Trump, Robots and the 2016 Election
A strange trip through social media research in the era of Trump
Social media bots are real. I know, because I have the data.
During the Primary election contests in 2016, my research partner and TechRepublic senior writer Dan Patterson and I analyzed social media numbers ripped right from the Twitter API. We were looking for ways to use social media statistics to gain a theoretical handle on what unexpectedly turned out to be a new type of election. We knew that social media had been useful to previous campaigns in messaging and get-out-the-vote efforts during earlier elections, and this year, we wanted to see what role it would play, given the unexpected rise of the newest Twitter star, Donald J. Trump. So Patterson and I began collecting and analyzing Twitter data for each of the major candidates from both Parties.
In the early stages of our analysis, we had a working theory: Social media was a vehicle for voter activation, and we would see this as we compared social media activity to actual vote counts. The analyses we performed then were designed to investigate the plausibility of this working theory. As Patterson stated in his article published just ahead of Super Tuesday, from February 26, 2016: “ We’ve discovered that winning and high-performing campaigns know how to translate big data and social media activity into actionable tactics that [convert] supporters to voters.”
One of the data metrics we used to describe increasing or decreasing enthusiasm for the candidates was “Daily New Adds”: We believed that we could open a window to the future by looking at trends in increasing or decreasing enthusiasm over social media. Our data seemed to support this. As something major (e.g. a primary, a debate) occurred, we observed across-the-board spikes, for all candidates that we tracked, of new people beginning to follow the candidates. We then surmised, correctly, that major campaign events activated people’s interest in candidates, and those people expressed this interest as they flocked to the internet and social media to find out more about a candidate and his or her positions. From there, we could later assert that this formed a sort of feedback loop which tended to demonstrate growing or shrinking interest in the candidate observed in survey data and in actual vote counts in primary elections.
At first, Patterson and I were amazed at how well our working theory held up. After a debate among Republicans, interest in ALL candidates spiked at exactly the same time. The spikes were bigger for perceived winners of the event. A major win in a political contest, such as the Iowa Caucus, would lead to a large spike for Ted Cruz, though smaller spikes for other candidates, while a successful debate performance would lead to a huge jump for Bernie Sanders. We argued then that real world events were driving people to follow candidates on social media in general, and Twitter specifically, and then the candidates would predictably tout their jumps in social media as a demonstration that they had “won” or were gaining momentum. Our analysis appeared to be demonstrating something of a virtuous perception cycle, where people would follow perceived (or actual) winners, and then those winners would turn around and use social media following to demonstrate that they were winning. People say that “success breeds success”, but as it turns out, even perceived success breeds success.
But around this time, as we had collected almost two months of rather granular data, a new trend slowly began emerging.
The idea that there was a mutual variance between social media interest and public opinion had been demonstrated by our work, but this turned out to not be the most interesting finding of our analysis.
It was becoming clear to us that there was something strange about the data we were collecting on both Donald Trump’s and Hillary Clinton’s twitter accounts. The tip-off was in the “New Adds” metric itself that we were using. Following the South Carolina primary, I informed Patterson of some rather strange anomalies in the data I was analyzing: It seemed that no matter what, Trump and Clinton were adding followers faster than other front runners. Win or lose, or sometimes, after opting not to even play, Trump’s numbers (and Clinton’s, though to a lesser degree) dwarfed other candidates like Bernie Sanders, Ted Cruz and Marco Rubio. This was a trend which suggested that the rate of “new adds” was not justified by the events themselves. Why would a good showing in a debate lead any generic top tier candidate to add on average 3500 followers a day, but Trump to add an average of 20,000 new followers a day, twice Clinton’s and Sanders’ new add rate for the same period, for weeks and months at a time?
This result, which was discussed in several pieces we did on zombie followers, was first noticed, and mentioned in Patterson’s February 26th Article ahead of Super Tuesday. Patterson first correctly acknowledged that there was at least some truth to the numbers we were analyzing: The huge twitter following that Trump and Clinton commanded seemed to be turning out and voting for their chosen candidates as we expected. That fact supported our original thesis. But then he asked perhaps the most important question, in my opinion, of our extended analysis:
“A scan of third-party site TwitterAudit.com indicates that in addition to @realDonaldTrump, Democratic frontrunner @HillaryClinton also has a large percentage of junk followers… Where do zombie followers come from, and are the campaigns aware of the dead accounts?”
This question plagued us for the next month or so. The use of the “new adds” metric shone a light into a very strange corner of the 2016 election: Major candidates could deliberately manipulate the perception of the vast majority of American voters about their own popularity using social media. Election results demonstrated that there was a good reason for them to do this- It fueled the social media-driven campaign.
It soon became clear that there was something irregular about the numbers both Trump and Clinton were pulling on social media. We did some cursory analysis of other social media platforms, and discovered that they tracked pretty closely with Twitter in terms of overall increases in support, so we continued to look at Twitter. But I also altered my metric from “New Adds” to raw “Twitter Followers” and the anomaly around the rate of adds became obvious. This switch, which was described in Patterson’s April 8th article for TechRepublic came out at the same time that several other outlets were more explicitly charging Trump and Clinton with acquiring botnets (see, for example, one of the earliest reports released on the same day, in National Review, who were, at that time, part of the #NeverTrump movement on the political right, or the massive number of stories which dropped a few days later in places like FiveThirtyEight.)
