Engineering Inflammatory Content
A Memetic Analysis of Russian Social Media Propaganda
RTF 368S - MEDIA STUDIES THESIS
Radio - Television - Film
The University of Texas at Austin
July 12, 2019
Sharon Strover, Ph.D.
Philip G. Warner Regents Professor of Communication
Director, Technology & Information Policy Institute
Moody College of Communication | The University of Texas at Austin
The Russians sourced content already proven to be popular among certain demographics, from hashtags and sites already popular among those demographics, edited them to convey their target messages, and spread them through Russian owned parody pages targeting the same demographics said content was sourced from. They provided minimal funding to spur their propagation, in order to propagate those images in a peer to peer fashion (memetically).
"One of the downsides of spreadability is that bad information may spread just as fast and as widely as good information. In some cases, because misinformation can be constructed to provoke people's fears or gratify their prejudices, it provides better cultural currency than more accurate and nuanced information. This is why we all need to take greater ownership over the information that we circulate. We need to insure its accuracy before we circulate it. We need to weigh its potential negative impact before we insert it into our conversations. We must insure the integrity of our information flow."
- Henry Jenkins for Engineering Inflammatory Content
Memetic Distribution vs. Viral Distribution
An often debated dichotomy within internet culture is the difference between content “going viral” or “becoming a meme.”
Therein lies a significant difference between "viral distribution" and "memetic distribution".
Key Differences between Memetic Distribution and Virality:
- A Single Cultural Unit
- Limited to one digital medium (eg, a viral video)
- Rapidly expanding popularity on a single popular hypermedia platform (“this video reaches 3 Billion views on YouTube)
- Content rapidly reaches pop culture status and is widely known; can quantify maximum popularity by some generally available metric easily (views, likes, retweets, sales)
- Generally distributed through hyperlinks and word of mouth; the content itself is generally has too large of a file size to be shared in a peer-to-peer manner as attachments and must be hosted on a platform for wide access (“I’ll send you a link to this video; Go look up “Gangam Style” on YouTube”)
- Always a collection of texts, involving multiple cultural units
- Content often takes advantage of an absurdist misuse of a social networking feature; often exploits the medium of distribution to increase psychovisual impact of the content in a humorous manner “Do you know da wae” “Storm Area 51, They Can’t Stop All Of Us” Spongebob Squarepants, The Communist Manifesto on Pornhub.com
- Expands upon convergent media; the meme can be originated, conform to, or be extended by many different hypermedia and traditional media formats, including but not limited to Facebook Events, YouTube Videos, Tweets, Profile Pictures, 3D Modeling software, books, print, and audio/video.
- Values convenience over quality; degrade in quality often used as a cultural embodiment of the subculture and a view of identity and understanding
Three key lenses for analyzing memetic internet content:
- Access - what groups are circulating which content? Was the media distributed via a link, by sharing the actual image, or other? Was there a paywall for access?
- Attribution - Who originally created the meme? Who posted the meme? Who took credit for the meme?
- Ownership - What platform is the meme present on, who has monetized the distribution, who has control of the deletion of the post?
1. Research Questions:
How did the Internet Research Agency effectively strategize, source, and deploy deceptive social media campaigns tasked with distributing misinformation, while maintaining anonymity?
Where did the images used to compose the Russian ads circulate previously to being used as a Facebook ad?
Were the sources of the images intertextually related to the demographics targeted by the ads they were contained within?
In order to analyze the memetic propagation of the Russian Ads dataset, I utilized node-mapping application Maltego XL as a host environment for running the analysis, and reverse-image search service Tineye to search the web.
The U.S. House of Representatives Permanent Select Committee on Intelligence released a dataset of 3,500 Facebook ads in PDF format, downloadable at intelligence.house.gov with direct links reproduced below.
irads, available on Github from the Maryland Institute for Technology in the Humanities, parses the original U.S. House of Representatives PDF's into structured data, available in JSON and CSV format. This structured dataset was the starting point for my analysis.
In order to import this data into Maltego XL, which only allows for the import of properly structured CSV files, I restructured the irads parsed dataset partially, most significantly sorting the objects in ascending order by Ad ID.
For each ad, the dataset contains the following fields:
id image title description facebook_url impressions clicks created ended cost currency location residence match interest behavior politics multicultural_affinity employer industry field_of_study exclude language age placement
In order to properly store this data in Maltego XL, I created a Maltego Entity entitled "Russian Facebook Ad" (mitchaiet.RussianFBAd). This allowed each ad and it's associated metadata to be stored as a single node within Maltego XL. The entity is included as part of this project's repository. Download the entity below and import it into Maltego XL if you are repeating this process.
In order to facilitate the quality of the search results I found, I manually cropped each source image to remove the Page Name, Ad Description, and Reactions. Every cropped image is available in the repository.
