The Executive Summary of this report is also available in Arabic and French.
Memes have become an integral tool of communication in partisan online spaces. Internet memes are defined 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.” There has been extensive research on the impact and properties of misinformation, as well as methods for platforms to moderate misinformation. However, moderation policies and algorithms capable of filtering out meme content with extremist foundations are lacking. Moderating supremacists’ memes through algorithms is difficult as images are not as easily detectable as text. Moreover, memes often embody inside jokes among users or have an unassuming image form that hides political intent. Building on prior research on identification of terrorist imagery, this report explores algorithmic approaches to classifying harmful meme content and exploring their patterns of diffusion.
Using state-of-the-art deep learning image and visual rhetorical analysis, we examine memes on an alternative (or fringe) platform by categorising them into themes of gender, race, partisanship and violence. This method further reveals the transmission rates of the memes associated with these themes. We propose a unique methodology that combines automatic image clustering with network analysis, developing a framework to compare the transmission rates of memes at different timepoints. In so doing, we provide experts with a working model of meme content filtering to help platforms to identify and filter memes with supremacist topics; the model also allows for the testing of image attributes to aid the development of a toolkit to understand which memes are spread in alt-tech platforms.
The key findings are:
- There is no direct correlation between engagement and the number of memes in a topic cluster.
- Memes with intersectional themes of gender and race with partisanship had the highest virality and diffusion.
- In 2020, the five meme clusters with the highest impact factors were Climate Change, George Soros, Pro-Trump 1, Pro-Trump 2, and Michelle Obama. By way of comparison, in 2021, five clusters on Maga and Soldiers, Against Political Lobbying, Children and Gender, Leftists, and Missing Votes had the highest impact factor.
- The degree of violence and the transmission rate of violent memes increased, starting mid-2020 up to January 2021.
- Memes with high engagement levels often were branded by a group emblem or logo.