Lingua Ferro Iter: Insights Gained Through Linguistic Analysis of Iron March
Rex Bray III and Jacques Singer-Emery
Over the past several years, far-right extremists have leveraged online platforms ranging from social media to Stormfront to congregate with one another, convert new members, and concoct violent plans. They have been so effective that the FBI recently elevated “to top level priority racially motivated violent extremism so it is on the same footing in terms of our national threat banding as ISIS and homegrown violent extremism.”
Among the most extreme and violent of these websites was IronMarch.org, a forum/message board founded in 2011 by a Russian user calling himself Slavros. According to the Southern Poverty Law Center, Slavros exercised tight control over the website, and he was certainly responsible for its direction: he alone authored 11% of the public posts and initiated 24% of the private conversations. Iron March slowly waxed in popularity, reaching over 1,200 regular users, before it was taken down in November 2017 following repeated hacks. Two years later, an unknown individual publicly leaked a SQL database containing the entirety of IronMarch.org.
The final front page of IronMarch.org. Compiled from the SQL database by ZDNet.
What made Iron March in particular so dangerous was that its influence extended beyond the cyber realm: The website was either “affiliated” with or “supported” at least nine real-world neo-Nazi groups spread all over North America and Europe. These include National Action, a British group banned after glorifying and supporting the assassination of Member of Parliament Jo Cox; the Atomwaffen Division, an American group of which several members have been charged or convicted of plotting terrorist attacks; and the Antipodean Resistance, an Australian group using the slogan “We’re the Hitlers you’ve been waiting for.” The latter two groups were born on Iron March, a testament to the platform it provided white supremacists.
We analyzed Iron March’s public posts and private messages in an attempt to understand how and why they, as a group, thought the way they did. On Lawfare we described the “average” Iron March user, the various factors that drove activity on the website, and how effective it was at connecting far-right extremists. Here we explore (1) how was influence distributed on the website? (2) Why did some users become content leaders, while others almost never posted? And finally, (3) what insights can be gained from the way users and content leaders wrote their posts and messages?
How was Influence Distributed on Iron March?
We profiled the “average” Iron March user in the aforementioned Lawfare piece; to summarize, 52% of the website was American, another 20% came from other English-speaking countries, and as with most of the Internet, English was the lingua franca for the rest of the users (while normal, it does contrast with Iron March’s successor forum, Fascist Forge, where the lingua franca is German). Here, we sought to understand how influence and content leadership were distributed on the website. Based on their public posting activity, users fell into one of four distinct tiers.
The top 5% of posters, wielded the most influence over the site’s content by far. They authored 54% of public posts, posting approximately 1,838 times each. They contributed slightly more of the content based on word count, at 57%, writing 128,000 words each. For context, the Jane Austen novel Sense and Sensibility clocks in around 126,000. Slavros was an obvious outlier writing almost 21,000 posts, accounting for over two million words (approximately three times the King James Bible). But even if Slavros is separated from the top tier, the remaining top 5% was still more influential than the next tier, and has almost as many posts as the other three tiers combined. Among the notable tier 1 users was Brandon Russell, the founder of the Atomwaffen Division, who has since been convicted of federal explosives charges.
Tier 2, the next 15% of posters, were also very influential, composing 33% of posts and 30% of the content. Combined with tier 1 they authored 87% of both the posts and words on Iron March. Tier 3, defined as all users in the bottom 80% who posted at least 10 times, was responsible for 12% of the activity and of the verbiage. Finally, tier 4, the users who posted less than 10 times, was responsible for less than 1% of posts and less than 2% of total words. The following table contains descriptive statistics about the posting activity of each tier.
While tier 1 did exert a considerable amount of influence over Iron March’s content, that influence waned over time: As we noted on Lawfare, the percentage of posts coming from the top tier decreased from 55% in the early life of the website to 15% in its final months. This change came primarily as a result of the website growing in popularity amid a surge in the number of American users. Over the site’s lifespan, the number of users posting in any given quarter grew almost ten-fold, and the posts-per-user fell by over 60%. For the first five years of the website’s life, that top user was always Slavros, but in the last five quarters, four different users held that dubious honor, and Slavros only once.
Long live IM, praise Slavros, Hail victory.
-- American tier 3 user, 9/15/2016.
It is curious that the private messages were much less concentrated: while Slavros was responsible for starting 24% of them, no other individual was responsible for any more than 1.5%. There were under 22,000 private messages, as contrasted with almost 196,000 public posts, but these differences in behavior can easily be explained. Iron March was a public forum, not a social media network.
But why was public posting activity on the website highly concentrated in a few individuals? Why did some of those who signed up post three or four times and leave, and others stick around and write novels’ worth of material? What was different about the content leaders, and did they bring that trait to the website or were they converted there? In the next section, we analyze the life cycles of the content leaders and what made their posting activity different.