Here is Ad ID 1 as an uncropped source image:
Here is Ad ID 1 as a cropped image:
The input for each Tineye search was the cropped image.
In order to run reverse image searches via the Tineye API, I created a Maltego Transform which allowed me to select a Russian Facebook Ad entity, and perform a search using the cropped version of the image.
The Tineye reverse image search results were returned to Maltego XL as nodes connected in a directed manner to the source Russian Facebook Ad entity. Each Tineye Result node contains the following data:
- a count of the total number of matches
- a direct URL to each matching image
- the URL of the page where each match was found
- if the source is a collection or stock site
- the score for the individual match
- a link to the ‘compare images’ overlay image
Conveniently, if two discrete ads returned a Tineye match to the same exact URL and Crawl Date, both Russian Facebook Ad entities will be linked to the same Tineye Result node.
In order to visualize the results of the searches outside of Maltego XL, I export selections from the large map using Maltego XL's "Export Graph to Table" function. I exported each selection using the option "Entity property flat map (machine readable)" which produces a formatted CSV file. I created a data flow using parabola.io in order to automatically format the Maltego XL formatted CSV into a separate CSV file formatted for import into node visualization software Kumu which I initially used to create the embedded visualizations for this report. On May 26th, 2020, the data was exported from Kumu in favor of recreating the node maps in D3.js for the sake of preservation, optimization, and portability.
The dataset produced three patterns of Russian memetic content distribution:
Propagation of unedited, memetically circulated images by sourcing them from popular, thematically linked hashtags concerning a particular demographic, and presenting those images through Russian-owned pages to the same source demographic in order to court viewership and establish brand trust.
Propagation of edited, memetically circulated images by watermarking the content with a specific Russian-owned brand, then circulating the content both through sponsored ads and in peer-to-peer groups in order to subvert those demographics. (why you can find blacktivist content on me.me)
The creation of original content (colloquially referred to as “OC” in online communities) using internet-sourced ingredient assets and using formatting styles typically associated with memetic content.
The IRA effectively weaponized Poe’s Law to build channels of distribution which propagated inflammatory misinformation to susceptible audiences. These patterns follow the following steps:
Produce a Target Demographic
Research the Target Demographic
Create a manifest of content sources which the target engages with passionately Extracting content from verified sources provides credibility to fabricated sources when compared side by side (“this page of unknown ownership is showing me the same content as Buzzfeed, it must be credible”)
Extract text and images from those sources Creating remixed content from publicly available sources reduces paper trail (not having to buy stock photos, or attribute photographs)
Create state-owned original content and distribution channels mimicking original content and distribution channels (pastiche/parody)
Seed a target demographic using paid ads Sponsored posts forced into news feed interpolates seed content alongside native posts Poe’s Law describes the act of mistaking seed content for native content Ads ran for 24 -72 hours at a time, similar to how inflammatory Twitter accounts stay up long enough to be screenshotted
Recruit audience members to that page Convert impressions into traditional ad clicks and page likes They built brand identity and audience through memetic distribution; i.e, a more valuable goal was for users to screenshot the image, share it to their own pages, and convert their followers by promoting the branded content in a peer-to-peer fashion
Promote misinformation through “severed seeding” Once an audience member is part of a community, discredit factual/mainstream media outlets, promote conspiracy and skepticism of other channels; Call to action: Spur audience members to seek out their own information and independent research while simultaneously discrediting proper outlets “check everything you read” “do your own research” “take the red pill” “down the rabbit hole”
Memetic Propagation Maps
The following maps show Russian Facebook Ads and their respective Tineye reverse image search results. Russian Ads are highlighted in Red, and Tineye results are highlighted in green. Click on each node for metadata information.
The Russians sourced the headline and two images from a black-focused Buzzfeed article:
These Stunning Women Are Shutting Down A Ridiculous Beauty Double Standard
They then created a meme out of the images:
and ran it as a sponsored ad, under a Russian-owned and branded Facebook Page "Black Matters (BM)" targeting Facebook users with an interest in the following terms:
- Martin Luther King
- African-American culture
- African-American Civil Rights Movement (1954‚1968)
- African-American history
- Malcolm X
AD ID's 2575, 2593, 2595, 2605, 2608, 2616, 2631, 2643, 2682, 2697, 3242, 3243 were all run under the Facebook Page "Brown Power", targeting hispanics. Here we can see that multiple images used in ads were sourced from the hashtag #BrownBerets on Instagram.
AD ID's 952, 2118, 2003, 1131, target black-related topics all link back to #africanamericans on Instagram.
American-Hispanic Intertextual Clusters: A network of American-only focused ads (491, 26, 1928, 2793, and 21) and a network of Hispanic-only focused ads (2626, 2624, 2667, 2584, 2736, 2666, 3244 + 2679) share a pattern of propagation through a single ad containing both American and Hispanic themes (2667). This produces an intertextual link between both demographics, as proven by the propagation of the image content eventually sourced by the Internet Research Agency.
