and the Imagined Audience. ” New Media & Society 13 (1): 114 – 133.
Meddaugh, P.M. and Kay, J. 2009. Hate speech or “Reasonable Racism?” The Other in
Stormfront, Journal of Mass Media Ethics: Exploring Questions of Media Morality, 24, 4
Murthy, D. 2012. Twitter: Social Communication in the Twitter Age. Cambridge, UK: Polity.
Nakamura, L. and Chow-White, P. eds. 2012. Race After the Internet, London, Routledge.
Pérez, R. 2013 Learning to make racism funny in the 'color-blind' era: Stand-up comedy
students, performance strategies, and the (re)production of racist jokes in Public, Discourse
Society, 24: 478-503
Potter, J. and Wetherell, M. 1987. Discourse and Social Psychology: Beyond Attitudes and
Behaviour. Newbury Park, CA: Sage.
Roversi, A. 2008. Hate on the Net – Extremist Sites, Neo-fascism On-line, Electronic Jihad,
Ross, S. nd. Hashtags, Algorithmic Compression, and Henry James’s Late Style (Draft).
http://www.academia.edu/attachments/30315675/download_file [Accessed December 5,
Picca, L.H. and Feagin, J.R. 2007. Two-Faced Racism: Whites in the Backstage and Frontstage,
Saldanha, A. 2007. Psychedelic White: Goa Trance and the Viscosity of Race, London, University
of Minnesota Press.
Sharma, S. 2013. Black Twitter? Racial hashtags, networks and contagion. New Formations: a
journal of culture/theory/politics. pp. 46 - 64.
Shifman, L. and Blondheim, M. 2010. The medium is the joke: Online humor about and by
networked computers. New Media & Society, 12( 8), pp. 1348-1367.
Sue, D. W. 2010. Microaggressions in everyday life: Race, gender, and sexual orientation. John
Wiley & Sons.
Vander Wal, T. 2005. Explaining and Showing Broad and Narrow Folksonomies.
http://www.vanderwal.net/random/entrysel.php?blog=1635. [Accessed December 5,
van Dijk, T.A. 1992. Discourse and the Denial of Racism. Discourse & Society, 3(1), pp. 87–118.
Zapavigna, M, 2011. Ambient Affiliation: A Linguistic Perspective on Twitter. New Media &
Society, 13(5), pp. 788–806.
Michele Zappavigna, M. (2015) Searchable talk: the linguistic functions of hashtags, Social
Semiotics, 25:3, 274-291
Although the Demos study discovered that some slurs are used in a non-derogatory manner
aimed at a sender's own community.
It is beyond the scope of this article to explore the how radicalized hashtags are produced
within Twitter in relation to its range of techno-cultural assemblages (i.e. as part of a wider
sphere of internet activity involving other social media services, online video or audio clips web
browsers and URLs and so on, all of which may feature).
We do not claim to have captured a complete dataset of all tweets containing the #notracist
hashtag during the time-period, because collecting data from the Twitter Search API is rate
limited (number of search requests per 15 minute interval). Nonetheless, as the frequency of
notracist tweets were relatively low, it is likely we captured a comprehensive set of tweets.
See the Chorus project website for further details and to download the software:
Our intention in introducing the timeline graph is to demonstrate how this visualisation
facilitated the decision to pursue other modes of analysis.
All tweets have been anonymized, both in terms of their user names and the tweet content
itself. Where URLs feature in tweets, key identifying characters are changed to "*".
The single significant display of communication - where the @mention convention (boyd,
Golder & Lotan, 2010) is used to directly address other Twitter users - is visible in some Twitter
users retweeting messages considered as containing racist content to the account
@YesYoureRacist. This account publishes tweets which claim to be not racist yet appear to
feature a racist statement of some kind.
Noortje Marres and Caroline Gerlitz (2015: 9) offer an important discussion of how digital
sociology methodologies are innovating forms of co-occurrence/word analyses which render
‘text amenable to network analysis, whereby empirically occurring associations among words in
a given data set provide an immanent criterion of relevance’. See also the work of Roberto
Franzosi (2010) for developing inductively-orientated quantitative textual analyses of large
As an aid to analyze the cluster map of Figure 2, the two radials have been added to the
Chorus visualization by the researchers.
Table 1 explores each radial in turn and noting key hashtags down to a minimum frequency
of 20 usages.
Common usage terms such as ‘like’ and ‘just’ have been included in the dataset to indicate
their relative frequency in relation other more charged terms such as ‘Black’ and ‘White’. As the
research focus was not on analyzing the content of tweets, only a limited ‘stop-list’ of common
words was used in the analysis (that exclude terms such as ‘a’, ‘the’, ‘and’ etc.).
It is interesting to note the multimodality of social media and internet usage for Twitter
users, which features as part of the creation of their own internet assemblages as part of a
broader field of activity: Twitter users do not just use Twitter to do their tweeting. It was not
within the scope of the research project investigate the content of URL (links) within tweets.
#loop refers specifically to videos posted on Vine, which are six seconds long and indefinitely
looped such that they repeat until the viewer moves on to the next one or closes the
Chorus computes collocations of terms, with co-occurrence values from 0 to 1 based on the
relative frequency with which those words occur together in single tweets. The co-occurrence
value is the probability, local to the dataset, of finding two terms occurring together in a tweet
(where 0 equates to zero probability and 1 signifies absolute certainty).
To make such a claim does not the beget an analysis exploring the meaning of humour-based
tweets. Rather, it points to ‘meaning’ being located in the hashtags; and only exploring these
operators semantically is a limited mode of analysis of a Twitter racialized assemblage.