A picture is worth a thousand (s)words: classification and diffusion of memes on a partisan media platform

By Esteban Villa-Turek, Rod Abhari, Mowafak Allaham, Chloe Mortenson & Ayse D. Lokmanoglu


On November 3, 2020, as presidential votes were being cast around the United States, another form of political discussion was taking place on Parler, a far-right social media platform. Users had been using memes to shape the discussion around political events by employing an extremist out-group rhetoric, usually mocking racial or gendered attributes of adversary political figures. That day, it was the turn of President Barack Obama and former First Lady Michelle Obama, but they were not the first, and certainly not the last. In today’s hyper-connected online public spheres, the circulation of dangerous and extremist narratives has reached unprecedented levels. While years’ worth of research has focused on the role of written discourse in shaping online extremist conversations, we have devised a novel approach that employs algorithmic detection to discover the non-discursive means that extremist actors employ in online fora, particularly through the creation, modification, and circulation of memes.

But what exactly is a meme? Limor Shifman defines internet memes as “a group of digital items sharing a common characteristic … created with awareness of each other … circulated, imitated and/or transformed via the internet by many users.” In other words, memes are inter-related digital artifacts that are meant to be created, recreated, and circulated across digital fora with the aim of shaping the narratives of such online spaces. Therefore, a meme is not only a symbolic representation of digital culture but also an evolving digital object that shapes online discourse.

However, there has been a proliferation of novel kinds of extremist, supremacist memes circulating across social media platforms. Integrating the meme characteristics highlighted in previous studies, our report defines supremacist memes as participatory digital objects and persuasive messages that shape public discourse and evoke motivations of activism. As we discovered, this is particularly true in moderate and partisan alternative platforms targeting general audiences, such as Parler.

Our research introduced a mixed‑method toolkit that facilitates the analysis of memes found on alternative platforms, tested on a publicly available dataset of Parler posts collected between August 2018 and January 2021. The dataset consisted of 183.1 million Parler posts made by 4.08 million users. To extract memes from Parler posts, all URLs to images were retrieved from the metadata fields of the 183.1 million posts. Only 3% of the posts had images (6,626,622 images). Approximately 92% of these images (6,148,168) are no longer available or accessible. Out of the remaining 478,454 images, 2.7% (13,225) of these images were excluded from our analysis because the files were distorted. The total number of remaining images included in the analysis was 465,229 images.

The complexity of identifying memes

We used a combination of computational and qualitative methods. The former allowed us to group images thematically into image clusters based on content similarity, while the latter enabled a detailed examination of the generated clusters to determine the types of topics present in the most engaging meme clusters.

To analyse and label the content of each cluster, we selected only clusters that fell within our definition of memes, in other words those that contained digital images and text created to be circulated with the purpose of amending or altering meaning in a message. We manually coded the meme content of the top 25 clusters that generated the most engagement from each of the years 2019, 2020 and 2021, for a total of 75 clusters. We then labeled the clusters according to the presence of themes of gender, race, violence and partisanship content in each of them.

Intersectionality, violence and engagement

We found that the intersectional topics of gender and race were the most violent in terms of imagery, while partisanship‑only topics had the least violent imagery. Moreover, memes with intersectional themes of gender and race combined with partisanship had the highest virality and diffusion, as with the prominent case of President Barack Obama and former First Lady Michelle Obama. This means that, as content creation and circulation follow the trend of increased violence in memes about gender and race, the public discourse within alternative platforms will potentially become more violent in the intersection of gender and race themes, a source of great concern.

In other instances of gendered violence, appeals to diverse forms of masculinity often included the depictions of the outgroup with feminine qualities to demonstrate presumed weakness. This finding illustrates how masculinity appeals are not only aimed at hyper, hegemonic, and toxic masculinity, but also to specifically target women, vis-à-vis perceptions of what constitutes weakness or a ‘lesser person’. To exemplify this around the Obamas, memes with Michelle Obama were photoshopped as Time’s Man of The Year, and Barack Obama was photoshopped into a ‘pussy hat’, an icon from the Women’s March on Washington, but more specifically a gendered insult in that moment’s electoral context.

Our results indicate that no direct correlation exists between the number of memes in a topic cluster and engagement levels. In fact, the meme clusters that had the highest levels of engagement and diffusion correlated with events on the ground (such as elections, regulation changes, and platform censorship) and with individuals. Moreover, the degree of violence within memes increased between mid‑2020 and January 2021, and the transmission rate of violent memes also increased. This illustrates the importance of identifying and monitoring trends within memes, as it was possible to see violence within the public discourse increasing parallel to the events and build up towards the events of January 6th, 2021.

Overall, these findings indicate that Parler memes sort into three categories: memes with gradual but steady attention patterns, memes with sharp attention patterns and memes with little attention change over time. As a mode of cultural discourse, memes allow online audiences to position themselves in relation to current events. The high fluctuations in the engagement of most meme topics indicate that meme topics are generally event‑based and can be partly described through external patterns of influence, including seasonal events and specific news items. In comparison, memes with consistent attention tended to represent the persistent interests of site users. As a site with an apparent right‑wing ideology, it is not surprising that these memes tended to focus on partisan agendas in general and Donald Trump in particular, with the persistence of specific meme formats like founding fathers’ quotes, which represented motifs that resonated with both site users and American conservatives alike.

Ultimately, memes are a ubiquitous form of online engagement that may motivate extremist upheavals, such as the January 6th insurrection. Our research effectively shows how machine learning can be used to identify the content characteristics of memes and trace the prevalence of problematic content over time. Naturally, meme proliferation tends to echo political happenings, and the content of these memes tend to violently target perceived weaknesses of the out-group. The extremism and virality shown by memes against the Obamas during the presidential election of 2020 is a testament of the nugatory consequences of online supremacist narratives.

Ayse D. Lokmanoglu is an assistant professor in the Communication Department at Clemson University, and a member of the Media Forensics Hub. Her work takes a mixed methods approach that integrates computational methodologies and critical cultural theory to examine information campaigns in digital media associated with racial, gender, and religious supremacy.

Rod Abhari analyzes the diffusion of scientific (mis)information across online platforms. In doing so, he aims to both improve the accuracy of scientific reporting and combat the spread of science-based populism. He holds an M.A. in Communications from the University of Wisconsin – Madison and an M.A. in Science and Technology Studies (STS) from the University of Maastricht in the Netherlands.

Mowafak Allaham is a PhD student in Northwestern’s Technology & Social Behavior program interested in the areas of computational communication and computational social science. He combines methods from computer science and statistics with theories from social sciences and communication studies to study the various forms of public engagement with news media and its implications online and on society.

Chloe Mortenson is a PhD student in the Media, Technology and Society program, working in the Center for Communication and Public Policy with Dr. Erik Nisbet. Her research is focused on political communication and comparative politics, specifically on the relationship between media ecosystems and regime types, public understanding of democracy, misinformation, and political violence. Chloe received her M.A in Communication Studies at The Ohio State University, and her B.A at Duquesne University in International Relations and Communication Studies.

Esteban Villa-Turek studies online information diffusion processes with special attention to their policy implications, focusing on political and scientific misinformation in Latin America and the Global South with the help of social network analysis techniques and applied computational methods. He holds a bachelor’s degree in law from Rosario University (Bogotá, 2014), a master’s degree in public policy from the Hertie School of Governance (Berlin, 2019), a Master in Communication (Northwestern University, 2022) and is currently finishing a Master in Analytics from Georgia Tech.

Image Credit: PEXELS