Russian Ads by Social Network
The following maps show Russian Facebook Ads whose content linked to various posts, users, and communities on major social networks. Russian Ads are highlighted in red; hover over them to reveal their respective results.
Note: Certain social networks prevent direct scraping by web crawlers. Notably, 4Chan deletes all posts after 24 hours, and Instagram is a closed community which does not allow for direct scraping. Therefore, maps for these communities are based on links to service-specific archival and proxy sites, like imgrum.org and 4archive.org.
Virality: “a word-of-mouth-like cascade diffusion process wherein a message is actively forwarded from one person to other, within and between multiple weakly linked personal networks, resulting in a rapid increase in the number of people who are exposed to the message.”
- (1) a person-to-person mode of diffusion;
- (2) great speed, which is enhanced by social media platforms; and
- (3) broad reach, which is achieved by bridging multiple networks.
Link Rot: Link rot (or linkrot) is the process by which hyperlinks on individual websites or the Internet in general tend to point to web pages, servers or other resources that have become permanently unavailable.
Meme Page: An internet account, page, group, or forum, specifically dedicated to sharing relevant internet memes, generally based around a certain theme, topic, or performance.
Memetic Information: user generated content created in a production environment without any need for fact checking or journalistic verification standards. It is propagated in a peer-to-peer manner through mediated channels such as group chats, comment threads, or user-uploaded video, and is propagated to the end viewer through a trusted peer source - like a friend, relative, or mutual chat/group member.
Memetic Distribution: the act of propagating a meme to various persons or groups via strong and weakly linked social networks in a peer-to-peer manner; the physical file is the one passed around, not just a link, usually through direct transmission of a file without alteration (texting the image to someone, screenshotting or screen recording and sending the file
Memetic Abstraction: When a screenshot is taken of a post or article and distributed memetically, the content continues to spread while attribution to the original source URL is severed. When a screenshot of a news article is taken and propagated eslewhere online, the mechanisms tracking the URL for the source article are blind to the propagation of the secondary image version.
Intertextuality: the shaping of a text's meaning by another text. It is the interconnection between similar or related works of literature that reflect and influence an audience's interpretation of the text. Intertextuality is the relation between texts that are inflicted by means of quotations and allusion. Intertextual figures include: allusion, quotation, calque, plagiarism, translation, pastiche and parody. Wikipedia
Diegetic UI Elements: The image contains embedded UI elements within the image which meet the following criteria:
- (1) The UI elements are non-interactive
- (2) They are not a critical or functional element of the image, it's messaging, or it's construction
- These are commonly found in smartphone screenshots, where status bars and app elements can be found along with the post, article, or image being screenshotted.
Artifacts of Distribution: visual embodiments of memetic propagation, including but not limited to crop edits, watermarks, changes in compression, filtering, attempted removal of identifiable information (scratching out usernames or faces in a screenshot). These artifacts pile onto memes in an archaeological manner.
Evergreen Media: the act of resurfacing available internet media posted previously
Deep Fried Meme: a style of meme wherein an image is run through dozens of filters to the point where the image appears grainy, washed-out, and strangely colored
Screenshot: an image that shows the contents of a computer display
ScreenLife: a storytelling framework for the creation of content, generally among the mediums of videos and film, which tells a complete narrative within the confines of a digital screen utilizing apps, websites, software functions, and internet-connected experiences. It’s main idea is that everything that the viewer sees happens on the computer, tablet or smartphone screen. All the events unfold directly on the screen of your device. Instead of film set — there’s a desktop, instead of protagonist’s actions — a cursor. ScreenLife content is created entirely within the confines of a digital device, often mirroring the screen the protagonist views.
F-Shaped Pattern for Reading Web Content: Eyetracking visualizations show that users often read Web pages in an F-shaped pattern: two horizontal stripes followed by a vertical stripe.
Poe's Law: an Internet axiom which states that it is difficult to distinguish extremism from satire of extremism on the Internet unless the author clearly indicates his/her intent. This notion is most frequently observed with highly polarized discussion topics, such as gender equality, religious or political fundamentalism and other social justice-related issues.
Morgan's Corollary to Godwin's Law: As soon as such a comparison occurs, someone will start a Nazi-discussion thread on alt.censorship.
DarkShikari's Theorem: Any community that gets its laughs by pretending to be idiots will eventually be flooded by actual idiots who mistakenly believe that they're in good company.
Emoji: pictographs (pictorial symbols) that are typically presented in a colorful form and used inline in text. They represent things such as faces, weather, vehicles and buildings, food and drink, animals and plants, or icons that represent emotions, feelings, or activities. To the computer they are simply another character, but people send each other billions of emoji everyday to express love, thanks, congratulations, or any number of a growing set of ideas.