Who Became a Content Leader and Why?
Based on posting activity, we concluded that tier 1 users were the content leaders on the website. Their life cycles, and especially as compared to those of other users, confirm that notion. Tier 1 users joined Iron March much earlier than other members: the median top tier poster had signed up and begun posting by quarter 2 of 2013, when only 174 people had ever posted on the website. By contrast, the median poster from the other three tiers first appeared in quarter 3 of 2016 – during the surge of Americans to the website. Furthermore, tier 1 users were more devoted to the website from the start. On average, they each posted 256 times in their respective first quarter. This contrasts with the others, who posted only 28 times in their first quarter.
In terms of longevity, the content leaders spent far longer on the website than any other tier: tier 1 users were on the website for approximately 12 quarters, whereas the other tiers, on average, spent less than 3 (for the purposes of this analysis, we counted 2011 and 2012 as ‘quarters’ due to the sparse activity during those time periods). At first glance, their longer ‘life spans’ might seem to explain why tier 1 users have so many more posts. However, they also posted far more frequently than did other users, at 158 times per quarter, as opposed to 28 for the others.
This pattern suggests that a core nucleus of extremely devoted users started the website, kept it going, and then later, there was a large influx of less devoted users. Over half of the users who joined in the first two years of the website’s existence became tier 1 and tier 2 users, whereas over half of the users who joined in the last year were tier 4. This occurrence makes perfect sense: early adopters are frequently the most devout, and only the people who are most dedicated to and most strongly believe in white supremacy would have likely found Iron March in its earliest days. This trend also aligns with the trend discussed earlier, of tier 1 users having less influence and more individual users posting in the forum, particularly around the time of the surge in Americans in 2016.
So, what was different about the users who joined earlier, posted more, and stuck around longer? And what was different about the users who showed up, posted less than ten times, and left? Our research suggests that the tier 1 users were significantly more likely to affiliate with real-world neo-Nazi groups than were tier 4. A primitive keyword search for “Atomwaffen” revealed that tier 1 users were 2.9 times more likely to discuss the violent group as were other users of the message board (of whom 11% used the name of the group), and 4.2 times more likely than were tier 4 users. This result comes even though tier 4 users were more likely to be American (where the Atomwaffen Division is located) and that tier 1 users made up a decreasing portion of the site’s posting activity starting fourth quarter 2015, when the Atomwaffen Division was formed.
Members of the Atomwaffen Division. Picture from Twitter, courtesy of the Daily Beast.
Knowing that tier 1 users were more devout and more likely to discuss real world white supremacist organizations, we wondered whether those traits translated into a difference in the rhetoric they used. To answer that question, we employed LIWC2015, a simple yet empirically robust linguistic analysis tool. It consists of a series of dictionaries, such as ‘past focus,’ ‘first-person personal pronouns,’ and ‘sad.’ The operator feeds a body of text into the program, and the program then counts the number of times words from each dictionary are used. In addition to the various dictionary frequencies, LIWC2015 outputs four summary variables: Analytical Thinking, Clout, Tone, and Authenticity. Analytical Thinking measures formal, logical thinking, and is known to correlate to college performance. Clout measures the relative social status of the speaker. Tone measures how positive or negative the speech is, with higher values indicating more positive speech. Finally, Authenticity measures how honest the author is being. LIWC2015 reports each as a percentile rank against the corpus of text on which it was trained. From that one can draw various conclusions about the author(s), which we do in the next section; here, we simply look for differences between the tiers of posters. We did that by comparing each tier’s summary variables and word frequencies within each dictionary.
The summary variables indicate that tier 1 users were more extreme and likely set the tone for the website, that tier 2 and 3 users acclimated to this tone, and that tier 4 users left before they could acclimate. To reach this conclusion, we first analyzed the entire corpus of text, learned what features distinguished the various tiers, and then compared those findings to the first posts each tier of users made.
Across the entire corpus of text, we observed several trends. Analytical Thinking was almost always slightly above average. Clout increased with tier (although Slavros scored only a 51st percentile. This actually makes since because he authored so much of the website’s content). Tone, for which a higher score indicates a more positive tone, was below average and decreased with tier – perfectly in-line with a website that thought Stormfront, the most infamous neo-Nazi forum, was too moderate. But interestingly, Authenticity decreased with tier, and tier 1 was in the 30th percentile. Low authenticity scores indicate that the authors used similar language to individuals engaging in deceptive writing. However, we were not able to discern exactly what the most active users were trying to hide.
Looking at Clout, Tone, and Authenticity shows that tier 4 users were significantly different from the other users. Specifically, Tier 4 users were 8 percentile points lower than tier 3 user on Clout, 8 percentile points higher on Tone, and 16 percentile points higher on Authenticity. Those results suggest they did not feel they had a lot of influence in this community, were more positive in their writing than most users, and likely wrote more honest posts.