All told the same story we were telling. Some mentioned specific numbers, such as “7% of Trump’s Twitter support is fake”, while others hinted at Russian or Ukrainian involvement. Whatever the headline in other reporting on this subject, the story we had been working up for more than a month, which was supported by our own, completely in-house data analysis broke across the nation in the month of April. I was quite proud to see my own suspicion about Trump’s army of Twitter support vindicated, not only in Patterson’s excellent, and “breaking news” reporting for TechRepublic, but also in the wave of outlets which published stories on the same topic in the following days. Other people were finally seeing what we had seen, and reported on, in the data, and were digging into the story.
On the Question of “buying” followers
Patterson’s April 8th article was among the first inklings that something was wrong with the narrative around Trump’s vaunted “twitter army”. But embedded in the article was the next question, begged by the observation that many of Trump’s followers were really either bots, zombies, or just straight out fake. As Patterson put it:
Are candidates using or buying “fake” Twitter followers? It’s not a crazy proposition. Celebrities and social media “specialists” have been buying fake followers for years. Though using phony followers seems off-message for political campaigns trying to leverage authenticity as an election tactic, it’s not unreasonable to assume candidates would want to bolster the perception of success by inflating accounts.
Obviously, at this early stage, we could not come right out and say, “This is happening- Candidate Trump is buying twitter followers.” As with most covert efforts, we lacked a “smoking gun” necessary to responsibly and ethically make such a bold assertion in the press. But this did not mean we could not investigate, and so we did.
At around this time, Patterson was contacted by an individual, Vlad Shevtsov, living in Novosibrsk, Russia who was familiar with the process of selling “bot nets”, or whole groups of bot followers on Twitter. Shevtsov filled Patterson in on independent research his firm had conducted on Trump’s botnet. Meanwhile, I compiled new data from a TwitterAPI site called Election Scorecard which reaffirmed our own discoveries through June- Trump’s twitter following had grown at a significantly faster rate than any of the other candidates in the Primary season.
For us, the combination of our own data and Shevtsov’s data demonstrated that Trump, and to a lesser degree, Clinton, possessed a network of fake followers, which could alternately be called “zombie” followers, who create accounts specifically to follow a few of the same accounts (A deep look into the people who these zombie accounts follow is illuminating: Apparently the same people who like Trump also love Katy Perry and also “Natural Area Rugs,” a rug store who currently has more than 7,000 followers) and then go dormant, or bots, who exist specifically to retweet and amplify the message of their clients.
I say clients, because this is precisely what a person who purchases thousands of fake accounts is: If Trump did indeed purchase thousands, or millions of followers, he did so cheaply, and he did so instrumentally. As Patterson reported, celebrities do this all the time, for the purpose of inflating the perception of their fame. Both Trump and Clinton appeared to be acquiring robots to not only increase the perception of their success, but also to retweet their posts across the social media landscape. The bots could be bought by the thousands, and would then do the work for free that used to be done by countless cadres of paid and unpaid staffers and volunteers. It was mass media on the cheap, a way for the front runners to simply dominate all conversations anyone was having about them. And so the bot net was very useful indeed.
We did not have the “smoking gun” we needed to directly tie Trump to bot purchases- that would have likely been a receipt, which may or may not have ever been written. But from where we sat, there certainly was an appearance of collusion between Trump and botsellers, and with Shevtsov’s help, the trails seemed to lead back to Ukraine and/or Russia.
The Past as Prologue: The 2016 Election and Trump’s America
As happened so often during 2016, the campaign seemed to turn on a dime. As soon as we were publishing what turned out to be one of our final collaborations over social media and the 2016 Primary’s, the campaign was off to something else- The Conventions. The whole story about possible collusion between Trump, and maybe Clinton as well, and bot brokers in Ukraine and Russia got pushed to the side as the GOP convention presented the US as a dystopia, and Trump blasted his critics via unforced errors on Twitter. Suddenly, the content of Trump’s tweets became important, regardless of how many fake followers he had on Twitter, because now his tweets were being fetishized by the Corporate Press and the 24 News channels. As summer faded into fall, and Trump’s lewd statements about women were forgotten in the wake of yet another hint that Clinton may have sent one or two inappropriate emails to her personal aide, we forgot all about a point we could not exactly prove and wrote about other things.
Before the election, even I fell victim to the general misreading of the environment and the trust in polling data. I warned in Patterson’s August 23rd story on polling data: “ Nothing can predict the future, and polls won’t tell you what’s going to happen… Polls only provide a snapshot of the current date and climate,” though apparently, two and a half months later, I did not follow my own advice. But then Trump was elected, and everything that happened as a result of that election happened.
At any rate, I can conclude that the 2016 election was an election quite like no other- This election was the first one where social media drove much of the narrative, rather than serving as a benign platform for candidates to merely reach new voters, and those candidates who did not understand that basic fact (see, for example, Jeb Bush), did not fare very well in our new environment. Meanwhile, those candidates who embraced it, Donald Trump and Bernie Sanders, did much better than anyone ever expected or predicted.
But Patterson and I have been vindicated one final time, I am happy to report. Stories about the Trump campaign’s outright collusion with Russians to build a Twitter Army of fake news robots are still being written and are now generally accepted as fact, by all but the hardest of hardcore supporters of the President, the ones for whom the President cannot possibly do any wrong. I would venture to say, these supporters are the actual human beings who have been following Trump’s occasionally mindless, often masturbatory, always incendiary Twitter feed. If we want to know who is Trump’s “Twitter Army” really was, beyond all the bots and zombies, we need look no further than the people walking about in the society, still religiously wearing their “MAGA” hats, the members and true believers of Trump’s cult of personality.
After it all, I don’t need a dataset to know that whatever Trump’s actual numbers of real, live breathing human followers on Twitter, that man or that woman was probably one of them.