However, tier 4 users posted very little: as discussed earlier, only four times on average, and 51 of them only posted once. Therefore, the disparity could be due either to tier 4 users being inherently different or to other users acclimating to the site’s extreme nature. To determine which, we looked at every user’s first post. Like the tier 4 users, the first post of users in tiers 2 and 3 showed lower clout, higher authenticity, and a more positive tone. However, this was not the case with tier 1 users. Compared to the other tiers, tier 1 users’ first posts scored significantly higher in clout (by 7 percentile points), significantly lower authenticity (7 by percentile points), and slightly lower tone (by 2 percentile points) than the others.
I actually came from SF, of course my view of NS has radicalized greatly ever since I joined this site.
-- American tier 2 user, 8/30/2017.
To summarize, upon joining the website, the data suggests that tier 1 users were already radicalized when they joined and were able to successfully set the overall tone for the site. Many less fanatical users acclimated to this as they became tier 2 and 3 posters, but individuals who fell into tier 4 did not and left the website.
What Insights Can We Gain from the Rhetoric Used by Iron March?
Analyzing the rhetoric from Iron March provides two insights: first, some of its users might be depressed and second, the level of “angry” speech this site exposed its users to.
Given the harshness and extremity of their rhetoric, we suspected that at least some members of the Iron March userbase exhibited signs of suffering from mental illnesses, and that the prevalence thereof may vary by tier. Specifically, we believed that while tier 1 users were the most extreme and violent of the group, that many of the lower tier users had joined this movement out of loneliness or hopelessness. Based on previous psychological studies that correlated diagnoses of mental illness to patterns in a subject’s LIWC2015 word use, we compared Iron March users’ language use to those of individuals with various mental illnesses. In some instances, we found that certain tiers of Iron March have writing patterns similar to individuals suffering from depression.
When we compared the rhetoric used by Iron March posters to known markers of depression, Iron March users only showed two similarities: negative emotion (e.g. “sad,” “upset”) and discrepancy (e.g. “should, “would”). Tier 1 and 2 users showed low levels of negative emotion, hinting that depression among these users was unlikely. Similarly, tier 3 users also showed abnormal levels of discrepancy, but still nothing that different from the average person. On the other hand, tier 4 posters exhibited four signs of depression: first-person singular pronouns, negative emotion, discrepancy, and health. This finding supports the theory that Tier 4 users may have been people who felt isolated and hopeless and, with nowhere else to turn, looked to join the white supremacist movement. However, because they did not originally harbor hateful beliefs and Iron March is extreme even by neo-Nazi standards, that type of person would probably be less likely to acclimate to the website, coinciding with our earlier observation that these users did not stay on the site for long.
The Iron March database also provided us with over 18 million words of casual discourse between some of the most extreme and dedicated white supremacists in the world. With that large of a dataset, any abnormalities in the LIWC2015 dictionary scores can indicate markers of hateful rhetoric.
Our findings confirm earlier research about the words that racist organizations are most likely to disproportionately use. In a 2018 article, Nicholas Faulkner and Ana-Maria Bliuc used LIWC2015 to analyze mission statements of racist groups and non-activist groups, and then compared certain dictionary scores to identify such markers. In particular, they found that the racist groups had higher scores for first-person plural pronouns, anger, and religion, and lower scores for achievement and the summary variable Analytical Thinking. Faulkner and Bliuc conducted their research on mission statements, which tend to be skewed as compared to ordinary writing. For that reason, rather than comparing Iron March to the Faulkner and Bliuc’s results, we compared them to the English language averages within the context of LIWC2015’s dictionary, looking at the same dictionary frequency variables.
Iron March posters were found to be significantly more likely to use anger (e.g. “hate,” “kill,” “annoyed”) and religious words (e.g. “altar,” “church”) than were general speakers of the English language. Furthermore, tier 1 users were 50% more likely than tier 4 users to user anger words, which supports our conclusion that they harbored more extreme tendencies. As Dobratz found in 2001, religion is often used by racist groups to create an identity and strengthen their position, both affirmatively (linking the white identity to Christianity) and aggressively (anti-Semitism and Islamophobia). As a result, it comes as no surprise that Iron March users were 3.8 times more likely to discuss it than were general speakers.
The posting habits and rhetoric of Iron March provide us with several valuable insights into some of the world’s most extreme and violent white supremacists. First, it shows that a small number of individuals had an out-sized impact in this community. Second, it shows that those individuals were more extreme than the others who signed up for the website. We also learned that many of the less extreme individuals were able to acclimate, but that some chose to leave instead. Finally, the rhetoric hinted to us that those users that chose to leave may have been depressed and looking for a ‘home.’ Future directions for research into the Iron March dataset include why those users left instead of acclimating, and whether the Iron March language can be used to predict neo-Nazi terrorism.